c _ Copyright 2009 by Xenophon Papademetris and Alark Joshi

All Rights Reserved

This book is an edited collection of class handouts that was written for the graduate seminar
Programming for Medical Image Analysis (ENAS 920a). This class was taught at Yale University,
Department of Biomedical Engineering, in the Fall of 2006 and again in the Spring 2009 semester. Some
the comments in this draft version of the book reect this fact. For example, see comments beginning
at Yale. Furthermore, many of the references that will appear in the nal version are still omitted. It
is made available at this stage in the hope that it will be useful.
Draft Sep. 2009
Contents
Preface 5
Preface to the First Edition 6
I Introduction 8
1. Introduction 9
2. Revision Control with Subversion 14
II Programming with Tcl/Tk 19
3. Introduction to Tcl 20
4. Advanced Topics in Tcl 28
5. An Introduction To Tk 35
6. Tk Part II 43
7. Object Oriented Programming with [ Incr ] Tcl 53
8. Iwidgets: Object Oriented Programming for GUIs 63
III The Visualization Toolkit I Using Tcl 70
9. An Introduction to the Visualization Toolkit 71
10. Curves and Surfaces in VTK 75
11. Images in VTK 85
12. Displaying Images in VTK 92
13. Transforming Surfaces and Reslicing Images 104
3
CONTENTS Draft Sep. 2009
IV Interfacing To BioImage Suite using Tcl 114
14. Leveraging BioImage Suite Components and Classes 115
15. Writing your own BioImage Suite Tcl Application 125
16. BIS Algorithm Framework 152
V C++ Techniques 169
17. Cross-Platform Compiling with CMAKE 170
18. C++ Techniques and VTK 177
VI VTK Programming with C++ and Tcl 189
19. VTK Code in C++: Procedural Code and Derived Classes 190
20. Implementing Point-Based Registration 202
21. Implementing Intensity Based Segmentation 220
22. More Advanced Image Algorithm/Filters and Templated Execution 239
23. Copying Data Objects and Breaking the Pipeline 251
24. The Insight Toolkit 256
VII Appendices 267
A. Introduction to Computer Graphics - Part I 268
B. Introduction to Computer Graphics - Part II 275
C. Final Exam 282
D. Code License 283
4
Draft Sep. 2009
Preface
This is an updated version of this book based on material used for the graduate seminar Programming
for Medical Image Analysis (ENAS 920a) that was taught at Yale University, Department of Biomedical
Engineering, in the Spring of 2009. This second edition was prepared jointly with Alark Joshi, who also
helped teach the class.
5
Draft Sep. 2009
Preface to the First Edition
This book is an edited collection of class handouts that I wrote for the graduate seminar Programming
for Medical Image Analysis (ENAS 920a) that was taught at Yale University, Department of Biomedical
Engineering, in the Fall of 2006. My goal for the class was to provide sucient introductory material for
a typical 1st year engineering graduate student with some background in programming in C and C++
to acquire the skills to leverage modern open source toolkits in medical image analysis and visualization
such as the Visualization Toolkit (VTK)[?] and, to a lesser extent, the Insight Toolkit (ITK)[?].
One obviously has to acknowledge that there are many programming books out there that cover the
material discussed in this book in more depth (and more correctly perhaps); I list many of my sources
in Chapter 1. However, placed one on top of the other, these books form a pile about 1-2 feet high,
which can be discouraging to a beginner. This book is meant to be an introductory guide. Many of the
denitions are informal, much like one teaches math at 1st grade. As the student progresses and begins
to understand the introductory material, then he/she should look at more specialized books which oer
a deeper insight into what is going on. The manual pages for many of these toolkits also become useful
at this point.
Most of our graduate students the intended audience for this book and class while having a strong
applied mathematics/signal processing background, are not expert programmers. Frequently, they would
have had some programming classes at the undergraduate level and would have been, most likely, exposed
to C/C++ at some point. However, with rare exceptions, a dive into the combination of object-oriented
and generic programming model used in ITK, for instance, would leave most of them befuddled.
Such students begin their graduate research in semester long projects called special investigations.
This is part of the process of identifying a topic for their research as well as a lab in which they will
pursue their dissertation work. In our own research in medical image analysis, the typical product of a
doctoral dissertation is a mathematical framework for attacking an image analysis problem which has
to be translated into computer code for testing and validation.
Most of the students, in these special investigations, prototype ideas in MATLAB [?]. While MATLAB
is a wonderful prototyping tool, it leaves much to be desired in terms of the development of the
programming habits needed to write a large, sustainable, and reusable body of code. Unfortunately,
many students ending up in the trap of developing the best algorithms that can be implemented in
MATLAB as opposed to focusing on what on optimal algorithmic strategy would be. This is especially
apparent once large 3D and 4D datasets enter into the picture, and their algorithms end up taking hours
and days to run.
At this point in the game, a helpful professor suggests that they should probably look to move to a
more ecient language such as C++. However, one look at straight C++ without any of the additional
toolkits, makes them realize that switching to C++ is easier said than done. There are very few default
6
CONTENTS Draft Sep. 2009
operations for things like linear algebra, image processing, image display etc. Then, perhaps, another
helpful person suggests that they take a look at VTK and/or ITK. While now, they can see that there
is a ton of functionality out there, they are often lost as to where to begin. VTK and ITK are natural
tools once one is used to them but they can be imposing and scary to the beginner. While there are
some books out there (especially the VTK Users Guide) which are very helpful, they are often only
obliquely related to what they really need to learn how to do: implement image analysis methods, learn
how to (properly) display their results, and learn how to put a graphical user interface to enable them
and their potential users to interact with the methods. The goal of the course, and this book, is to
precisely provide the necessary guidance for a new graduate student in order to achieve these goals.
The selection of the material, as well as its presentation, is naturally colored by the authors own
experience. I coordinate the BioImage Suite project www.bioimagesuite.org, which is a large medical
image analysis utility developed in a mixture of C++ and [ Incr] Tcl.
1
In part, the motivation for
teaching this class and writing this book, is directly derived from the needs of the BioImage Suite
project. In particular, the driving question was, how does one get a new programmer up to speed with
the skills he/she will need in order to contribute? Perhaps somebody else teaching this class would have
used Python instead of Tcl as the scripting language; this is as much a matter of taste (and endless
discussions and sadly ame wars in this internet era) and experience with the particular language as
anything else. Also, while I mention ITK towards the end see Chapter 24, the primary toolkit used
here is VTK. The template-free interface of VTK makes it easier for beginners, and I nd that ITK can
be downright user-hostile to the less experienced programmer.
One of my complaints with many introductory texts is that they never attempt to teach how one goes
about learning how to put together decent sized applications. They focus too much on grammar and
too little on writing stories. However, graduate students need to write stories useful ease-to-use
tools that both they and clinical and/or basic science collaborators can use and maintain. Hence, as
part of the class, I have also made an attempt to introduce some software engineering tools such as
Subversion and CMAKE. As any experienced developer often learns the hard way, these tools can be
just as critical as the choice of programming language or toolkit.
Finally, a disclaimer: While VTK is now at version 5.0 we still focus on VTK 4.4 this is the version in
use in BioImage Suite at this point. As we migrate the software to VTK 5.x (a good rule of thumb is
never to use .0 release of any toolkit) this book will be updated to reect this.
Acknowledgments: BioImage Suite is supported by the National Institutes of Health (NIH)/National
Institute of Biomedical Imaging and Bioengineering (NIBIB) under grant R01 EB006494. The author
would also like to acknowledge the valuable help of Thomas Teisseyre in preparing this book.
1
Interfacing with BioImage Suite is discussed in Chapters 14 and 15.
7
Draft Sep. 2009
Part I
Introduction
8
Draft Sep. 2009
Chapter 1
Introduction
The goals of the rst chapter are (i) to ensure that all students reading this book have an appropriate
software environment to work with and (ii) to present a list of possible helper textbooks for consideration.
These lecture notes are intended as a roadmap for material made available on the class webpage.
1.1 Introduction
It has been said (see/nd the Perl web page), that the three great virtues of a programmer are laziness,
impatience and hubris. Randal Schwartz explains these in more detail in the Camel book
1
The key
one for the original class, whose lecture notes form the core of this book, is the rst one:
LAZINESS: The quality that makes you go to great eort to reduce overall energy expen-
diture. It makes you write labor-saving programs that other people will nd useful, and
document what you wrote so you dont have to answer so many questions about it. Hence,
the rst great virtue of a programmer.
This description, together with an explicit commitment to not reinventing the wheel, is a good summary
of my overall philosophy for this class, which results in the following three goals:
Provide students with resources so that they can practically teach themselves.
Do not reinvent the wheel, there are many excellent tutorials available for most of these topics,
so simply point students to them and provide supplementary material where necessary.
Provide a safety net for those who get stuck on various topics.
And critically: Ensure that the instructor does the minimum amount of teaching necessary see
denition of laziness above.
Book Structure: Programming requires hands-on work, one can take all the programming classes in
the world but will never learn how to program until one sits in front of a computer and types stu.
Hence (at least in the beginning) I plan to structure each chapter as:
1. Prepare a 4-25(!) page writeup for each chapter that outlines the material to be presented.
1
OReily publishers has a series of programming books each with a dierent animal on the cover.
9
CHAPTER 1. INTRODUCTION Draft Sep. 2009
2. Hand out the assignment for the topic, at which point the students retreat to their own computer
(e.g. laptop or one of the workstations in the MRRC) and work on the assignments. During this
time, I will be available for any additional help needed etc.
I fully expect, and will give extra credit for, students (or any readers) correcting mistakes in the book
(which I will try to keep to a minimum.)
Regular Assignments: Once the assignment is completed it will be graded (mostly by looking at
the code and a working program). The students will be required to upload all solutions to the class
subversion repository. 75% of the nal grade will be computed from the individual assignment grades.
Final Assignments: A larger nal assignment/project will be used instead of a nal. This will have a
25% towards the nal grade.
1.2 The Goals for Chapter 1
The fundamental for chapter 1 is to ensure that:
All students have a properly congured computer to work on, ideally their own laptop.
Create personalized accounts for the students for the class wiki and source code repository.
Identify students with no background in C/C++ and provide supplementary material to enable
them to catch-up by the time we get to the second half of this class.
1.3 Software Needed for this book
The following software is needed to create a proper working environment for this book:
1.3.1 Image Analysis Specic Tools
While one can get the various tools needed individually (Tcl/Tk, ITcl, Iwidgets, VTK, ITK etc.) my
suggestion (see laziness above) is to simply download and install the Yale BioImage Suite software
package www.bioimagesuite.org, which includes all of the above properly compiled to work together.
It will save you days (if not weeks) of work. If you can compile and install all of these by yourself, you
are most denitely over-qualied for this class/book and should not be reading it!
On linux make sure you install the version that matches the compiler installed on your system, if in
doubt ask! Ask your questions on the BioImage Suite Forum, which can be found at
http://research.yale.edu/bioimagesuite/forum/
1.3.2 General Tools
A good text editor: The choice of text editor often has quasi-religious connotations. In the old days
the two major editors where emacs and vi, each of which were almost equally powerful and user-hostile.
10
CHAPTER 1. INTRODUCTION Draft Sep. 2009
My own preference has always been to use emacs were possible and one is willing to put up with the
steep learning curve for this, it is highly recommended. Emacs will run on just about anything, a more
user friendly version called XEmacs might be more suitable for beginners. (Extra credit assignment:
Read about the messy Emacs vs XEmacs ame war that led to the split of the XEmacs project from
core Emacs).
On unix-like operating systems (which for desktop use means linux these days), good alternatives are
nedit and gedit. These are installed on most linux machines, I can help you install on of them if needed.
Nedit is also available for windows but requires the installation of the cygwin environment which is not
a trivial exercise.
For Windows, I have heard good words about WinEdt, TextPad, PFE (programmers File editor) &
Alpha tk. Wordpad (which comes free with windows) is also decent, avoid notepad though it is too
limited.
A C++ Compiler: On linux, this is usually installed by default (unless you are using one of the
more user-friendly modern distributions like Ubuntu). You must know what version of gcc is installed
on your system, simply type gcc -v and note the answer down. Red Hat distributions (and Red Hat
like distributions such as Fedora and Centos) often ship with two compilers, so you have a choice. On
Windows, the best option is Microsoft Visual Studio 2008 (VS 2005 is also OK for the most part). At
Yale, this can be downloaded for free! from the Yale MSDN Academic Alliance home page
(http://www.yale.edu/msdnaa/), please register and download this. The Visual Studio editor is also
very nice and can be used as a generic text editor if desired. You will need to access this from a machine
on the Yale Network, if doing this from home use Yale VPN rst to connect to this. (VPN = virtual
private network, see
http://its.med.yale.edu/software/remoteaccess/vpn/ for details/software.)
On Mac OS X (10.4 +) you will need to install the development tools from the system DVD, also install
X11 while you are at it. Other than that, Mac OS X, is very similar to linux.
Subversion: Subversion, described in detail in the Subversion Book this can be found online at
http://svnbook.red-bean.com/). Subversion is the revision control system we will use for this class.
Please make sure that it is installed on your system.
Windows users should install the more user friendly TortoiseSVN package instead.
See http://tortoisesvn.tigris.org/.
CMake: CMake or cross make www.cmake.org is also needed for the C++ portion of this class.
This is included with the windows version of BioImage Suite (see above).
1.4 Useful Books
First, a note that many of the books are available on-line from the books24x7 webpage which is at:
http://library.books24x7.com/login.asp?ic=0. This webpage is only available for machines on
the Yale network (or use Yale VPN, see above).
11
CHAPTER 1. INTRODUCTION Draft Sep. 2009
C++
C++ Demystied: A Self-Teaching Guide by Je Kent McGraw-Hill/Osborne. If you have no
experience with C++, this should get you started.
C++: A Beginners Guide, Second Edition (Beginners Guide) (Paperback) by Herbert Schildt
Publisher: McGraw-Hill Osborne Media; 2 edition (December 3, 2003)
Professional C++ by Nicholas A. Solter and Scott J. Kleper Wrox Press.
Tcl/Tk
Tcl/Tk, Second Edition: A Developers Guide (The Morgan Kaufmann Series in Software En-
gineering and Programming) (Paperback) by Clif Flynt Publisher: Morgan Kaufmann; 2 edition
(May 5, 2003)
Practical Programming in Tcl and Tk (4th Edition) (Paperback) by Brent Welch, Ken Jones,
Jerey Hobbs Paperback: 960 pages Publisher: Prentice Hall PTR; 4 edition (June 10, 2003)
VTK/ITK/CMake etc
The VTK Users Guide, Version 4.4 (Paperback) by Kitware Inc. This is out of print right now.
The ITK Software Guide 2.4 (Paperback) by Luis Ibanez; William Schroeder Publisher: Kitware,
Inc. This is freely available as .pdf le
Mastering Cmake 2.4, 4th Edition (Paperback) by Ken Martin; Bill Homan Paperback: 393
pages Publisher: Kitware, Inc. (February, 2008)
Other
Version Control with Subversion (Paperback) by C. Michael Pilato, Ben Collins-Sussman, Brian
W. Fitzpatrick Paperback: 304 pages Publisher: OReilly Media (June 22, 2004) This is freely
available online.
1.5 Syllabus
This class was taught last in the Spring of 2009. The notes from that class are available as an online
book at: www.bioimagesuite.org/vtkbook/. We will follow the same 6 part division for this class,
namely:
1. Introduction
2. Programming with Tcl/Tk
3. The Visualization Toolkit I using Tcl
4. Interfacing to BioImage Suite using Tcl
5. C++ Techniques
6. VTK Programming with C++ and Tcl
The rst two parts will be more or less identical to what was taught then. Parts 3-6 will be updated for
VTK 5.2 and BioImage Suite 3.0.
12
CHAPTER 1. INTRODUCTION Draft Sep. 2009
Assignment
Part I: Ensure that you have access to a workstation with all the necessary software for this class
(either your own laptop or access to a networked workstation in the MRRC). If you need help let me
know.
Part II: Start BioImage Suite BrainSegment and load one of the default images. This will ensure that
everything is in place.
Part III: Enter (in a text editor), save and execute the following trivial tcl script. This will ensure that
you know how to start tcl scripts on your computer. (On Windows use the BioImage Suite Console as
a command line terminal.)
#!/bin/sh
# the next line restarts using wish \
exec tclsh "$0" "$@"
set m0 $tcl_platform(user)
set m1 $tcl_platform(os)
set m2 $tcl_platform(osVersion)
set m3 [ info nameofexecutable ]
set m4 $tcl_version
puts stderr "\nHello User $m0 on [ info hostname ]"
puts stderr "You are using $m1 ($m2)"
puts stderr "Your interpreter is $m3 version $m4\n"
exit
13
Draft Sep. 2009
Chapter 2
Revision Control with Subversion
The goal of the rst chapter is to introduce the key concepts of revision control and to demonstrate its
use using the subversion package (http://subversion.tigris.org/).
2.1 Introduction
In setting out to teach this class, one my goals was not only to teach a programming techniques, but to
introduce (however cursorily) additional topics/tools that are useful for large-scale software development.
As stated in the Subversion book:
1
Version control is the art of managing changes to information. It has long been a critical tool for
programmers, who typically spend their time making small changes to software and then undoing those
changes the next day. But the usefulness of version control software extends far beyond the bounds of
the software development world. Anywhere you can nd people using computers to manage information
that changes often, there is room for version control. And thats where Subversion comes into play.
Revision control enables, especially within the context of software development in a research setting,
the following:
The ability to return to an earlier (presumably functional) version of the software following a failed
attempt to improve upon it this can be a real life-saver.
The ability to return to an older version of the software for the purpose of quantifying improve-
ments.
The ability to collaborate with other developers on a complex project and easily share code
updating, editing.
A build in backup mechanism for key documents (if the code repository is on a dierent server).
1
Much of the contents of this handout is derived from the book Version Control with Subversion by Ben Collins-Sussman,
Brian W. Fitzpatrick and C. Michael Pilato, which is freely available online (http://svnbook.red-bean.com/).
14
CHAPTER 2. REVISION CONTROL WITH SUBVERSION Draft Sep. 2009
2.2 Alternative Revision Control Systems
Subversion is but one of many choices in a revision control systems, although in my opinion it is as good
as anything out there. It is fairly user friendly (within the bounds of a complex system), it works on just
about any operating system and there are nice user-friendly GUI versions, especially the TortoiseSVN
package.
See http://tortoisesvn.tigris.org/.
Subversion was started to address issues with the CVS (http://www.nongnu.org/cvs/) which is
probably the most popular open source revision control system. CVS in turn builds upon RCS (Revision
Control System, http://www.gnu.org/software/rcs/rcs.html, which was the probably the rst
freely available source control setup).
The Microsoft Visual Source Safe system
http://msdn.microsoft.com/vstudio/products/vssafe/default.aspx is another alternative.
All of the above work much better than no source control setup, but my recommendation is for subver-
sion.
2.3 Key Concepts
Repository: The place where the ocial version of the code sits. In the case of subversion, the
repository can be either a separate location within the local computer lesystem, or accessed remotely
from a server. A repository structure can resemble that of a lesystem and multiple projects can share
a single repository.
What dierentiates a repository from simply a le server (or another directory) is that it remembers
every change ever written to it: every change to every le, and even changes to the directory tree itself,
such as the addition, deletion, and rearrangement of les and directories.
When a client reads data from the repository, it normally sees only the latest version of the lesystem
tree. But the client also has (unless one is using a web browser) the ability to view previous states of
the lesystem. For example, a client can ask historical questions like, What did this directory contain
last Wednesday? or Who was the last person to change this le, and what changes did he make?.
These are the sorts of questions that are at the heart of any version control system: systems that are
designed to record and track changes to data over time.
Repositories are accessed using their address which has a similar form to a web-page URL. There are
two types of repositories, local repositories that live on the lesystem of your own computer that have
a prex (le://) or remote repositories that live on the lesystem of a server and are accessed through
a web-server (and have a prex http:// an example of this is the class repository) or even a custom
subversion server (svn://).
Revision Number: Unlike CVS and RCS (for those familiar with these) subversion maintains a single
global revision number. Any changes to the repository result in the version number being increased.
Each revision number represents a potential version of the les and it can be used to obtain the state
of all les at that point in time. For example, consider the case when one is working on a project and
at revision 77 they have reached a good working set of code. This is used to then generate results for
15
CHAPTER 2. REVISION CONTROL WITH SUBVERSION Draft Sep. 2009
Figure 2.1: An overview of subversion. There are additional svn commands such as svn del, svn
copy, svn status.
a paper. Subsequently improvements are made to the code (raising the revision number to say 110)
and the question is raised as to how eective those changes were. From the repository we can obtain a
version of the code from revision 77 and test it against the current version.
When downloading BioImage Suite some of you may have noticed that the comment at the top of the
page SVN repository version 3467. This is a reminder of the exact cut of code in this release, and
it enables us to go back to it, were subsequent improvements to break something!
Access: First prior to starting work on a project a user checks-out a copy of the code from the
repository. (If this is a new project, then a new directory is made in the repository and then checked
out to which les may be added). The key command here is svn checkout.
Workow: Prior to starting work, the user rst synchronizes their local copy with the repository (svn
update). Then he/she performs changes to it (using a text editor presumably) and/or adds or deletes
les as needed (svn add, svn delete). Then he/she (if it all goes well) commits (svn commit)
the changes to the repository.
A dierent user then updates their local copy to re-synchronize with the repository.
Conicts and Merges: On synchronizing the local version of the les with that of the repository one
of two things can happen. If you have made any changes on a particular le that conict with the
same regions of that le on the repository, then the les are said to be in conict. You cannot commit
your changes until the conict is resolved. You can either revert your changes and accept the copy in
the repository as the master copy or you can integrate the changes by using your version. Subversion
handles conicts by creating 3 new les: lename.mine (your working copy - version before svn update),
lename.rOLDREV (older version of the le you had checked out) and lename.rNEWREV (the newest
version in the repository). Using TortoiseSVN, conicts can be resolved by going through each conicting
16
CHAPTER 2. REVISION CONTROL WITH SUBVERSION Draft Sep. 2009
region in the local and repository versions of the le and selecting which of the two regions you would
like to keep. After the conicts has been resolved, you have to execute svn resolved to tell subversion
about the resolution.
Alternatively, if the changes that you have made in your le are in a dierent region as compared to the
version in the repository, then the changes are merged and a single merged version of the le is then
created in your local repository. You still have to commit your changes to update the version of the le
on the repository.
See also the Subversion book Chapter 3.
Notes: While source control comes into its own for source code, it is also useful for storing and keeping
track of other documents (e.g. text), binary objects (images, powerpoint slides).
Users of TortoiseSVN should use the graphical user interface to accomplish these operations. The
buttons have the exact same name as the commands described above.
2.4 Repositories used in This Class
In this class each student will need to access (at least) two directories on the class subversion repository.
The rst is located at:
https://boreas.med.yale.edu/repos/enas920/00handouts/
Note the https instead of the usual http. It contains all the ocial class notes, examples etc. Please
create an empty directory somewhere (lets call it handouts) and then execute
cd handouts; svn checkout https://boreas.med.yale.edu/repos/enas920/00handouts .
The trailing dot is critical! This will place all les on the server in this directory. Subsequently as I
add les there, you can re-synchronize the les with the server by simply typing svn update in the
handouts directory. You have no permissions to modify les in this directory!
The second repository is (will be after the end of this class!) located at:
https://boreas.med.yale.edu/repos/enas920/netid/,
where one should replace netid with their netid. This is your own private directory for uploading
solutions to assignments. I suggest you make a directory on your computer (lets call it homework) and
then execute as above:
cd homework; svn checkout https://boreas.med.yale.edu/repos/enas920/netid .
The trailing dot is critical! Once you have a completed solution, place it in your homework directory
(i.e. copy the le) and then add it to the repository
svn add mysolution.tcl
svn commit -m adding solution
17
CHAPTER 2. REVISION CONTROL WITH SUBVERSION Draft Sep. 2009
The svn add adds the le the todo list for the next synchronization. This happens when the svn
commit command is executed.
You can also create your own repositories on your local machines.
Assignment
Skim Chapters 1 and 2 of the Subversion Book and read carefully chapter 3.
Part I: Checkout the handouts repository from the class server
Part II: Checkout your own private homework directory from the class server.
Part III: Create a le called assignment2.txt in your homework directory contain the line.
This is the solution to assignment 2
Add this to the repository and commit the changes.
Then edit the le to read like:
This is the solution to assignment 2.
Here is another line.
Compare this version with the version in the repository. Can you tell what has changed?
Change this again to:
This is the solution to assignment 2.
This is a replacement for the second line.
Repeat the comparison. Commit this new le to the repository.
Delete the le assignment2.txt from your local directory. Then type svn update. What happens in
this case?
Note: We will use subversion extensively for the rest of the semester so please play with it and get
comfortable using it! Once you get used to subversion you will never know how you survived without it.
18
Draft Sep. 2009
Part II
Programming with Tcl/Tk
19
Draft Sep. 2009
Chapter 3
Introduction to Tcl
The goal of the this session (and the next one) is to introduce the Tcl (pronounced tickle) script-
ing language. We will make extensive use of TclTutor, a computer aided instruction package for
learning Tcl. The TclTutor can be obtained, either from the class subversion repository or from
http://www.msen.com/ clif/TclTutor.html. The clearing house on the web for all things Tcl
is the Tcl Wiki, which can be found at http://wiki.tcl.tk.
3.1 Introduction
Learning a new programming language can be a challenge, however there are few programming languages
much easier than Tcl. For the next two sessions we will work rst on basic Tcl concepts (today) and
more advanced (modular programming) concepts (next session). The main learning tool is the TclTutor
which has 44 dierent lessons covering just about everything in the core Tcl language. In table 3.1, I
list the lesson titles and mark some of them as optional the hint is that you should focus on the other
ones rst.
Tcl is an interpreted language. The commands are executed inside a special shell (the interpreter).
There are three dierent shells that we will use for this class:
tclsh the basic Tcl language shell.
wish an extended shell which includes the Tk graphical interface toolkit.
vtk an extended shell that includes the Visualization Toolkit libraries (VTK)
All these shells are available as part of BioImage Suite. (On Windows use the BioImage Suite Console
as a command line terminal.)
3.2 Rationale
In order to develop largely accessible and user friendly applications, it is critical to identify a good
combination of a programming language (facilitates large software development), scripting language
20
CHAPTER 3. INTRODUCTION TO TCL Draft Sep. 2009
Figure 3.1: Scripting Languages can be powerful complements to more traditional compiled languages
(e.g. C++). They enable a more exible development environment.
and GUI (graphical user interface) libraries for creating easy to use user interfaces. Accessibility can
be addressed by developing cross-platform (i.e. Windows/Unix/Mac OS) applications which reduces
the choices one has as regards the GUI libraries that can be used. A number of choices exist including
WxWidgets, FLTK, Qt etc. Additionally, the ability of the GUI libraries to interface with a programming
language such as C++ further increases their use for large scale development.
In this class, I will present a hybrid programming approach that leverages both a scripting language
(Tcl) and a more traditional compiled language (C++). The rationale for the hybrid approach is shown
in Figure 3.1. While there are other scripting languages, such as Python, Ruby and Perl, Tcl in my
mind is preferable because it is the one with the best integration with a graphical user interface toolkit
(Tk). In fact, most of the other languages, directly or indirectly invoke Tk and hence an understanding
of Tcl is helpful in learning these languages as well.
BioImage Suite uses the Tcl/Tk combination together with a custom C++ wrapper around Tcl to
enable the creation of user interfaces in both a scripting language (Tcl) and C++ (as needed for
complex components).
3.3 A Guided Tour of Tcl
This section presents a guided tour of Tcl. It is meant to highlight key features at a glance and to
supplement the TclTutor. In all examples below, lines starting with % indicates the commands typed
at the command prompt and the following line(s) is (are) the expected output.
Printing: puts is the print statement in Tcl. It can also be used to print to a le. It is the Tcl
equivalent of the C commands printf and fprintf.
% puts "Hello World"
Hello World
21
CHAPTER 3. INTRODUCTION TO TCL Draft Sep. 2009
Variables: Tcl stores data as strings, and only converts to numbers when necessary for calculations
or comparisons. The assignment operator in Tcl is set. When set is called with two arguments, as in:
% set foo 10
10
% set bar "Hello"
Hello
% puts stderr "$foo $bar"
10 Hello
% set temp hello$foo
hello10
% set temp foo
foo
Flow control: For loops in Tcl have the form:
for { set i 0 } { $i < 10 } { incr i 1 } {
puts stdout $i
}
where incr is the increment operator which increases the value of the counter variable i by one. An
alternative is the while loop which has the syntax:
set i 0
while { $i < 10 } {
puts stdouti "i=$i"
incr i 2
}
Conditionals: Tcl has both an if .. elseif ... else and a switch construct for conditional
operations. These take the form
for { set a -3 } { $a<=3 } { incr a } {
puts stderr "Testing for a=$a"
if { $a > 2 } {
puts stderr "\t a is greater than 2"
} elseif { $a >= -2 } {
22
CHAPTER 3. INTRODUCTION TO TCL Draft Sep. 2009
puts stderr "\t a is between -2 and 2"
} else {
puts stderr "\t a is less than -2"
}
}
Both the elseif and else constructors are optional, note though that unlike C/C++ you must use elseif
as one word as opposed to else if.
The switch statement has the form:
switch -exact -- $osmode {
"unix" {
set libraryprefix "lib"
set librarysuffix ".so"
}
"windows" {
set libraryprefix ""
set librarysuffix ".dll"
}
}
Math Operations: All math operations make use of the expr command. The following statement is
not valid unfortunately because by default everything is a string!
% set i 2+2
2+2
% set i [ expr 2+2 ]
4
expr parses the operation numerically i.e. not as a string!
Lists: Lists are a complex variable structure, and they are literally everywhere in Tcl. If in doubt as
to what something is, it is a list! A new list can be created using the list command.
% set l [ list 1 2 3]
1 2 3
23
CHAPTER 3. INTRODUCTION TO TCL Draft Sep. 2009
We can access an element of the list (rst element has index 0) using the lindex command:
% set a [ lindex $l 1 ]
2
To get the length of the list (i.e. how many items it has) use the llength command.
% set a [ llength $l ]
3
There are other commands such as lappend, lsort, linsert etc. A good understanding of how lists
work is critical for understanding Tcl!
Strings: Every variable is implicitly a string this is the key strength and weakness of Tcl. Strings
can be manipulated using the string command
% set l [ string length Hello ]
5
% string range "Hello" 2 4
llo
% string index "Hello" 1
e
There are lots of other options.
Associative Arrays: Another complex variable structure in Tcl is the associative array. To create an
array (implicitly) use:
% set a(1) "Hello"
% set a(2) "Help"
% set a(other) "Complicated"
% puts stdout "$a(1), $a(2), $a(3)"
Hello, Help, Complicated
Arrays can be modied using the array command.
24
CHAPTER 3. INTRODUCTION TO TCL Draft Sep. 2009
File Input/Output: Tcl uses the two standard output channels stderr and stdout (standard output
default and standard error). New les can be created using the open command.
set fileid [open "/winnt/temp/testfile" w]
puts $fielid "This is the first line"
puts $fielid "This is the second line"
close $fileid
To read an existing le use:
set fileid [open "/winnt/temp/testfile" r]
gets $fielid firstline
gets $fileid secondline
puts stdout "Read\n $firstline \n $secondline"
close $fileid
Filename Manipulation: The file command can be used to parse lenames. For example consider
the following:
set fname /home/papad/vtkpxcontrib/CMakeLists.txt
set a [ file extension $fname ]
.txt
% set a [ file rootname $fname ]
/home/papad/vtkpxcontrib/CMakeLists
% set a [ file tail $fname ]
CMakeLists.txt
% set a [ file dirname $fname ]
/home/papad/vtkpxcontrib
% set a [ file size $fname ]
1024
There are lots of options for copying les, deleting les, moving les/directories etc.
Arguments to a Script The arguments passed to the script on execution are stored in the global
variable argv which is a list of all arguments. Unlike C/C++ the rst item in the list is NOT the name
of the script itself, this is stored in the variable argv0.
If a script is executed as:
tclsh myscript.tcl a.hdr 10.0
25
CHAPTER 3. INTRODUCTION TO TCL Draft Sep. 2009
then argv is the list a.hdr,10.0.
Assignment
1. Download the TclTutor package from handouts repository
2. Also download the tkconsole.tcl from the same location
The URL is you need to do a manual download is
https://boreas.med.yale.edu/repos/enas920/00handouts/session3/, if you have suc-
cessfully installed subversion and checked-out the handouts directory a simple svn update will
place both of these les into the session3 subdirectory.
3. Unzip the TclTutor package and execute it using
wish TclTutor.tcl
4. Work through lessons 1-9, 14-19, 22-24. Most of these are fairly short and should not take up
too much time. Just try to grasp the basic concepts.
Note: This assignment is really the rst half of the assignment to be handed out at the end of the
next session. At that point I will provide unnished programs for you to complete to do specic tasks.
26
CHAPTER 3. INTRODUCTION TO TCL Draft Sep. 2009
Lesson 0 Introduction
Lesson 1 Simple Text Output
Lesson 2 Assigning values to variables
Lesson 3 Evaluation & Substitutions 1: Grouping arguments with
Lesson 4 Evaluation & Substitutions 2: Grouping arguments with
Lesson 5 Evaluation & Substitutions 3: Grouping arguments with []
Lesson 6 Results of a command - Math 101
Lesson 7 Textual Comparison - switch
Lesson 8 Numeric Comparisons 101 - if
Lesson 9 Looping 101 - While loop
Lesson 10 Looping 102 - For and incr
Lesson 11 Adding new commands to Tcl - proc
Lesson 12 Variations in proc arguments and return values
Lesson 13 Variable scope - global and upvar
Lesson 14 Tcl Data Structures 101 - The list
Lesson 15 Adding & Deleting members of a list
Lesson 16 More list commands - lsearch, lsort, lrange
Lesson 17 String Subcommands - length index range
Lesson 18 String comparisons - compare match rst last wordend
Lesson 19 Modifying Strings - tolower, toupper, trim, format
Lesson 20 Optional Regular Expressions 101
Lesson 21 Optional More Quoting Hell - Regular Expressions 102
Lesson 22 Associative Arrays.
Lesson 23 More Array Commands - Iterating through array
Lesson 24 File Access 101
Lesson 25 Information about Files - le, glob
Lesson 26 Optional Invoking Subprocesses from Tcl - exec, open
Lesson 27 Optional Learning the existence of commands and variables ? - info
Lesson 28 Optional State of the interpreter - info
Lesson 29 Optional Information about procs - info
Lesson 30 Modularization - source
Lesson 31 Building Libraries of procs - unknown, info library
Lesson 32 Creating Commands - eval
Lesson 33 More command construction - format, list
Lesson 34 Substitution without evaluation - format, subst
Lesson 35 Changing Working Directory - cd, pwd
Lesson 36 Optional Debugging & Errors - errorInfo errorCode catch error return
Lesson 37 Optional More Debugging - trace
Lesson 38 Optional Command line arguments and environment strings
Lesson 39 Optional Leftovers - time, unset
Lesson 40 Optional Channel I/O & socket, leevent, vwait
Lesson 41 Optional Time and Date - clock
Lesson 42 Optional More channel I/O - fblocked & fcongure
Lesson 43 Optional Child interpreters
Table 3.1: List of Lessons in the Tcl Tutor.
27
Draft Sep. 2009
Chapter 4
Advanced Topics in Tcl
The goal of the this session is to introduce key modularization concepts in Tcl. In particular I will discuss
the following: procedures, packages and namespaces. Just like the previous session , we will make ex-
tensive use of TclTutor, a computer aided instruction package for learning Tcl. The TclTutor can be ob-
tained, either from the class subversion repository or from http://www.msen.com/~clif/TclTutor.html.
A reminder that the clearing house on the web for all things Tcl is the Tcl Wiki, which can be found at
http://wiki.tcl.tk.
4.1 Introduction
Writing a long program in a single chunk, while doable (I inherited some cardiac code in C that was a
single main function that run over 1,000 lines!) is not the best idea in terms of maintenance, exibility
and code reuse. Most programs of any length consist of multiple pieces that can then be reused for
similar tasks in ways more elegant than simple copy and paste (although all of us do a fair amount of
this too!).
The most elementary form of modularization is the packaging of code into specialized functions that
usually perform simple tasks. Functions are termed procedures in Tcl and are specied using the proc
command.
The next step is to move some of these functions into a separate le (or collection of les) to form a
reusable library, code that is invoked by multiple programs. These separate les can either be invoked
directly (using the source command which the rough C-equivalent of #include) or by using the package
mechanism in Tcl which provides additional functionality.
The use of multiple libraries in large programs can often lead to name conicts, as for example in the
case of using two separate libraries each of which denes a procedure called LoadImage. Namespaces,
are an elegant way of avoiding such conicts. In some respects namespaces are similar to classes in
object-oriented languages in that they can provide for code and data encapsulation.
4.2 Procedures
The core common method for modularizing code is the use of procedures. A procedure in Tcl is dened
using the proc command. The syntax is:
28
CHAPTER 4. ADVANCED TOPICS IN TCL Draft Sep. 2009
proc myfunction { argument1 argument2 } {
... some code
}
For example, a trivial script (script4-1.tcl) which contains and addition procedure looks like:
proc addtwonumbers { x y } {
return [ expr $x + $y ]
}
set a [ addtwonumbers 2 3 ]
set b [ addtwonumbers 3 $a ]
puts stderr "a=$a, b=$b"
Returning Multiple Values: As a general rule, all complicated issues in Tcl are resolved using lists.
This applies to the present case. A procedure can only return one thing, if you need to return multiple
things, make a list and return the list! Consider the following lename parsing procedure (script4-2.tcl
has a complete example)
proc parsefilename { filename } {
set path [ file dirname $filename ]
set tail [ file tail $filename ]
set rootname [ file root $tail ]
set extension [ file extension $tail ]
set output [ list $path $rootname $extension ]
return $output
}
Scoping: The use of procedures (and namespaces as we will discuss later) introduces an additional
complication, the issue of scoping. Variables dened inside a procedure are, by denition, local variables
to that procedure the main program (global variables) are not available in a procedure, unless explicitly
accessed using the global command. This is dierent from other languages, such as C, where global
variables are always available.
Consider the following case (script4-2b.tcl)
set defaultextension .hdr
29
CHAPTER 4. ADVANCED TOPICS IN TCL Draft Sep. 2009
proc changeextension { filename } {
global defaultextension
set rootname [ file root $filename ]
set extension [ file extension $filename ]
set output [ list "${rootname}${defaultextension}" ]
return $output
}
puts stderr [ changeextension myfile.img ]
Here the variable defaultextension is made available to the changeextension procedure using the global
command, which tells the procedure that the variable defaultextension is the global variable defaultex-
tension, as opposed to a local variable to be used within the procedure.
In general, global variables are a bad idea. They are a prime source of what is often termed name
pollution. A good metric of code quality is the total number of global variables, the fewer the better.
If you need to use global variables, I suggest that you group them together in an associative array. For
example the script above could have been rewritten as (script4-2c.tcl)
set parameters(defaultextension) .hdr
proc changeextension { filename } {
global parameters
set ext $parameters(defaultextension)
set rootname [ file root $filename ]
set extension [ file extension $filename ]
set output [ list "${rootname}${ext}" ]
return $output
}
The associative array trick has the advantage of keeping the total number of globally accessible names
to one! There are other constructs available (particularly upvar) which I recommend you avoid!
Note: There are ways to make procedures take variable numbers of arguments (using the args argument)
and to dene default values for the arguments.
30
CHAPTER 4. ADVANCED TOPICS IN TCL Draft Sep. 2009
4.3 Using Multiple Files
The source command Consider the case where the parselename procedure is placed in a le
mylib.tcl. A separate script (script4-3.tcl) can invoke this using:
# The next commands includes mylib.tcl
source mylib.tcl
set testfilename "c:/temp/hello.txt"
set results [ parsefilename $testfilename ]
puts stderr "\tThe path name is: [ lindex $results 0 ]"
puts stderr "\tThe main name is: [ lindex $results 1 ]"
puts stderr "\tThe extension is: [ lindex $results 2 ]"
Packages An additional level of exibility (or complexity) is introduced using the package mechanism.
A package is a collection of procedures, or even more simply a name for the contents of a .tcl le that
is separate from the actual lename.
To create a formal package we need two things: (a) the package denition using the package
provide, typically the top le in the .tcl le containing the code and (b) Create an index le (pkgIndex.tcl)
in the directory containing the .tcl le (or les) that matches package names to lenames.
Consider the mylib.tcl le. Its rst few lines look like:
package provide mylibrary 1.0
proc parsefilename { filename }
The rst line species that this le contains the abstract entity mylibrary version 1.0. A dierent le
could contain mylibrary version 2.0 etc.
The pkgIndex.tcl le can either be created automatically using the pkg mkIndex command (in the tclsh
interpreter) or by hand, which is my preferred option. In the simple case of the directory containing a
single package this takes the form:
package ifneeded mylibrary 1.0 [list source [file join $dir mylib.tcl]]
which means: if you need mylibrary version 1.0, this is located in the le mylib.tcl in the same directory
as pkgIndex.tcl. If there are more than package in the directory, each will have an entry in pkgIndex.tcl.
The script (script4-4.tcl) can be rewritten to make use of the package mechanism as:
31
CHAPTER 4. ADVANCED TOPICS IN TCL Draft Sep. 2009
# Append to auto_path the directory containing the current script
lappend auto_path [ file dirname [ info script ]]
package require mylibrary 1.0
The variable auto path contains a list(!) of paths in which the interpreter will look for packages. We
append to this the directory containing the current script. If we want to invoke other packages (e.g.
from BioImage Suite which are stored in /usr/local/bioimagesuite/base) we could add the command
lappend auto_path /usr/local/bioimagesuite/base # (unix)
lappend auto_path c:\yale\bioimagesuite\base # (windows)
The second line (package require) tells the interpreter to look in all the pkgIndex.tcl les in the directories
listed in auto path to nd which one has a match for the requested item mylibrary version 1.0.
The main reason for using packages has to do with the ease with which dierent versions of the code
can be swapped in or out, or moved about. It also makes using other libraries (written by other people
easier) as there is no need to know exactly which le to source, only which package to request. In a
newer version of the libraries the lenames may change but with an appropriate new pkgIndex.tcl le old
code can remain functional. This provides a level of abstraction between the code interface and the code
implementation, which is a common theme in many of the more advanced programming techniques that
we will encounter. In general the interface (what the user uses to access code), need not directly match
the implementation (the underlying code). The interface often acts as a tidy easy to see showroom for
a messy storage room (implementation) where the real work is done!
4.4 Namespaces
Just like global variables pollute the name space, procedures can also have the same eect. Consider
the aforementioned example of needing to use two separate libraries each of which denes a procedure
called LoadImage. In such a case, we have a naming conict which means that only of the libraries
can be used. There are two ways around this: (i) the use of suitably long names for all procedures (i.e.
xpMyUtilLoadImage) to ensure unique names (but unreadable code) or the (ii) the more elegant way of
using namespaces.
A namespace is a collection of procedures and variables that are grouped together and to which the
interpreter adds implicitly a prex that distinguishes them from other code. An example (newname.tcl)
hopefully will clarify the concept:
namespace eval mynewname {
# All code from this line
variable counter
32
CHAPTER 4. ADVANCED TOPICS IN TCL Draft Sep. 2009
set counter 0
proc newname { } {
variable counter
incr counter
return "new$counter"
}
proc longnewname { } {
set a [ newname ]
set b long${a}
return $b
}
# to this line is enclosed in the mynewname namespace
}
A namespace is created using the namespace eval namespacename command whose argument (a
list!) is the entire code in the namespace. The variable counter and the procedure newname live inside
the namespace.
Inside the namespace we can call the procedures newname and longnewname using their regular names
(i.e. newname and longnewname). Variables dened inside the namespace (e.g. counter) can be made
accessible inside each procedure using the variable command which is the namespace equivalent of
the global command.
Accessing the code from outside the namespace is a dierent game (script4-4.tcl).
lappend auto_path [ file dirname [ info script ]]
package require mynewname
set firstname [ ::mynewname::newname ]
set secondname [ ::mynewname::longnewname ]
puts stderr "The firstname is $firstname, the second name is $secondname"
puts stderr "The current value of the counter is $::mynewname::counter"
Here all functions and variables dened in the mynewname namespace can be accessed only by adding
a ::mynewname prex to their names. Hence the counter variable becomes ::mynewname::counter and
newname becomes ::mynewname::newname etc. While this may seem unwieldy, the general idea is
that most functions in a namespace are only meant to be accessed from within the namespace (the
implementation!) and only one or two functions are to be accessed from the outside (the interface).
This type of functionality is already a step towards object oriented programming, that we will discuss
later.
33
CHAPTER 4. ADVANCED TOPICS IN TCL Draft Sep. 2009
Also note, that (i) symbols can be exported from a namespace into the global scope (a bad idea, so
I will not cover this) and (ii) that namespaces can be nested, i.e. a namespace can be dened inside
another one. In this case we can have paths like ::mynewname::mysubnumspace::myfunction and so on.
Again though, the general idea is that stu in a nested namespace is not meant to be accessed from
outside the namespace itself or at most its parent namespace.
Assignment
General hint. The construct:
while { [ gets $fileid line ] >=0 } {
...
}
will read a line at a time from le $leid and place it into the variable line until there is nothing left to
read from the le.
Turning in assignments: Make a directory chapter4 and add it to you repository directory. Place all
solution in this.
1. TclTutor: Work through lessons 11-13, 24,30,34, in addition to those recommended in the last
assignment.
2. File search/replace: Write a script that opens le script4-4.tcl and replaces all stderr with stdout
and save the result in script4-4-stdout.tcl. Hint: Take a look at the regsub command.
3. Package Use: Write a short library (name it mymath.tcl) with two functions add and subtract.
Make this declare a package (package provide mymath) and add this to the index le (pkgIndex.tcl).
Then write a second script that reads a set of numbers from a provided le (numbers.txt) and for each
line print out the original numbers, as well as their sum and their dierence. (Hint: you may want to
investigate the scan command for parsing each line as it is read from the le.)
4. Arguments: Modify the script in exercise 2 so that the lename of numbers.txt is not hard-wired
but read as an argument on the command-line, i.e. it should be executed as tclsh script3.tcl
numbers.txt.
5. Conditionals: Modify the script in exercise 3 so that the program only prints the dierence of the
numbers if it is positive, otherwise it prints the second number is bigger than the rst.
In the next chapter, we will examine the creation of graphical user interfaces using the Tk Toolkit.
1
1
Incidentally, since somebody asked, the domain Dot TK (.tk), is the domain of the island of Tokelau. Extra Credit:
Where is Tokelau?
34
Draft Sep. 2009
Chapter 5
An Introduction To Tk
Tk is an open source, cross-platform widget toolkit, that is, a library of basic elements for building
a graphical user interface (GUI). This chapter will introduce the basic concepts in Tk, while the next
chapter will handle more advanced/esoteric aspects of the toolkit. The clearing house on the web for
all things Tcl is the Tcl Wiki, which can be found at http://wiki.tcl.tk. A good introduction to
Tk is available in chapter 18 of Brent Welchs book Practical Programming in Tcl and Tk the chapter
is freely available online at http://www.beedub.com/book/2nd/TKINTRO.doc.html. Much of this
handout is based on a set of lecture notes by Marcel Jackowski that was presented in an IPAG seminar
a few years back.
5.1 Introduction
The task of writing a program with a graphical user interface (GUI) for the rst time can be a major
challenge. GUI-based programming is a vastly dierent concept from standard command-line (CLI)
programming. In CLI programming the program (and by implication the programmer) is in control.
There is only one entry point into the program and the general ow through the program (i.e. order
in which statements are executed) is more or less predictable with few permutations. In GUI-programs,
in contrast, the user is really in control, as the master. The program (and hence the programmer) is
really the slave/butler whose job is to react to events initiated by the master (e.g. clicking on a button)
with appropriate responses (often captured in terms of callback procedures). A key part of the job of a
programmer in this environment, is to ensure that only valid actions are taken. For example, an image
can be displayed once it is loaded but not before. In such a case, clicking the DisplayImage button
prior to the LoadImage button should result in an error (a helpful user friendly hand-holding message)
as opposed to an attempt to display a non-existent image which will likely crash the program.
The execution of GUI-based program consists of two distinct phases (i) initialization which includes
the creation of the GUI and (ii) the event loop which for the most part is the process of waiting
on the user to generate some event which requires an appropriate response (event handling). Once an
event is handled the program just sits there until the next event is generated, or the user issues an exit
command which stops the program (or in many cases the program crashes, but one should try to avoid
this scenario!)
35
CHAPTER 5. AN INTRODUCTION TO TK Draft Sep. 2009
5.2 What is Tk?
Tk is a freely available open-source GUI toolkit implemented with Tcl and C. It runs on multiple
platforms: X/Motif, Win32 GUI, Mac GUI. Its simplicity enables fast development of GUIs with few
lines of code. It also allows for easy creation of new GUI controls. Tk is used both in commercial
packages (e.g. Mayo Clinics Analyze) and other large scale Medical Image analysis tools (e.g. Slicer,
BioImage Suite).
The following are the basic concepts in Tk Programming
Windows (such as dialog boxes)
Widgets: windows with a particular look and feel e.g. buttons, labels etc.
Class commands: create dierent widgets (e.g. button .a -text "Button" -command "exit")
Widget commands: congure widgets (e.g. .a configure -bg black)
Geometry management commands: place, pack & grid commands (e.g. pack .a -side top)
-side top)
Event bindings (bind .a <ButtonPress> puts %b )
A First Example: This script (script5-1.tcl) creates two buttons (Print and Exit) and waits for them
to be pressed.
proc PrintButtonCallback { } {
puts "Print Button Pressed!"
}
button .b1 -text "Print" -command { PrintButtonCallback }
button .b2 -text "Exit" -command { exit }
pack .b1 .b2 -side top
The two lines beginning with the button command create two button widgets (Print and Exit). The
button command, much like all widget creating commands, has the syntax:
button widgetname -optionname optionvalue -optionname2 optionvalue2 ...
The widget name has the form parent.newwidget, where parent is the parent widget (which contains
the new widget) and newwidget is the particular name of the new widget. The base widget (which is
created by default) has the name . (dot), and in this case the widgetnames of the two buttons are
.b1 and .b2. The button command has a number of options, two of the most common ones are
-text (which species the name of the button) and -command (which species the action to be taken
36
CHAPTER 5. AN INTRODUCTION TO TK Draft Sep. 2009
when the button is pressed.) While the widgets are created using the button command, they do not
appear until they are managed using the pack command which adds them to the display.
5.3 Widgets - Part I
The Core Tk (in a couple of chapters we will also discuss extensions such as the Iwidgets package) has
the following core widgets:
1) button 2) canvas 3) checkbutton 4) entry 5) frame 6) label 7) labelframe 8) listbox 9) menu
10) menubutton 11) message 12) tk optionMenu 13) panedwindow 14) radiobutton 15) scale 16) scroll-
bar 17) spinbox 18) text 19) toplevel - Create and manipulate toplevel widgets .
In addition to these, Tk also provides the following complex popup dialogs to simplify processes such as
lename selection and generating message boxes to alert the user: i) tk chooseColor - pops up a dialog
box for the user to select a color. ii) tk chooseDirectory - pops up a dialog box for the user to select
a directory. iii) tk dialog - Create modal dialog and wait for response iv) tk getOpenFile - pop up a
dialog box for the user to select a le to open or save. v) tk messageBox - pops up a message window
and waits for user response. vi) tk popup - Post a popup menu .
Finally, Tk oers three dierent types of geometry managers, namely: a) place - which positions widgets
at absolute locations, b) grid - which arranges widgets in a grid, c) pack - which packs widgets into
a cavity, . These managers control how dierent widgets managed by a single parent appear on the
screen. The most common manager is the packer we will use this exclusively in this chapter.
In the next sections, we will briey demonstrate the use of these widgets. Each widget can have
additional options not covered here, the Tcl Wiki is a good source of information for this, or any Tk
book.
5.3.1 Frames and Toplevels
The Frame Widget: The frame widget is a simple widget which simply acts as a container for other
widgets. Frames are extremely useful and are used constantly in Tk. For example, consider the case
when one would like to create two rows of buttons. The easiest way to do this is to (i) rst create two
frames, a top frame and a bottom frame and pack them into the parent window. Then (ii) the top row
of buttons can be packed into the top frame and the bottom into the bottom frame. Frame widgets
are also useful as dividers by explicitly setting their height and/or width and their background color.
Class Command: frame, e.g. frame .top
Common Options: -bg (background), -height, -width
An example of the above settings is given in the script (script5-2.tcl) below.
proc PrintButtonCallback { number } {
puts "Print Button number $number Pressed!"
}
frame .top;
37
CHAPTER 5. AN INTRODUCTION TO TK Draft Sep. 2009
frame .middle -height 10 -bg black;
frame .bottom
pack .top .middle .bottom -side top -expand false -fill x
button .top.b1 -text "Print 1" -command { PrintButtonCallback 1 }
button .top.b2 -text "Print 2" -command { PrintButtonCallback 2 }
pack .top.b1 .top.b2 -side left -expand true -fill x
button .bottom.b1 -text "Exit" -command { exit }
pack .bottom.b1 -side left -expand true -fill x
The configure method can be used to set an option after the widget is created, e.g. to change the
color of the frame .middle to blue use
.middle configure -bg blue
The cget command is used to return the current value of an option. To get the current value of the
background color use:
set bgcolor [ .middle cget -bg ]
The Toplevel widget: This is similar to a frame widget with the key addition that instead of it
being placed inside a parent widget, it is created as a standalone window. Consider the example below
(script5-3.tcl)
wm protocol .dlg WM_DELETE_WINDOW { grab release .dlg ; wm withdraw .dlg}
wm title . "Main Window"
wm title .dlg "Second Dialog"
wm geometry . 800x60
wm geometry .dlg 400x40
pack .top .middle -side top -expand false -fill x
button .top.b1 -text "Print 1" -command { PrintButtonCallback 1 }
button .top.b2 -text "Print 2" -command { PrintButtonCallback 2 }
button .top.b3 -text "Show Exit Dialog" -command { wm deiconify .dlg }
pack .top.b1 .top.b2 .top.b3 -side left -expand true -fill x
38
CHAPTER 5. AN INTRODUCTION TO TK Draft Sep. 2009
button .dlg.b1 -text "Exit" -command { exit }
pack .dlg.b1 -side left -expand true -fill x
wm withdraw .dlg
The wm title command can be used to set the title of the window and the wm geometry command
can be used to set the dimensions of the window. The dialog is shown using wm deiconify and hidden
(minimized) using wm withdraw.
Note:The line beginning with wm protocol is a critical piece of boiler plate code that ensures that the
dialog box is not destroyed when closed. (This is particularly the case in Windows!)
5.3.2 Labels and Buttons
The Label Widget: The label widget is a simple widget that simply places some static text in a
window.
Class Command: label, e.g. label .top -text "Some Text"
The Button Widget: This is eectively extension of the label widget which executes a command
when clicked. It is important to remember that the command is executed in global scope. I strongly
suggest avoiding lengthy code in the -command option; instead I recommend creating explicit callback
handling procedures and using the -callback to call these.
The Checkbutton Widget: This presents the user with a check that can be set to on or o. The
key additional option is that of a variable (global scope, hence global variable) whose value (1=on,0=o)
reects the state of the widget.
The Radiobutton Widget: This is similar to the checkbox widget ,but here, the widget variable is
shared among multiple radiobuttons, each of which, sets it to a unique value and only one of which
can be on!
The Option Menu Widget: This is a dierent form of the radiobutton widget.
The following example (script5-4.tcl) demonstrates the use of the widgets mentioned this far.
proc Reset { } {
global parameters
set parameters(usegradient) 1
set parameters(usehessian) 0
set parameters(color) "red"
set parameters(smoothness) "Low"
}
39
CHAPTER 5. AN INTRODUCTION TO TK Draft Sep. 2009
proc PrintStatus { } {
global parameters
puts stderr "Using Gradient : $parameters(usegradient)"
puts stderr "Using Hessian : $parameters(usehessian)"
puts stderr "Color: $parameters(color)"
puts stderr "Smoothness: $parameters(smoothness)"
}
Reset
frame .top; frame .middle -height 10 -bg black;
frame .middle2; frame .bottom
pack .top .middle .middle2 .bottom -side top -expand false -fill x
button .top.b1 -text "Print Status" -command { PrintStatus }
button .top.b2 -text "Reset" -command { Reset }
pack .top.b1 .top.b2 -side left -expand true -fill x
button .bottom.b1 -text "Exit" -command { exit }
pack .bottom.b1 -side left -expand true -fill x
set w1 [ frame .middle2.left ]
set w2 [ frame .middle2.middle ]
set w3 [ frame .middle2.right ]
pack $w1 $w2 $w3 -side left -expand true -fill both
checkbutton $w1.1 -variable parameters(usegradient) -text "Use Gradient"
checkbutton $w1.2 -variable parameters(usehessian) -text "Use Hessian"
pack $w1.1 $w1.2 -side left -expand true -fill x
radiobutton $w2.a -variable parameters(color) -text "Red" -value "red"
radiobutton $w2.b -variable parameters(color) -text "Green" -value "green"
radiobutton $w2.c -variable parameters(color) -text "Blue" -value "blue"
pack $w2.a $w2.b $w2.c -side left -expand true -fill x
label $w3.a -text "Smoothness:"
tk_optionMenu $w3.b parameters(smoothness) "Low" "Medium" "High"
pack $w3.a $w3.b -side left -expand true -fill x
Note that all parameters are stored in global scope in the single global associative array parameters.
This is to keep name pollution down. The original (default) values of the array are set by a call to the
procedure; this is to allow for easy restoration of default values. Note also that any widget command
returns a value which is the same as the actual widgetname. It is best in complex scripts to not hardwire
the names of the widgets but to actually use variables.
The Entry Widget: The entry widget enables the input of a single line of text. The value of the
widget is captured in the associated textvariable. Setting the variable changes the widget, and conversely,
typing text in the widget changes the value of the variable. We can change the example above to switch
from using the tk optionMenu to an entry widget as follows (script5-5.tcl):
40
CHAPTER 5. AN INTRODUCTION TO TK Draft Sep. 2009
entry $w3.b -textvariable parameters(smoothness) -width 10 -relief sunken
An often useful modication is to disable text input into an entry widget and use it purely as a place to
display the value of a variable. This is accomplished using:
$w3.b configure -state disabled; # to disable
$w3.b configure -state normal; # to enable
The Scale Widget: This is a slider widget that allows the setting of a variable by dragging a scale.
It has the syntax:
scale widgetname -variable variablename -from lowvalue -to highvalue
We can modify the previous example to add a scale as follows (script5-6.tcl):
scale $w3.b -variable parameters(smoothness) -from 1.0 -to 5.0 -digits 2 \
-orient horizontal -resolution 0.1
Also the default value of parameters(smoothness) is now 1.5, as a scale only takes numerical values!
Note: In the next chapter we will continue with more complex widgets such as the listbox, the text
and the canvas widget. We will also take a look at some of the more complex controls available for
selecting lenames and colors, as well as providing feedback to the user in the form of messageboxes.
Assignment
Note: A more detailed Tk assignment will be handed out at the end of the next chapter. The goal of
this assignment is to get to be familiar with the basic concepts in Tk.
Download Marcel Jackowskis presentation (jackowski tk.ppt) and work through it.
Take script5-3.tcl as a starting point, and rearrange the buttons such that Print 1 is to the right
of Print 2.
Take script 5-4.tcl as a starting point and modify the PrintStatus method to save the current
parameters in a le parameters.txt in addition to printing them on the screen.
Take script 5-4.tcl as a starting point, and modify the PrintStatus method to set the color of the
.middle variable to be the same as the value of the parameters(color).
41
CHAPTER 5. AN INTRODUCTION TO TK Draft Sep. 2009
Take script 5-4.tcl as a starting point, and rearrange the widgets so that the graphical user interface
is transposed i.e. buttons are going top to bottom instead of left to right.
42
Draft Sep. 2009
Chapter 6
Tk Part II
In this chapter, building on the foundations laid by chapter 5, we will explore some additional topics
in Tk. This chapter is divided into two parts.First we will rst examine some additional widgets, such
as menus and common dialogs. Then we will look at a fairly complete application, a simple text
editor, which demonstrates how a complex GUI-based application looks and illustrates some defensive
programming concepts. A reminder that the clearing house on the web for all things Tcl is the Tcl
Wiki, which can be found at http://wiki.tcl.tk. A good introduction to Tk is available in chapter
18 of Brent Welchs book Practical Programming in Tcl and Tk the chapter is freely available online
at http://www.beedub.com/book/2nd/TKINTRO.doc.html.
6.1 Common Dialogs
Tk has a number of precongured complex common dialog boxes. These can be used (this is not an
exhaustive list) for:
Asking yes/no type question and giving warning/error messages tk messageBox
Getting a lename to load from tk getOpenFile
Getting a lename to save into tk getSaveFile
Selecting a directory tk chooseDirectory
Selecting a color tk chooseColor
We examine each of these in turn, next, in the context of an example script (script6-1.tcl) shown below:
# Procedure for setting global values
proc Reset { ask } {
global parameters
if { $ask == 1 } {
set ok [ tk_messageBox -type yesno -default no -title "Think again"
-message "Reseting" -icon question ]
if { $ok == "no" } { return }
}
set parameters(readname) ""
set parameters(writename) ""
set parameters(directory) ""
}
43
CHAPTER 6. TK PART II Draft Sep. 2009
# Set Background Color
proc SetColor { widget } {
set color [ tk_chooseColor -title "Set Background Color" -parent . ]
if { [ string length $color ] > 0 } {
$widget configure -bg $color
}
}
# Execute one of the three dialogs to set filenames/directory
proc GetName { mode } {
global parameters
set filetype1 [ list "Text Files" [ list .txt .tex ] ]
set filetype2 [ list "All Files" "*" ]
set f ""
switch -exact $mode {
"directory" { set f [tk_chooseDirectory -title "Select Current Directory" ] }
"readname" { set f [tk_getOpenFile -title Load \
-filetypes [ list $filetype1 $filetype2 ] ] }
"writename" { set f [tk_getSaveFile -title Save -filetypes [ list $filetype1 ] ] }
}
if { [ string length $f ] < 1 } { return }
set parameters($mode) $f
}
Reset
# Create GUI one line at a time in frame.$i
for { set i 1 } { $i <= 3 } { incr i } {
set parent [ frame .frame$i ]
pack $parent -side top -expand false -fill x
# First set the things that are different for each line
switch -exact $i {
"1" { set name "Input File:"; set mode readname }
"2" { set name "Save File :"; set mode writename }
"3" { set name "Directory :"; set mode directory }
}
# Create the GUI for each line -- note that use of eval
label $parent.1 -text $name -width 20
entry $parent.2 -textvariable parameters($mode) -width 80
eval "button $parent.3 -text \"...\" -command { GetName $mode }"
# Pack in order to make sure what appears and what can stretch
pack $parent.3 -side right -expand false -padx 5
pack $parent.1 -side left -expand false -padx 2
pack $parent.2 -side left -expand true -fill x
}
# Create bottom button bar
frame .bottom; pack .bottom -side bottom -expand true -fill x -pady 10 -padx 20
button .bottom.1 -text "Reset" -command { Reset }
44
CHAPTER 6. TK PART II Draft Sep. 2009
Figure 6.1: A snapshot of script6-1.tcl.
button .bottom.2 -text "Color" -command { SetColor . }
button .bottom.3 -text "Exit" -command { exit }
pack .bottom.1 .bottom.2 .bottom.3 -side left -expand true -fill x -padx 2
Message Box: The message box control, invoked in the Reset procedure, can be used to inform the
user of some event (or error) and ask simple yes/no/cancel questions. Its key options are
-type this determines which buttons to display. Valid values include: yesno, yesnocancel,
ok, okcancel ...
-icon this determines the icon in the dialog. This is one of: error, info, question
,warning.
-default name of the default button (e.g. yes)
-message message to display in the dialog
-title the tile of the window
-parent the parent window of the dialog which determines where it will appear.
The output of this command is the name of the button that was pressed (e.g. yes).
Choose Directory: The choose directory control tk chooseDirectory can be used to select a di-
rectory. Its options include
-parent, -title same as above for message box.
-initialdir species the initial directory, if omitted the current directory is used.
-mustexist whether the directory must exist (set this to true or false)
The output of the tk chooseDirectory command is the name of the selected directory, or an empty string
if cancel was pressed.
Open File: This control (tk getOpenFile) invokes the custom dialog for selecting a lename to
open. Its options include:
-parent, -title, -initialdir same as above for choose directory, same goes for the out-
put.
-initialfile this can be used to set the selection.
-filetypes this is a list of extensions allowed. This has to be of the form of a list of lists!
Each component list has the format:
"Analyze Files" "*.hdr" where the rst argument is the human-readable type of the le,
and the second species the lter to use (e.g. extension .hdr). The second argument can itself
be a list(!) to specify multiple extensions as is done in the example above (script6-1.tcl).
45
CHAPTER 6. TK PART II Draft Sep. 2009
Save File: This control (tk getSaveFile) invokes the custom dialog for selecting a lename to write
to. Its options include:
-parent, -title, -initialdir,initialfile, filetypes same as above for open le,
same goes for the output.
-defaultextension this lets the program add a default extension (e.g. words .doc) when
none is specied
Choose Color: This control (tk chooseColor) brings up a dialog for the user to select a color. It
has the following options:
-parent, -title same as above for open le.
-initialcolor this lets the program specify the initial color. The output of the tk chooseColor
command is the name of the selected color, or an empty string if cancel was pressed.
The eval command Towards the end of script6-1.tcl, one would have expected to see a line of the
form:
button $parent.3 -text \"...\" -command { GetName $mode }
as opposed to:
eval "button $parent.3 -text \"...\" -command { GetName $mode }"
This is a subtle point of great importance. The problem with the rst statement (i.e. the one
without eval) is that the implicitly dened callback procedure GetName $mode is not executed at the
time the button command is executed, rather it will be executed when the button is pressed. At this
point in time, the variable $mode does not necessarily exist or have the appropriate value (it will always
be set to, in this case, directory), which will cause unpredictable problems (Try it!). What we really
want to specify, for example, during the rst iteration of the loop ($i=0) is something like GetName
readname. However, this is cumbersome, and sometimes impossible. A way around this is do to a
double parsing of the code using the eval command.
As the manual page states: Eval takes one or more arguments, which together comprise a Tcl script
containing one or more commands. Eval concatenates all its arguments . . . and passes the concatenated
string to the Tcl interpreter recursively, and returns the result of that evaluation (or any error generated
by it) . . .. In this case the eval command takes $mode and replaces it with its current value (readle)
before executing the button command, which achieves the desired result. This will be a critical step
when we start using the object-oriented extensions..
There is nothing really complex here, other than (i) recognizing that the callback function has variables
whose values are either not going to be available or uncertain at the time of the execution and (ii) using
the eval command to recursively parse the command (e.g. button) rather than parsing it directly.
46
CHAPTER 6. TK PART II Draft Sep. 2009
Figure 6.2: A snapshot of script6-2.tcl.
6.2 Menubars and Menus
The main menu appears at the top of a window, from which hang submenus, each of which contains
collections of buttons (both regular buttons, as well as checkbuttons and radiobuttons). Menus are very
useful for making a large number of options available in a small amount of screen real-estate. While
there are many subtle issues here, a simple example will clarify things. We will modify the previous
example to move the three buttons (Reset, Color and Exit) into a menu structure. The bottom part of
this new script (script6-2.tcl) is shown below:
# Create the Main Menubar and attach it to the main widget
set menubase [ menu .top ]; . configure -menu $menubase
# Create two submenus
set filem [ menu $menubase.file -tearoff 0 ]
set filee [ menu $menubase.edit -tearoff 1 ]
# Attach the submenus to the main menubar
$menubase add cascade -underline 0 -label "File" -menu $filem
$menubase add cascade -underline 0 -label "Edit" -menu $filee
# Add buttons to the file submenu
$filem add command -label "Color" -command { SetColor . }
$filem add separator
$filem add command -label "Exit" -command { exit }
# Add buttons to the edit submenu
$filee add command -label "Reset" -command { Reset 1 }
Creating a Menu: Both the main menu, and the submenus are created using the menu command.
A key option of the menu command is -tearoff which determines whether the menu can become a
standalone dialog if needed for easier access.
The main menu is attached to its parent widget by specifying it as the option -menu of the widget, i.e.
.configure -menu $menubase, where in this case . is the parent widget.
Adding Entries to a Menu: Items (entries) can be added to a menu using the add option. The
syntax for adding a submenu is:
$parentmenu add cascade -label Submenuname -menu $childmenu
Similarly other kinds of entries can be added. The most common ones are:
command which adds a button, e.g.
$filem add command -label Color command SetColor .
47
CHAPTER 6. TK PART II Draft Sep. 2009
separator which adds a line to space of items on the menu.
check which adds a checkbutton, e.g.
$filem add check -variable parameters(swap) -label Swap
radio which adds a radiobutton, e.g.
$filem add radio -variable parameters(displaymode) -label ShowAsSurface
If your program only has a few button like elements, then menus are probably not the best GUI-
element, in fact newer operating systems like Windows Vista seem to be going away from them. On
the other hand a complex program with lots of options will probably benet for having a menu setup.
An additional script (script6-3.tcl) illustrates the use of add check, I will not discuss it in any detail in
the handout.
6.3 A Complete Application
The following example (script6-4.tcl) illustrates the use of the text widget as the heart of a simple text
editor. The text widget can be used to store multiple lines of text (and has other advanced features).
The program consists of the following pieces:
Initialize Parameters
Create Scrolled Text Box
Procedures for Manipulating the Text Box
Procedures for Loading and Saving the contents of the Text Box
A Main GUI Generation procedure.
The main program itself
We examine them in turn, beginning with the initialize parameters function which is straightforward.
proc InitializeParameters { } {
global parameters
set parameters(enabletimestamp) 0
set parameters(enableediting) 1
}
Creating the Text Widget: Then we create the text using the text command. We can also attach a
scrollbar created using the scrollbar command. Note the use of eval to set the name of the scrolling
callback.
proc CreateScrolledText { basewidget } {
global parameters
Figure 6.3: A snapshot of the simple
text editor (script6-4.tcl)
48
CHAPTER 6. TK PART II Draft Sep. 2009
set w [ labelframe $basewidget.log -text "Text File" ]
pack $w -side top -fill both -expand t
eval "text $w.log -width 80 -height 10 -borderwidth 2 -relief raised \
-setgrid true -yscrollcommand {$w.scroll set}"
set cmd "scrollbar $w.scroll -command { $w.log yview} "
eval $cmd
pack $w.scroll -side right -fill y
pack $w.log -side left -fill both -expand true
return $w.log
}
Manipulating the Text Widget: The following three procedures illustrating simple clearing the wid-
get, adding text and setting its state. Unlike simpler widgets, there is no variable that automatically
contains the contents of this widget, the widget needs to be queried for its contents. The key item here
is the concept of index which has the format line.character and it starts at 1.0. The word end is
shorthand for the last character in the le.
proc ClearAllText { } {
global parameters
set numlines [ expr int([ $parameters(textbox) index end ]) -1 ]
if { $numlines > 1 } {
set ok [ tk_messageBox -type yesno -default no -title \
"Think again ..." -message "This will erase your current file?" \
-icon question ]
if { $ok == "no" } { return 0 }
}
# Delete from start (1.0) to end
$parameters(textbox) delete 1.0 end
return 1
}
proc AddANewLine { } {
global parameters
set line "Some More Text"
if { $parameters(enabletimestamp) } {
set currenttime [ clock format [ clock seconds ] -format "%H:%M:%S on %d %b %Y" ]
set line "$line at $currenttime"
}
$parameters(textbox) insert end "$line\n"
$parameters(textbox) see end
}
proc SetEditableState { } {
global parameters
if { $parameters(enableediting) == 1 } {
$parameters(textbox) configure -state normal
} else {
$parameters(textbox) configure -state disabled
}
49
CHAPTER 6. TK PART II Draft Sep. 2009
}
Loading/Saving the Text: The next two procedures illustrate loading and saving the text in a de-
fensive programming setup where you assume that everything that can go wrong will. The catch
command is a special tcl construct for handling exceptions (errors). Placing a command inside catch
ensures that no system errors are given. Rather the program has a chance to gracefully respond
to the error (e.g. we have no permission to read or write to a specic le).
To get the number of lines, we explicitly ask for the value of the end index, truncate it to its integer
part and subtract 1!
proc Load { } {
global parameters
set ok [ ClearAllText ]
if { $ok == 0 } { return 0 }
set fname [tk_getOpenFile -title "Load Filename" \
-filetypes { { "Text Files" {.tex .txt } }} ]
if { [ string length $fname ] < 1 } { return 0 }
set fileid 0; catch { set fileid [open $fname r] }
if { $fileid == 0 } {
tk_messageBox -type ok -title "Error ..." \
-message "Cannot read file $fname" -icon error
return
}
while { [ gets $fileid line ] >=0 } {
$parameters(textbox) insert end "$line\n"
}
close $fileid
wm title . "Simple Text Editor [ file tail $fname]"
}
proc Save { } {
global parameters
set numlines [ expr int([ $parameters(textbox) index end ]) -1 ]
set fname [tk_getSaveFile -title "Save Filename" \
-filetypes { { "Text Files" {.tex .txt } }} ]
if { [ string length $fname ] < 1 } { return }
set fileid 0; catch { set fileid [open $fname w] }
if { $fileid == 0 } {
tk_messageBox -type ok -title "Error ..." \
-message "Cannot save to file $fname" -icon error ]
return
}
50
CHAPTER 6. TK PART II Draft Sep. 2009
for { set i 1 } { $i < $numlines } { incr i } {
set ip [ expr $i + 1 ]
set txt [ $parameters(textbox) get $i.0 $ip.0 ]
puts -nonewline $fileid "$txt"
}
close $fileid
wm title . "Simple Text Editor [ file tail $fname]"
}
Main GUI Generation: This is a fairly straight forward GUI generation procedure, with a menu at
the top and a textbox at the bottom.
proc GenerateGUI { } {
global parameters
set menubase [ menu .menu ]
. configure -menu $menubase
frame .top; pack .top -side top -expand true -fill both
set filem [ menu $menubase.file -tearoff 0 ]
set filee [ menu $menubase.edit -tearoff 0 ]
$menubase add cascade -underline 0 -label "File" -menu $filem
$menubase add cascade -underline 0 -label "Edit" -menu $filee
$filem add command -label "Clear Text" -command { ClearAllText }
$filem add separator
$filem add command -label "Load" -command { Load }
$filem add command -label "Save" -command { Save }
$filem add separator
$filem add command -label "Exit" -command { exit }
$filee add check -label "Enable TimeStamp" -variable parameters(enabletimestamp)
$filee add check -label "Enable Editing" -variable \
parameters(enableediting) -command { SetEditableState }
$filee add separator
$filee add command -label "Add A New Line" -command { AddANewLine }
set parameters(textbox) [ CreateScrolledText .top ]
}
Main Program: Finally the main program itself, short and sweet.
InitializeParameters
GenerateGUI
wm title . "Simple Text Editor"
51
CHAPTER 6. TK PART II Draft Sep. 2009
Assignment
Naturally, read and understand the scripts referred to in this handout. Note that there are slight
dierences between the actual scripts and the versions presented in this handout, mainly to save
space.
Using script6-1.tcl as a base add a function to print the contents of the le whose lename is
specied under Input File.
Using script6-2.tcl modify the menu to add a Help Menu. Add a function under this labeled
About which gives (via a messagebox) some brief information about the application.
Using script6.3.tcl modify the GUI to add a status line (an entry widget) at the bottom. Add a
function to the edit menu called Info that gives information about the current le (e.g. number
of lines, and anything else that you can get).
52
Draft Sep. 2009
Chapter 7
Object Oriented Programming with [
Incr ] Tcl
Object-oriented programming (OOP) is a key concept. It represents a paradigm shift from procedural
(structured programming). The key dierence is that, while in procedural programming variables and
procedures are distinct entities, in OOP the functionality (i.e. procedures) is embedded into the data
structures (variables) themselves to create what are called classes. We will explore OOP using the [Incr]
Tcl extension to Tcl. The webpage for the [Incr] Tcl project is at http://incrtcl.sourceforge.net/itcl/.
Many new concepts in Tcl, including namespaces, began as part of [Incr] Tcl.
7.1 Introduction
In the evolution of programming some of the main milestones where (i) the transition to structured
programming, (ii) the use of object-oriented programming (OOP) and more recently (iii) generic pro-
gramming (we may get to this at the end of the semester). Each of this transitions was a result of the
need to be able to write bigger programs with less code. The Visualization Toolkit (VTK), which we
will get to in a couple of chapters, is an object-oriented library and understanding how object-oriented
libraries work and how to extend them is at the heart of this class. Rather than taking the more con-
ventional approach and describing OOP using C++, I will instead take a detour and use the [Incr] Tcl
OOP extension for the Tcl language instead to ease the transition.
Object Oriented Programming: In a Wikipedia article, OOP is distinguished from procedural pro-
gramming by the short mnemonic that in OOP the verb is always attached to the object! For example
consider the following statements:
1
Subject-oriented: The Sales Application saves the Transaction vs Object-oriented: The Transaction
saves itself upon receiving a message from the Sales Application
Subject-oriented: The Sales Application prints the Receipt vs Object-oriented: The Receipt prints itself
upon receiving a message from the Sales Application
1
In the discussion below, Subject-oriented=procedural for our purposes.
53
CHAPTER 7. OBJECT ORIENTED PROGRAMMING WITH [ INCR ] TCL Draft Sep. 2009
In OOP, instead of having simple variables which store values and are operated on by external pro-
cedures, we have objects. Objects are active variables which combine data storage and functionality.
Hence objects have the ability to do stu when instructed by the main program.
An example we will (perhaps over) use extensively in this chapter is the distinction between a four-month
old infant and four year old child. Consider the all-important (once you have children you will appreciate
this more!) concept of going to the bathroom (or more likely cleaning up the situation once this is
completed!). For all intents and purposes the infant is, a mostly, passive participant. The action is
on the part of the parent Change the baby. On the other hand a four year old (when things are
going well), can be told to Go to the bathroom and she has the capability to do so and take care of
things. In procedural programming, variables are infants they can do nothing of themselves. Rather,
the program (the adult) operates on them using dierent procedures. In OOP, variables are children,
they have the ability to do things as prompted by the main program.
A key rst concept is to distinguish between class and object. The class is the generic (e.g. infant)
where as the object represents a concrete incarnation or instance of the class (e.g. Alex).
To stretch this metaphor a little bit, infants and four-year-olds have some properties in common (e.g.
instances of both groups have a name). In programming terms, we can rst design the simpler of the
two classes (infant in this case) and then write the second class (fouryearold) as explicitly extending
(or deriving from) infant. In this way some common functionality need only be implemented once, thus
eliminating potential bugs in the code. Also if down the road we need to add functionality that is
common to both (e.g. a social security number), this needs only to be added to the infant class and it
will automatically be inherited by the derived fouryearold class.
Object hierarchies and their design are critical concepts. In programming with VTK one needs to under-
stand how the object hierarchy in VTK is put together for tasks as simple as reading the documentation.
Some of the functionality of classes such as Images is simply inherited from parent classes (e.g. DataOb-
ject) and one needs to read both pages to get a full handle on what the Image class can do and how
to interact with it.
What is [Incr] Tcl [Incr] Tcl is probably the most commonly used object-oriented extension for Tcl.
The Tcl core language is not object-oriented at its core (unlike say Java or Python), and OOP is added
using extensions such as [Incr] Tcl, Others include Xotcl and snit.
7.2 The Infant Class
A First Class Denition: A class consists of the following:
1. Member variables these store data (e.g. myName)
2. Member functions these are the procedures that are embedded in the class. Two special functions
are:
The constructor: This method is automatically called when the object is created.
The destructor: This method is automatically called when the object is deleted.
An additional feature is that the methods can be dened in one of two ways. (i) Either explicitly in
the class denition (we will use this for the constructor and destructor), or (ii) outside the main class
denition (we will use this convention for everything else).
54
CHAPTER 7. OBJECT ORIENTED PROGRAMMING WITH [ INCR ] TCL Draft Sep. 2009
A nal concept, before looking at a real example, is the distinction between public and protected
(and private). Public variables/methods can be accessed from outside the class, where as protected
variables/methods can only be accessed from inside the class. It is a good idea to make most variables
protected and to provide methods for accessing their value and/or modifying them.
Without further ado, here is an example class (infant) from the script le (infant.tcl). First there are the
usual directives for package handling. The second line (package require Itcl) loads the OOP extensions.
package provide Infant 1.0
package require Itcl 3.2
Next comes the class denition itself. This is performed using the itcl::class command whose
arguments are the name of the class (e.g. Infant) and the class denition.
itcl::class Infant {
# Class Variables
protected variable myWeight 2.0
protected variable myName "Anonymous"
# Constructor and Destructor
constructor { newname } {set myName $newname }
destructor {
set myName ""
set myWeight ""
}
# Interface, public methods
public method GetName { }
public method GetWeight { }
public method SetWeight { wgt }
public method DailyRoutine { }
public method PrintSelf { }
# Protected methods
protected method Cry { }
};
The nal ; ends the class denition. The denition consists of four parts:
The member variables (myWeight and myName) both declared protected.
The constructor and the destructor.
The Interface dened in terms of the public methods. These are the methods that outsiders can
invoke (e.g. the main program).
The Protected methods which can only be invoked from within the class.
Finally we move to the implementation both the public and the private methods. This is very similar to
implementing procedures with two key exceptions: (i) use the syntax itcl::body Classname::methodname
55
CHAPTER 7. OBJECT ORIENTED PROGRAMMING WITH [ INCR ] TCL Draft Sep. 2009
as opposed to proc methodname and (ii) all class member variables (e.g. myWeight and myName in
this case) are explicitly available as if a global-like command had been executed. A nal variable
this is also always dened, this is the name of the class instance or object itself. For example, the
method Cry is invoked from within the DailyRoutine method using $this Cry.
::itcl::body Infant::GetName { } { return $myName }
::itcl::body Infant::GetWeight { } { return $myWeight }
:itcl::body Infant::SetWeight { wgt } { set myWeight $wgt }
::itcl::body Infant::DailyRoutine { } { puts stderr "infant ... Daily Routine Start"
$this Cry; puts stderr "infant ... Daily Routine End"
}
::itcl::body Infant::Cry { } { puts stderr "infant ... WaWa!"}
::itcl::body Infant::PrintSelf { } {
puts stderr "infant ... myName = $myName"
puts stderr "infant ... myWeight = $myWeight"
}
Interacting with a Class: The following script (script7-1.tcl) demonstrates the use of the infant class.
lappend auto_path [ file dirname [ info script ]]
package require Itcl 3.2; package require Infant
set leanboy [ Infant \#auto "A" ]
$leanboy SetWeight 7.0
set fatboy [ Infant \#auto "B" ]
$fatboy SetWeight 11.0
puts stderr "\nLets see the details on $leanboy"
$leanboy PrintSelf
$leanboy DailyRoutine
puts stderr "\nLets see the details on $fatboy"
$fatboy PrintSelf
itcl::delete object $leanboy; itcl::delete object $fatboy
The rst thing that needs to happen in order to leverage a class is to create a specic instance of the
class. (The generic concept infant does not cry, it is only real infants that do!). This is accomplished
by calling the constructor as follows:
set leanboy [ Infant \#auto "A" ];
56
CHAPTER 7. OBJECT ORIENTED PROGRAMMING WITH [ INCR ] TCL Draft Sep. 2009
Figure 7.1: An example class hierarchy. The classes get pro-
gressively more complex with new functionality being added
at each level.
The constructor is invoked using Infant command (which is the name of the class). The constructor
takes an implicit argument which is the name of the object (in C/C++ terms this is the address of the
pointer). While one can specify this manually, it is best to let the system generate an automatic pointer
name using the #auto construct. The rest of the arguments are those specied in the constructor
denition itself, in this case there is only one (newname) which is the (human) name of the infant. In
this case we will call this baby A. The result is stored in the variable leanboy which can be used to
access the class.
Baby A can be made to do all the things dened by its public methods (e.g. GetName, GetWeight,
SetWeight, DailyRoutine and PrintSelf). One thing we can do (following script7-1.tcl) is to set its
weight, using:
$leanboy SetWeight 11.0
We can also tell leanboy to print itself using $leanboy PrintSelf or to go through his daily routine
$leanboy DailyRoutine. We can also instantiate a second baby (fatboy) and interact with it similarly.
When an instance of a class is no longer useful, it can be removed from memory, by calling its destructor.
This is invoked using the itcl::delete object command which takes one argument, the (pointer) name of
the instance.
7.3 Inheritance Trees
Consider the case where you have some code (either set of procedures or a class) that is almost what
you need but needs a slight tweak or some additional functionality. In procedural code doing this would
involve: (i) copying the code, (ii) renaming the functions and (iii) change the piece(s) that needs
adaptation. A downside of this method is that should you need to add functionality that is common to
both the original and the copy, this needs to added twice (in each piece).
The same task in an Object Oriented Programming environment reads: (i) derive a new (child) class
from existing class. (we will call this the parent class) (ii) Add new methods and override the methods
that needs changing. Crucially if new code, that is common to both classes, needs to be added later, it
can simply be added to the parent class and the child class automatically inherits it.
57
CHAPTER 7. OBJECT ORIENTED PROGRAMMING WITH [ INCR ] TCL Draft Sep. 2009
Figure 7.2: A comparison of
infant and fouryearold.
The inherit statement: At the top of the class denition there is the usual class command and
then the key statement inherit which species that the class FourYearOld is derived from the class
Infant.
itcl::class FourYearOld {
inherit Infant
Constructor and Destructor: Following this, there are some new member variable denitions (my-
FavoriteFood, myFavoriteColor) and then the next key step is the denition of the nested constructor
as follows:
constructor { name favcol } {
::Infant::constructor $name
} {
set myFavoriteColor $favcol
set myWeight 20.0
}
The constructor for FourYearOld rst calls the constructor for Infant and passes the name argument
to it. Then the constructor for Infant itself is executed. The destructor is also implicitly nested in
the opposite order. First the explicit destructor for the FourYearOld class is called followed by the
destructor for the Infant class.
New Methods: The following methods are new: GetFavoriteFood,SetFavoriteFood,
GetFavoriteColor, Eat and GotoBathroom. These represent functionality that is absent in the Infant
class (in a developmental interpretation, these are new skills that develop later in life!).
58
CHAPTER 7. OBJECT ORIENTED PROGRAMMING WITH [ INCR ] TCL Draft Sep. 2009
Redened or Overridden Methods: The methods DailyRoutine, PrintSelf and Cry exist in both
Infant and FourYearOld but the versions in the derived class either redene or amplify the functionality
of the original. Consider the denitions or Cry and PrintSelf respectively given below. The new Cry
method simply redenes the old Cry method.
::itcl::body FourYearOld::Cry { } {
puts stderr "fouryearold ... Mom I would like to watch a movie!"
}
The new PrintSelf method on the other hand explicitly amplies the old one and it explicitly calls
it to do half the work. This is accomplished by the invocation of ::Infant::PrintSelf which prints
the values of myName and myWeight. Then the new method adds the information that is new to
FourYearOld but not present in infant. In this way if down the road a new member variable (e.g.
myHeight) is added to the Infant class, it will automatically be printed for an instance of FourYearOld
without needing to modify the laters PrintSelf method directly.
::itcl::body FourYearOld::PrintSelf { } {
::Infant::PrintSelf
puts stderr "fouryearold ... myFavoriteFood = $myFavoriteFood"
puts stderr "fouryearold ... myFavoriteColor = $myFavoriteColor"
}
The example script below (script7-2.tcl) illustrates the use of the FourYearOld class. It is similar to
the previous example script (script7-1.tcl) and will not be described in detail here.
}
lappend auto_path [ file dirname [ info script ]]
package require FourYearOld
set girl [ FourYearOld \#auto "M" "Pink" ]
$girl SetWeight 35.0
$girl SetFavoriteFood "PB & J"
puts stderr "\nLets see the details"
$girl PrintSelf
puts stderr "\nHere is [ $girl GetName]s daily routine"
$girl DailyRoutine
itcl::delete object $girl
A third class in the hierarchy Teenager is also dened which adds additional functionality. The code
is in the le teenager.tcl and an example invocation is in the example script7-3.tcl. We will not discuss
them in any detail here. (In any event given my general lack of experience with teenagers I would rather
not comment on the matter too much at this stage, I may have to revisit this nine years from now!)
59
CHAPTER 7. OBJECT ORIENTED PROGRAMMING WITH [ INCR ] TCL Draft Sep. 2009
7.4 Static/Common Methods and Variables
Often there is a need for functions or variables that are tightly associated with a class. Such helper
procedures are often needed by code in the class denition but do not need an active instance of
the class for them to function. To revisit a previous statement, while the generic concept infant does
not cry, but only real infants do, there are such things that are properties of the generic class infant as
opposed to individual infants. One such function would be the statistic How many infants are there?.
While such functionality could be implemented in separate functions/variables, a more elegant way
is to implement them within the class denition as explicit properties of the collective. Such meth-
ods/variables are called either static (C++ terminology) or common ([Incr] Tcl) terminology.
Consider the example below (script7-4.tcl). To simplify, both the class denition/implementation and
the main code are in the same le. At the top we have a variant of the Infant class called IrsInfant
(IRS=Internal Revenue Service) which has an explicitly social security number (mySSN) eld. A chal-
lenge is to make sure that all instances of IrsInfant have unique SSNs (otherwise this would mess-up
the tax collection). This is accomplished by maintaining a common counter in the class (NumInfants)
as well as an explicit common method (NewSSN) for generating a new social security number.
The rst part of the class denition is unremarkable and follows closely the original Infant denition
with some omissions to keep the code short.
package require Itcl 3.2
itcl::class IrsInfant {
protected variable myName "Anonymous"
protected variable mySSN "0"
constructor { newname } {
set myName $newname
set mySSN [ ::IrsInfant::NewSSN ]
}
destructor {
set myName ""
set mySSN ""
}
# Interface
public method GetName { }
public method GetSSN { }
public method PrintSelf { }
The common/static methods and variables are dened next. The variables are declared common using
the common directive in place of the variable directive. The methods are declared common using the
proc directive instead of the method directive, as is shown below. Note that the constructor above
calls the static function NewSSN using the explicit designation ::IrsInfant::NewSSN. In some respects
common/static variables and procedures use the class as an implicit namespace to cut down on name
pollution.
60
CHAPTER 7. OBJECT ORIENTED PROGRAMMING WITH [ INCR ] TCL Draft Sep. 2009
# Common
protected common NumInfants 0
protected proc NewSSN { }
public proc GetNumInfants { }
};
The implementation of the non-static methods is very simple:
::itcl::body IrsInfant::GetName { } { return $myName }
::itcl::body IrsInfant::GetSSN { } { return $mySSN }
::itcl::body IrsInfant::PrintSelf { } {
puts stderr "myName = $myName"
puts stderr "mySSN = $mySSN\n" }
Finally the static/common methods are implemented. The assigned SSN is equal to 1000 plus the
number of infants; the number of infants is incremented each time a request for a NewSSN is issued.
The public common method GetNumInfants simply returns the current value of the NumInfants
variable.
::itcl::body IrsInfant::NewSSN { } {
incr NumInfants; return [ expr 1000+ $NumInfants ]
}
itcl::body IrsInfant::GetNumInfants { } { return $NumInfants }
The main script (below) simply allocates 5 new infants with names C1,...,C5 and prints their description.
for { set i 1 } { $i <=5 } { incr i } {
set baby($i) [ IrsInfant \#auto "C$i" ]
$baby($i) PrintSelf
}
puts stderr "The Total Number Of Infants = [ ::IrsInfant::GetNumInfants ]"
7.5 Additional Notes
Object Hierarchies can also come in tree like structures where two classes can have the same parent
class but each is going their separate ways (e.g. my brother is in banking although we share common
genes, we also have properties that are dierent e.g. salary!)
There is also the possibility of mixed inheritance, i.e. a class can have two parents. I suggest you avoid
this, it makes things really messy in my opinion.
61
CHAPTER 7. OBJECT ORIENTED PROGRAMMING WITH [ INCR ] TCL Draft Sep. 2009
Figure 7.3: An more complex class hierarchy (VTK 4.4 transformation classes).
Assignment
Make sure you understand the following concepts:
1. Object Hierarchy
2. Inheritance
3. Overriding
4. Static/Common
5. Instance vs Class
6. Constructors and Destructors,
7. Protected vs Public (read also about protected vs private)
In particular write a two sentence description of each. Place your answers in a text le and upload
this in the repository in the usual way.
Download the .pdf le itclitk.pdf. This contains two chapters by Michael McLennan who is the
original author of [Incr] Tcl. Read chapter 1. The namespace material (pages 47-63) is outdated
and to some extent confusing. You can also skip (if you dont have time) anything past page 37.
Chapter 2 is interesting but can be skipped for now.
Taking script7-2.tcl as a base, create an additional fouryearold object and set its favourite food
to something and also its weight. Use PrintSelf to verify that the properties have been set.
Merge some of the changes from IrsInfant into Infant (i.e. the social security number stu) and
re-run script7-3.tcl to observe whether these changes also apply to an instance of the Teenager
class.
Create an additional class (collegestudent) derived from Teenager.
1. Add at least one additional member variable with appropriate Get/Set Methods for modifying
and obtaining its value. Appropriately modify the PrintSelf method to print this also.
2. In addition change the GotoSchool method to only print I am going to school if the current
day is a Monday through a Friday. You can get the current date using
set a [ clock seconds ]; set d [ clock format $a -format %w ]
where the value of d is the weekday as a number (Sunday = 0, Saturday = 6).
3. Modify pkgIndex.tcl to reect the presence of the new class and write a script similar to
script7-3.tcl to exercise the new class.
62
Draft Sep. 2009
Chapter 8
Iwidgets: Object Oriented Programming
for GUIs
Object oriented libraries are key tools in large programming tasks. In the next part of the book,
we will explore the Visualization Toolkit with its multifaceted functionality for image processing and
visualization. In this chapter, we explore a smaller but extremely useful library the Iwidgets toolkit.
This provides a number of so-called megawidgets complex combinations of basic widgets packaged as
[ Incr ] Tcl classes which enable the quick and easy construction of complex graphical user interfaces.
Libraries such as Iwidgets form the backbone of most large applications. The iwidgets.pdf le in the
repository contains an excellent introduction to the topic, this handout is meant to supplement this
material and provide a roadmap.
8.1 Introduction
The Iwidgets library leverages the [Incr] Tcl object oriented extensions to the basic Tcl language. Before
launching into a full scale description of Iwidgets, there is one additional concept from [Incr] Tcl that
needs to be introduced, namely the use of cget and configure to read and modify public member
variables respectively. Consider the following script (script8-1.tcl) which denes the class PubInfant and
invokes it. Note that PubInfant has two member variables, one protected (myName) and one public
(myWeight). To get the value of myName we need to invoke the explicitly dened GetName member
function. Since, myWeight, is a public variable it can be directly accessed using the congure method
to set its value and the cget method to get its value, as illustrated in the script.
package provide Infant 1.0
package require Itcl 3.2
itcl::class PubInfant {
public variable myWeight 2.0
protected variable myName "Anonymous"
constructor { newname } { set myName $newname }
destructor { set myName ""; set myWeight "" }
public method GetName { } { return $myName }
};
63
CHAPTER 8. IWIDGETS: OBJECT ORIENTED PROGRAMMING FOR GUIS Draft Sep. 2009
# --------------------------- Main Code ---------------
set leanboy [ PubInfant \#auto "Alex" ]
$leanboy configure -myWeight 8.0
puts stderr "The weight of [ $leanboy GetName ] is [ $leanboy cget -myWeight ]"
Note also that unlike previous examples, in this case, the code for GetName is placed right in its
denition as opposed to having a separate code block for it. This is often desirable for short methods.
8.2 Introducing Iwidgets
Figure 8.1: Top The Iwidgets class hierarchy. Middle: The demo pro-
gram which can be used as a tutorial for all Iwidgets, and Bottom: Invok-
ing the demo.
Iwidgets is a collection of so-called
mega-widgets. These are combinations
of Tk widgets that are used as build-
ing blocks to enable the creation of
more complex graphical user interfaces.
Consider the example of the need for
the combination of a label and an en-
try widget, which is a fairly frequent
occurrence (this is illustrated in iwid-
gets.pdf). In ordinary Tk, this would
require three widgets, (i) a frame, (ii)
a label and (iii) an entry. Iwidgets pro-
vides a single mega-widget called entry-
eld which presents precisely this com-
bination and some additional features
as well.
Figure 8.1(top) presents the hierarchy
for the Iwidgets library. A critical skill in
Object Oriented Programming and es-
pecially when leveraging outside code
is the ability to read such hierarchy
diagrams. For example, consider the
case of the class Entryeld. This is a
child of LabeledWidget which in turn
is a child of itk::Widget which is -
nally in turn a child of itk::Archetype.
Typically most class libraries have a
single (or a very small number) of
toplevel parent(s) (e.g. itk::Archetype)
which denes functionality common to
all classes. Then, progressively func-
tionality is both specialized (via over-
riding) or added to create new classes.
Note also that the parent classes (e.g.
itk::Archetype) are often never meant
64
CHAPTER 8. IWIDGETS: OBJECT ORIENTED PROGRAMMING FOR GUIS Draft Sep. 2009
Figure 8.2: Some of the many widgets in Iwidgets.
to be instantiated (these in C++ are called abstract classes) but are simply present to encapsulate
common functionality by classes further down the hierarchy.
Figure 8.1(middle) shows the catalog demo that comes with the Iwidgets package. This can be used
to illustrate the individual widgets. On the left the user selects the widget that he/she is interested
in. In the right pane, there is a four tab notebook. The four tabs display: (i) the example (ii) the
example code (iii) the inheritance tree for the widget and (iv) the detailed manual page description.
This is a great learning tool for the Iwidgets package. The bottom part of Figure 8.1 shows how
to invoke the catalog demo in Windows. On linux, the path is dierent, simply change the path to
/usr/local/vtk44 yale/lib/iwidgets4.0.1/demos.
8.3 A few key Iwidgets
There are lots of mega-widgets in Iwidgets. The following are particularly useful (at least I end up using
them a fair amount). For example code for these widgets look at the catalog demo described previously.
I suggest you read this section with the catalog demo program open.
The entryeld widget: This widget combines a label and an entry eld together. It also has function-
ality for validating the input in the entry eld, such as making sure that it is an integer, a real-valued
number, or a string. See Figure 8.2A.
The optionmenu widget: This is essentially a combination of a label and a tk OptionMenu widget.
See Figure 8.2B.
65
CHAPTER 8. IWIDGETS: OBJECT ORIENTED PROGRAMMING FOR GUIS Draft Sep. 2009
The combobox widget: This is a drop-down list control that again enables the selection of an element
from a list. See Figure 8.2C.
The scrolledlistbox widget: This allows the selection of an element (or group of elements) from a
list. See Figure 8.2D.
The tabnotebook widget: This is used by the catalog demo itself. It allows for the creation of
complex graphical user interfaces using tabs in the window. See Figure 8.2E.
The scrolledhtml widget: This is a widget in which you can display HTML formatted text, it could
be used to construct a mini web-browser. See Figure 8.2F.
8.4 Object Oriented GUIs
The following example (script8-2.tcl) is an object oriented version of a previous example (script6.1.tcl).
Functionally, with the exception of the nice alignment of the entryelds, this is equivalent to script6-
1.tcl. However, the encapsulation of this graphical user interface into a class, allows for more than one
instance of the class to be created in a program in a clean manner.
Figure 8.3: A snapshot of script8-2.tcl.
The rst part of the script invokes both the Itcl and Iwidgets
extensions in the usual manner.
package require Itcl
package require Iwidgets
Key Trick: Next comes the class header. The only novel aspect of
this is the declaration of the parameters variable (which will turn out
to be a two-dimensional associative array) as private common. If
you go back to script6-2.tcl you will notice that, in that case, a global array called parameters was used to
store all variables that were attached to widgets. In the case of [Incr] Tcl based programming, we use a similar
concept. However, this array can not be explicitly a member variable of the class, other than as a static/common
variable, because linking a variable to a widget requires that it has a static address. Hence the solution is to
use a static associative array attached to the class. Since, we frequently would like to have more than one
instance of the class active in a program, the trick here is to have two-dimensional associative array of the form:
parameters(objectidentifier,variable) where objectidentier points to the individual instance of the class.
This is most easily done using the implicitly dened this variable. Note that the constructor simply takes one
argument (parent) which is the parent widget and calls CreateGUI which creates the graphical user interface.
itcl::class myGUI {
private common parameters
protected variable basewidget 0
constructor { parent } { CreateGUI $parent }
destructor { }
66
CHAPTER 8. IWIDGETS: OBJECT ORIENTED PROGRAMMING FOR GUIS Draft Sep. 2009
public method Reset { ask }
public method SetColor { widget }
public method GetName { mode }
public method CreateGUI { base }
};
The next threee methods are practically identical to their non OOP-counterparts in script6-1.tcl with two excep-
tions: (i) There is no global statement at the top of each procedure since the parameters array, as it is a class
variable, is always available (ii) as mentioned before, the parameters array is now a two-dimensional associative
array.
itcl::body myGUI::Reset { ask } {
if { $ask == 1 } {
set ok [ tk_messageBox -type yesno -message "Resetting filenames" -icon question ]
if { $ok == "no" } { return }
}
set parameters($this,readname) ""
set parameters($this,writename) ""
set parameters($this,directory) ""
}
itcl::body myGUI::SetColor { widget } {
set color [ tk_chooseColor -title "Set Background Color" -parent $basewidget ]
if { [ string length $color ] > 0 } { $widget configure -bg $color }
}
itcl::body myGUI::GetName { mode } {
set fname ""
set filetype1 [ list "Text Files" [ list .txt .tex ] ]
set filetype2 [ list "All Files" "*" ]
switch -exact $mode {
"readname" { set fname [tk_getOpenFile -title "Load" -filetypes [ list $filetype1 $filetype2 ] ] }
"writename" { set fname [tk_getSaveFile -title "Filename" -filetypes [ list $filetype1 ] ]}
"directory" { set fname [tk_chooseDirectory -title "Select Directory" ] }
}
if { [ string length $fname ] < 1 } { return }
set parameters($this,$mode) $fname
}
Finally, here is the CreateGUI method. This replaces part of the main program in script6-1.tcl.
itcl::body myGUI::CreateGUI { base } {
set basewidget $base
$this Reset 0
67
CHAPTER 8. IWIDGETS: OBJECT ORIENTED PROGRAMMING FOR GUIS Draft Sep. 2009
set labellist ""
for { set i 1 } { $i <= 3 } { incr i } {
set parent [ frame $basewidget.$i ]
pack $parent -side top -expand false -fill x
switch -exact $i {
"1" { set name "Input File:" set mode readname }
"2" { set name "Save File :" set mode writename }
"3" { set name "Directory :" set mode directory }
}
eval "iwidgets::entryfield $parent.1 -labeltext \"$name\" -width 20 \
-textvariable [ itcl::scope parameters($this,$mode) ] -relief sunken"
lappend labellist $parent.1
eval "button $parent.3 -text \"...\" -command { $this GetName $mode }"
pack $parent.3 -side right -expand false -padx 5
pack $parent.1 -side left -expand true -fill x -padx 2
}
eval "iwidgets::Labeledwidget::alignlabels $labellist"
set w [ frame $basewidget.bottom ]
pack $w -side bottom -expand true -fill x -pady 10 -padx 20
eval "button $w.1 -text Reset -command { $this Reset 1 }"
eval "button $w.2 -text Color -command { $this SetColor $basewidget }"
button $w.3 -text Exit -command { exit }
pack $w.1 $w.2 $w.3 -side left -expand true -fill x -padx 2
}
There are three elements here that need explanation.
1. Attaching a class variable to a widget: The construct -textvariable [ itcl::scope parameters(this,mode)
] takes a class associative static array parameters(this,mode) and converts it to global scope using the
itcl::scope command. The output of itcl::scope is then used to set the -textvariable eld of the widget.
This construct is needed everytime a class member variable is attached to a widget. Note also that the
variable must be static/common. In addition, since the $this variable is not dened in global scope, using
eval to pre-parse this command is necessary.
2. Using the iwidgets entryeld class. This is simply invoked using iwidgets::entryeld. Nothing complex here,
other than the fact that it takes arguments related to both the label and the entry widgets it contains.
3. Using alignlabels: Labelwidgets such as the entryeld widget can be aligned so that they look nice using
the static class procedure Labeledwidget::alignlabels. We create a list of the labels to be aligned and use
the eval command to convert this list into separate arguments.
Finally, the main program. We create and pack a frame (.gui) and then create an instance of the myGUI class
into it.
}
frame .gui
pack .gui -side top -expand true -fill both
set element [ myGUI \#auto .gui ]
68
CHAPTER 8. IWIDGETS: OBJECT ORIENTED PROGRAMMING FOR GUIS Draft Sep. 2009
8.5 Assignment 8
1. Download the .pdf le iwidgets.pdf. This contains a chapter by Michael McLennan who is the original
author of [Incr] Tcl/Iwidgets. This is the best introduction to Iwidgets anywhere.
2. Spend sometime with the catalog demo. This is another key tool in understanding Iwidgets.
3. Using script8-2.tcl as a base, modify it so that two myGUI objects are instantiated. Place these next to
one another.
4. Change the last script so that each instance of myGUI is in a separate dialog box.
5. Using script6-2.tcl as a base, rewrite it such that it becomes a class (similar to the conversion from script6-
1.tcl to script8-2.tcl).
6. Create some text and save it in HTML format. Write a short script that invokes the scrolledhtml widget
to display this text.
69
Draft Sep. 2009
Part III
The Visualization Toolkit I Using Tcl
70
Draft Sep. 2009
Chapter 9
An Introduction to the Visualization
Toolkit
The Visualization Toolkit (VTK) has become one of the de facto standard tools in modern programming for
image analysis. It has rich functionality for both image/surface processing and visualization. Its design pioneered
the use of the combination of scripting and compiled languages in a single piece of software through the use
of wrapping techniques for making available the underlying C++ classes in VTK as new commands in scripting
languages such as Tcl and Python. The keys to successfully using VTK are (i) understanding the structure of the
object-oriented hierarchy and (ii) understanding its pipeline architecture. This document and those following it
will focus on VTK version 5.2.
9.1 Introduction to 3D Graphics
Most students rst experience to image display consists of simply displaying a two-dimensional image slice (most
likely in Matlab). While there are some issues, such as the relative orientation (row/column vs column/row)
and the position of the origin, it is for the most part a fairly straightforward and intuitive procedure. The same
applies to rendering curves and point landmarks on the image. The key behind this simplicity is the fact that,
unsurprisingly, two-dimensional computer monitors are well suite to displaying two-dimensional content!
The move to 3D graphics and visualization requires, in some way, getting around the fact that for the most part
our display units are two-dimensional. The most common illusion used in 3D graphics and the one that
is employed by the Visualization Toolkit (VTK) is that the program rst generates a 3D world consisting of
three-dimensional entities actors/props in VTK parlance which have various appearance properties. The world
is illuminated by a set of lights and the user looks at the world through the eyes of a virtual camera. The image
that gets displayed on the computer monitor is precisely the output of such a virtual camera. To recap, a world
consists of (i) Actors, (ii) Lights and (iii) a camera. The exact output naturally depends critically, in addition to
the Actors themselves, on the position and orientation of both the lights and the cameras. If a camera is looking
away from an actor, that actor will not be visible in the displayed image.
9.2 An Introduction to VTK
VTK is a large, complicated, powerful but often surprisingly easy to use toolkit for 3D image processing and
visualization. It is an object-oriented library. Many operations in VTK are performed using a pipeline architecture
where multiple elements are attached together to perform a complex task. A typical pipeline takes the form shown
in Figure 9.1. This is broken into two parts. The rst part (shown in Figure 9.1(left) consists of:
71
CHAPTER 9. AN INTRODUCTION TO THE VISUALIZATION TOOLKIT Draft Sep. 2009
Figure 9.1: The VTK Pipeline Parts 1 and 2.
1. Sources these classes produce data. For example, vtkJPEGReader reads a jpeg image from a le and
generates an image output.
2. Filters these operate on some data to produce a modied version. For example, vtkImageGaussianSmooth
acts on an image to perform Gaussian smoothing and to produce a new smoothed image.
3. Mappers these dene the interface between data (e.g. images) and graphics primitives or software
rendering techniques. A special kind of mapper like class are the writers which output the data to les
(e.g. vtkJPEGWriter). Multiple mappers may share the same input, but render it in dierent ways.
The second part of the pipeline consists of the elements that make up the virtual 3D world, namely:
1. Props/Actors these take as input the output of a mapper and know how to generate the visible repre-
sentation of data. The type of rendering produced is governed by an auxiliary data structure known as a
property (e.g. color, showing a surface as a wire-frame vs a full surface etc.). Props take as their input the
output of a mapper. Mappers should not be shared among props.
Volumes these are special kinds of props that are used to display volume rendered images.
2. Renderer Renderers are the classes that generate a 2D image from a 3D scene. They have attached
actors as well as lights and cameras. More formally, from the man page, renderer is an object that
controls the rendering process for objects. Rendering is the process of converting geometry, a specication
for lights, and a camera view into an image. vtkRenderer also performs coordinate transformation between
world coordinates, view coordinates (the computer graphics rendering coordinate system), and display
coordinates (the actual screen coordinates on the display device). Certain advanced rendering features
such as two-sided lighting can also be controlled.
3. Render Window the Render Window is the piece of screen real estate in which the virtual camera image
is displayed. An important auxiliary item attached to a render window is an interactor which handles
mouse/keyboard input to the window.
72
CHAPTER 9. AN INTRODUCTION TO THE VISUALIZATION TOOLKIT Draft Sep. 2009
9.3 Data Objects
VTK stores key items such as images, surfaces and meshes in classes derived from vtkDataObject. Data objects
consist of combinations of points (nodes) which dene the location of the data and cells (elements) which dene
the topology. For example a 20x20 2D image can be thought o consisting of 400 regularly spaced points
connected into squares. Similarly, triangulated surfaces consist of a set of points and a set of interconnecting
triangles.
Consider the examples shown in Figure 9.2. The rst is a polygonal curve, which consists of ve points whose
locations dene the position of the curve. The topology (cells) are the ve lines connecting the points. In this
case they would take the form (0, 1), (1, 2), (2, 3), (3, 4), (4, 0) where (0, 1) is the line connecting points 0 and 1.
The second example is that of a triangulated surface. This is similar to the curve, with the replacement of line
cells with triangle cells.
The case of the image is more interesting in that, in this case, both the topology and location of the individual
points are implicitly dened once the position of the corner of the image (the origin) and the dimensions of the
individual element (pixel or voxel spacing) are dened. Hence in the denition of an image (vtkImageData), the
only things that need to be dened in terms of geometry and topology are the dimensions (number of points),
the origin (position of the corner node) and spacing (the distance between the nodes), the rest are automatically
dened from these.
For many data objects, especially images there is an additional critical element the data associated with each
point (or cell in some cases). Such data is used to store information such as the image intensity etc. Two
data structures attached to each DataObject are used to store such information. These are the PointData and
CellData structures which contain in turn data arrays derived from vtkDataArray. In the case of surfaces, additional
information may include surface normals in addition to surface color etc.
9.4 Data Arrays
Ultimately most data, in VTK, is stored in Data Arrays. The parent class vtkDataArray (see hierarchy in gure
9.3) provides a generic interface to all arrays which are specialized forms of it for dierent data types. The
common data objects simply assume that their data is stored in a vtkDataArray and rely only on the generic
interface. The actual data is stored in derived classes of vtkDataArray (e.g. vtkFloatArray) which are specialized
for the specic data type. In this way, one can have image data structures that are independent of the data type,
which is dynamically dened during the runtime of the program.
A vtkDataArray is essentially a two dimensional array, the rst dimension (the row dimension) is the Tuple dimen-
sion and second dimension (the column dimension) is the Component dimension. In the case of vtkDataObjects,
the contents of each tuple (row) are used to specify the data associated with each point (or cell in CellData).
For example a 16 16 color image (RGB) would have its intensity information in a vtkDataArray consisting of
256 tuples, with 3 components per tuple. In this way the color information for the fourth voxel (in raster scan
order) is in the fourth tuple of the array (indexing begins at zero in VTK, hence this will have index 3) with
the red color stored in component 0, the green in component 1 and the blue in component 2. In fact, the very
name component originated in the need to store color images, vtkDataArrays prior to version 4 could only have
a maximum of 4 components. This has been since relaxed, and we can use the components to store the time
component of a 4D Image (in the case of fMRI, this can be a number greater than 100).
9.5 Concluding Remarks
I have tried in this session to focus on the fundamental components of VTK, understanding these is key to learning
how to use the toolkit. In the next session we will begin to use VTK for surface and image rendering and manip-
ulation. In lieu of an assignment, take a look at the VTK man pages at http://noodle.med.yale.edu/vtk/.
73
CHAPTER 9. AN INTRODUCTION TO THE VISUALIZATION TOOLKIT Draft Sep. 2009
Figure 9.2: Geometry and Topology of Common Data Objects.
Figure 9.3: The Data Array Hierarchy.
74
Draft Sep. 2009
Chapter 10
Curves and Surfaces in VTK
The Visualization Toolkit (VTK) has become one of the de facto standard tools in modern programming for image
analysis. It has rich functionality for both image/surface processing and visualization. In this session we describe
creating and visualizing surfaces and space curves. The examples in this handout require VTK 5.2 or later,
they will not work with the VTK 4.4 that is distributed with BioImage Suite 2.6. You can get a copy
of the BioImage Suite 3 base distribution which includes VTK 5.2 at https://boreas.med.yale.edu/base.
Simply unzip/untar these into /usr/local (or C:/yale).
10.1 Introduction
The standard introduction to VTK tends to have the following pattern: (i) create a simple object, e.g. a cone
1
(ii) display it and (iii) manipulate the display in some way. While this approach has some merit, it is more directed
towards users who are simply interested in visualizing existing data (i.e. read in from a le) as opposed to more
common tasks in image analysis in which we need to somehow create this data perhaps as the output of some
segmentation algorithm.
While visualization remains a critical task, in image analysis, we are often less interested in cosmetic eects
such as lights, perspective etc. and more at getting a handle of how a specic algorithm is performing. Hence for
most of these lectures, we will focus on more boring visualization tasks. In any event, once one is comfortable
with VTK, Hollywood like eects of high quality are not that hard to generate.
10.2 Tcl and VTK Objects
A common confusion when programming in VTK using the Tcl scripting language is the concept of object name.
A new object (e.g. an image) is generated using the syntax (remember to use the vtk interpreter from here on):
vtkImageData myimage
where vtkImageData is the class type and myimage is the object name. Later code can modify this object
using its name explicitly, in standard OOP usage, e.g. the example below changes myimage to have dimensions
1
In the repository, I copy the les from one of the tutorials included with VTK, this is the Cone series and consists of
six scripts, Cone.tcl,Cone2.tcl,. . .,Cone6.tcl. These are worth looking through.
75
CHAPTER 10. CURVES AND SURFACES IN VTK Draft Sep. 2009
100 100, sets its storage type to oat and allocates memory.
myimage SetDimensions 100 100 1
myimage SetScalarTypeToFloat
myimage AllocateScalars
The object name myimage must be absolutely unique in the whole program. If an object of the name myimage
exists, then any attempt to create a second object with the same name will fail. The object name is, in eect,
a super-global variable
2
and while explicitly naming objects in this way works for small scripts, it is a recipe for
disaster in large programs, especially when leveraging code written by other people.
The solution to this problem is to have a function that generates unique names (like the #auto construct in [ Incr
] Tcl) and using this to set the object name. Such a function could be similar to that used to generate unique
SSNs in the IrsInfant class (session 7). VTK accomodates this using the static New function.
Using this we can rewrite the script as:
set myimage [ vtkImageData New ]
$myimage SetDimensions 100 100 1
$myimage SetScalarTypeToFloat
$myimage AllocateScalars
In this case, myimage is an ordinary variable which contains the object name (probably something like vtkObj1).
The usual scoping rules apply etc. You can now access the object the sane way, using $myimage.
10.3 Data Arrays
All data in VTK is stored ultimately in one of the many derived classes of vtkDataArray. vtkDataArray is an
abstract superclass for classes representing arrays of vectors called tuples (or numbers treated as vectors of
length 1). Each tuple consists of a set of numbers or components. Derived Classes of vtkDataArray include
vtkUnsignedCharArray, vtkShortArray, vtkFloatArray, vtkDoubleArray etc. An abstract class is one which is never
instantiated itself, it is rather a parent class which captures the common interface for a set of derived classes.
vtkDataArray-derived classes can function either as a dynamic array (lower performance) or a xed length arrays.
For example, in the case of vtkImageData, the intensities are stored in a vtkDataArray having dimensions equal
to the number of voxels and vector length typically equal to 1 (3 for color images). In the case of multiframe
data such as fMRI or cardiac data, the number of components is equal to the number of frames.
Fixed Length Array: The following script (script10-1.tcl) creates a xed length array and manipulates its
contents.
wm withdraw .
2
In C++, this would be the equivalent of accessing variables directly using their memory location. While it can be
done, it is asking for trouble.
76
CHAPTER 10. CURVES AND SURFACES IN VTK Draft Sep. 2009
# Create the Array
set arr [ vtkFloatArray New]
$arr SetNumberOfComponents 2
$arr SetNumberOfTuples 12
# Set some defaults -- always a good idea
$arr FillComponent 0 0.0
$arr FillComponent 1 10.0
# Set some values
$arr SetComponent 10 0 3.0
$arr SetTuple2 11 9.0 2.0
$arr SetComponent 4 4 -2.1
# Print some info and then the array
puts stdout "The array has [ $arr GetNumberOfTuples ] tuples,
[ $arr GetNumberOfComponents ] components"
puts stdout "Here are its contents"
set nt [ $arr GetNumberOfTuples ]
set nc [ $arr GetNumberOfComponents ]
for { set i 0 } { $i < $nt } { incr i } {
puts -nonewline stdout "Tuple $i : ("
for { set j 0 } { $j < $nc } { incr j } {
puts -nonewline stdout "\t [ $arr GetComponent $i $j ]"
}
puts stdout "\t)"
}
# The exit command is explicitly needed in a vtk script
exit
We can access the elements in this array by using the SetComponent and GetComponent methods. All indices
start at 0. The value of a whole component can be set using the FillComponent method as shown above. The
SetTuple commands (SetTuple2, SetTuple3 etc.) can be used to set whole vectors and the corresponding GetTuple
commands can be used to get whole vectors as lists.
Dynamic Arrays: These are useful when one does not know ultimately how much data is coming. In this
case we only specify the type of the array and the number of components. The following script (script10-2.tcl)
illustrates this concept:
wm withdraw .
set arr [ vtkFloatArray New]
$arr SetNumberOfComponents 1
$arr InsertNextTuple1 5
$arr InsertNextTuple1 10
set b [ $arr GetComponent 1 0 ]
puts stderr "b=$b"
There are additional functions (InsertNextTuple2, InsertNextTuple3,4,9) for multi-component arrays.
77
CHAPTER 10. CURVES AND SURFACES IN VTK Draft Sep. 2009
Figure 10.1: Constituent parts of a Data Object. A data object consists of a set of points which dene
its geometry (these are implicitly dened in the case of images), and a set of cells (e.g. lines, triangles,
etc.) which dene the topology (again this is implicitly dene in the case of images). Both the points
and the cells can have associated attributes (e.g. the intensities in an image are stored as point-data
associated with each point).
Data Types: The following constants are often dened, as shorthand for the numerical value that VTK uses
to identify each data-type.
set VTK VOID 0 set VTK SHORT 4 set VTK LONG 8
set VTK BIT 1 set VTK UNSIGNED SHORT 5 set VTK UNSIGNED LONG 9
set VTK CHAR 2 set VTK INT 6 set VTK FLOAT 10
set VTK UNSIGNED CHAR 3 set VTK UNSIGNED INT 7 set VTK DOUBLE 11
In C++ one can use the static member function vtkDataArray::CreateDataArray(type) to create an array of type
to be specied at run-time. In Tcl there is a less elegant solution as vtkDataArray can not be accessed directly.
Here one creates a specic empty array (e.g. vtkFloatArray) and then uses this to create another array, e.g. ,
set emptyshell [ vtkFloatArray New]
set shortarray [ $emptyshell CreateDataArray VTK_SHORT ]
10.4 Creating Curves and Surfaces
In VTK, we can create curves and surfaces using the vtkPolyData class, which is a derived class of vtkDataObject
for storing polygonal data. The general structure of a vtkDataObject (see last sessions notes for more details) is
described in gure 10.1.
vtkPolyData is a complex class which has many members. The key ones are:
Points of type vtkPoints represents the geometry of the surface (i.e. the points)
78
CHAPTER 10. CURVES AND SURFACES IN VTK Draft Sep. 2009
Polys of type vtkCellArray represents part of the topology of the surface (i.e. the polygons)
Lines of type vtkCellArray represents another part of the topology (the elements which are simple lines as
opposed to faces).
PointData of type vtkPointData represents data associated with the points (e.g. normals, colors etc)
CellData of type vtkCellData represents data associated with the points (e.g. again normals, colors etc)
Creating a Polygonal Curve: A polygonal curve consists of a set of points and the lines that connect these.
First some basic preliminaries (script10-3.tcl):
wm withdraw .
set pi 3.1415
Next we create the points. These are stored in a vtkPoints data structure which is very similar to a DataArray
(it contains a data array and adds convenience methods such as GetPoint/SetPoint for manipulating this array).
Our curve is a planar circle consisting of eight segments.
# Creating a circle with 8 segments
# First the points
set pts [ vtkPoints New]
$pts SetNumberOfPoints 8
for { set i 0 } { $i <= 7 } { incr i } {
set rad [ expr 2.0*$i*$pi/8.0 ]
set x [ expr 10.0* sin($rad) ]
set y [ expr 10.0* cos($rad) ]
$pts SetPoint $i $x $y 0.0
puts stdout "Point $i = [ $pts GetPoint $i ]"
}
From the geometry, to the topology. The circle has eight line segments which need to be specied. These are
stored in a vtkCellArray data structure, which is a very powerful (=complex) object. This uses dynamic storage
and the second command Allocate 10 5, creates an initial memory allocation for 10 cells and instructs the array
to allocate any additional memory in increments of 5. Inserting a Cell in Tcl is probably more complex than C++
given the lack of pointers. First we specify the size of the cell (InsertNextCell 2) and then we insert the individual
point indices (InsertCellPoint). The point indices will link to points in the Points array (pts) above, with 0 being
the rst point. Since this is a closed curve, the last line segment consists of points 7 and 0.
# Next the line segments
set lines [ vtkCellArray New]
$lines Allocate 10 5
for { set i 0 } { $i <= 7 } { incr i } {
set p1 $i
set p2 [ expr $i + 1 ]
if { $p2 > 7 } { set p2 0 }
$lines InsertNextCell 2
$lines InsertCellPoint $p1;
79
CHAPTER 10. CURVES AND SURFACES IN VTK Draft Sep. 2009
$lines InsertCellPoint $p2
puts stdout "Set Line Segment $i = ($p1, $p2)"
}
Once we create the points and the lines, we can place them in a surface structure. We create this using the
vtkPolyData command and attach the points and the lines to it.
# Create the curve object and set the points and lines
set curve [ vtkPolyData New]
$curve SetPoints $pts
$curve SetLines $lines
The pts and the lines structures can now be deleted. VTK uses reference counted memory allocation (we will talk
about this more later), and by setting the pts and lines as parts of the surface structure their reference count is
increased, so issuing the Delete command does not actually delete them!
}
# Since VTK uses reference counting delete pts and lines
# as these are counted inside curve
$pts Delete
$lines Delete
Finally we can output the curve to a le, using the vtkPolyDataWriter class. All vtk writer classes have similar
structure. The SetInput method is used to specify the data-structure to be saved, the SetFileName method
species the lename and the Write method executes the operation.
# Save the curve to a file
set writer [ vtkPolyDataWriter New]
$writer SetInput $curve
$writer SetFileName curve.vtk
$writer Write
# Clean Up all objects and exit
$writer Delete
$curve Delete
exit
Creating a Surface: Our simple surface will be a cone which will consist of a circle and an apex. This is
accomplished by modifying the previous script to yield a new script (script10-4.tcl) as follows. Each triangle
consists of a line segment on the curve as a base with the apex as the third point. The key changes from the
previous le are:
80
CHAPTER 10. CURVES AND SURFACES IN VTK Draft Sep. 2009
Figure 10.2: The VTK Pipeline Parts 1 and 2.
...
set pts [ vtkPoints New]
$pts SetNumberOfPoints 9
for { set i 0 } { $i <= 7 } { incr i } {
...
}
# Set apex point as point 8
$pts SetPoint 8 0.0 0.0 10.0
# Next the line segments
set triangles [ vtkCellArray New]
$triangles Allocate 10 5
for { set i 0 } { $i <= 7 } { incr i } {
set p1 $i
set p2 [ expr $i + 1 ]
if { $p2 > 7 } { set p2 0 }
$triangles InsertNextCell 3
$triangles InsertCellPoint $p1;
$triangles InsertCellPoint $p2
$triangles InsertCellPoint 8
puts stdout "Set Line Segment $i = ($p1, $p2)"
}
set surface [ vtkPolyData New]
$surface SetPoints $pts
$surface SetPolys $triangles
The key change is that we now have triangles as opposed to lines which are attached to the vtkPolyData object
surface using the SetPolys method.
81
CHAPTER 10. CURVES AND SURFACES IN VTK Draft Sep. 2009
10.5 Displaying Polygonal Data
To display the curve and/or surface constructed previously we need to create a pipeline of the form shown in gure
10.2. The following script (script10-5.tcl) is similar to script10-4.tcl up to the point where vtkPolyDataWriter is
created. At that point in the code we replace the le output operation with the display pipeline as shown below:
The rst step is to create the mapper (mappers dene the interface between data and graphics primitives)
# Create the
mapper set map [ vtkPolyDataMapper New]
$map SetInput $surface
The next step is to create the actor which appears in the scene:
#Create the actor and set it to display wireframes
set actor [ vtkActor New]
$actor SetMapper $map
$map Delete
Next we tweak the display properties of the actor, using its property element. We set the color to white (1,1,1)
and the display to wireframe.
set property [ $actor GetProperty ]
$property SetColor 1 1 1 ; $property SetAmbient 1.0
$property SetDiffuse 0.0; $property SetSpecular 0.0
$property SetRepresentationToWireframe
The next step is to create the renderer and add the actor to it. While we are at it we switch the renderer to
parallel projection as opposed to perspective which is the default.
set ren [ vtkRenderer New]
$ren AddActor $actor
[ $ren GetActiveCamera ] ParallelProjectionOn
We next create a render window, and attach the renderer to it.
set renWin [ vtkRenderWindow New]
$renWin AddRenderer $ren
$renWin SetSize 300 300
82
CHAPTER 10. CURVES AND SURFACES IN VTK Draft Sep. 2009
Figure 10.3: A screenshot of the script10-5.tcl.
Finally we create an interactor to handle mouse/keyboard input and reset the camera
set iren [ vtkRenderWindowInteractor New]
$iren SetRenderWindow $renWin
$iren Initialize
$iren AddObserver SetExitMethod { exit }
$ren ResetCamera
Pressing the e key invokes the exit command of the renderer which exits the program.
10.6 Reference Counting Memory Allocation/Deallocation
When a new object is instantiated, memory is allocated e.g. set pts [ vtkPoints New]. As part of the
construction of the new object, a counter called the reference count of the object is set to 1.
Each time the object is used (contained) by another object e.g. $curve SetPoints $pts, the reference number
of the object $pts is increased by 1.
When the object $curve is deleted it will decrement the reference count of $pts by 1. If at this point in time
$pts has a reference count of 0 it will get removed from memory, if it not it will stick around.
Calling the delete method of an object, e.g. $pts Delete, also decreases the number of references by 1. When
the number of references is equal to zero, the memory is then released. Hence calling delete does not always
result in (immediately) deleting the object.
If an object (e.g. $pts) is not meant to have an independent existence from its container object (e.g. $curve in
this case), then once assigned to the container its delete method should be called to decrease its reference count
back to 1. That way, when the container object is deleted, the dependent object is also deleted.
83
CHAPTER 10. CURVES AND SURFACES IN VTK Draft Sep. 2009
Assignment
1. Review the contents of this and the previous handout.
2. Play with the Cone scripts in the repository.
3. Using script10-3.tcl as a base, change the code to create a helix, i.e. a curve where the z-coordinate
increase with each point as opposed to a constant 0.0. Save the output in a le helix.vtk
4. Using the last assignment as a base and the display code from script10-5.tcl (or your own version) write a
script that creates and displays a helix.
5. Modify the last assignment to display both a helix and a cone. Hint: you will need two actors.
84
Draft Sep. 2009
Chapter 11
Images in VTK
Finally, we come to the point we talk about images per se for the rst time. In particular, we set the foundations
for creating and manipulating images. This handout aims to set the stage for hands-on examples that will be
described in the next chapter.
11.1 Introduction
In their early steps in programming for image processing/analysis, most people tend to regard and store an image as
essentially a matrix of numbers. The only complication, at this stage, is the relationship between the row/column
indices of the matrix and the x- and y-axes of the image, respectively.
1
This type of radical simplicity disappears, unfortunately, once we have issues with mapping images. Examples
include, image registration applications when we are looking to estimate a transformation between two images,
statistical shape model building etc.
In general, in medical imaging, we need to keep track of not only the image intensities but additional attributes
such as voxel dimensions (voxels need not be isotropic), image orientation (the relationship of the x-,y- and z-axis
orientation to the human body see Figure 11.1), the position of the voxels etc.
In VTK, images are stored in a complex class vtkImageData. This shares much of the general structure of other
VTK data objects, but because of its regularity much of its geometry and topology can be specied using a very
small number of parameters. It is also worth pointing out that in earlier versions of VTK (earlier than 4.0), the
primary class for storing images was called vtkStructuredPoints which is now, for the most part, an empty
subclass of vtkImageData. If you see vtkStructuredPoints anywhere in the VTK documentation, you can safely
substitute vtkImageData at this point.
11.2 The vtkImageData Class
In VTK, images are stored using the class vtkImageData. VTK, implicitly at least, treats an image as a function
which takes values equal to those specied at the center of each voxel and interpolates in-between. The standard
interpolation scheme used is tri-linear interpolation. In some cases (vtkImageReslice) other interpolation schemes
are available. In addition to the image intensities themselves there are three key variables (arrays) in vtkImageData,
which are used to dene an image, namely:
1
Often a problem in MATLAB, when image matrices need to be transposed before display!
85
CHAPTER 11. IMAGES IN VTK Draft Sep. 2009
Figure 11.1: Left: Standard
Image Acquisition Orientations,
Axial or Transverse, Coronal
and Sagittal. The arrows indi-
cate the z-axis direction (which
may be inverted depending on
the acquisition protocol), the x-
axis and the y-axis are perpen-
dicular to this. Right: Ax-
ial, coronal and sagittal slices.
Oblique acquisitions are also
sometimes used, in which the z-
axis is rotated, or obliqued away
from one of the standard acqui-
sitions, e.g. coronal oblique.
Figure 11.2: Axes orientation
for a typical axial acquisition.
Figure 11.3: Voxel indices in a
144 image. Voxels are ordered
in raster-scan order.
Figure 11.4: Because of
the regularity of the underly-
ing geometry/topology, the only
aspect of the complex vtk-
DataObject data-structure that
needs to be explicitly dened in
vtkImageData is the PointData,
where we store the image inten-
sity. The geometry and topol-
ogy can be specied simply by
specifying the dimensions of the
grid, the position of the corner,
and the size of the voxels.
86
CHAPTER 11. IMAGES IN VTK Draft Sep. 2009
int Dimensions[3] the number of voxels in the x,y and z directions respectively. In the case of the
image shown in gure 11.3, this array takes values 14,4,1.
double Origin[3] the position in 3D space of the centroid of the rst voxel (with index 0, as dened
in gure 11.3.) Often this is specied with respect to the center of the scanner, although it is often simply
set to 0,0,0.
double Spacing[3] the physical size of each voxel e.g. 1 1 1.2mm. While sometimes images are
isotropic, this is not always the case, so one needs to be careful.
Once these variables are set, they completely dene the topology and the geometry of the image. Hence (as shown
in Figure 11.4), in a vtkImageData we do not specify the positions of the points or the topology explicitly, these
are simply implicitly dened by these three variables. The key element in the specication of a vtkImageData is
a Data Array that is contained in the PointData attributes.
This Data Array has as many tuples as the image has voxels, and each tuple has a number of components
corresponding to the components (in color images) or, frames in 4D images. The tuples are ordered in raster-scan
order as illustrated in Figure 11.3.
11.3 Creating and Manipulating Images
11.3.1 Creating an Image
Creating an image involves the following:
1. Creating the instance of vtkImageData
2. Specifying the Dimensions
3. Optionally specifying the Spacing if not (1.0,1.0,1.0).
4. Optionally specifying the Origin if not (0.0,0.0,0.0).
5. Optionally specifying the number of components if not 1.
6. Specifying the image type, e.g. short, oat etc.
7. Allocating Memory
In Tcl (using the vtk interpreter!) this can be accomplished using the following commands:
package require newname
set img [ vtkImageData New ]
$img SetDimensions 10 8 4
$img SetSpacing 1.0 1.0 1.0
$img SetOrigin 0.0 0.0 0.0
$img SetNumberOfScalarComponents 1
$img SetScalarTypeToFloat # or Short,Int,Double,Char,UnsignedChar etc.
$img AllocateScalars
11.3.2 Manipulating Image Intensities
The Slow Method: The easiest method for accessing and modifying image data is to use the pair of methods:
double GetScalarComponentAsDouble (int x, int y, int z, int component)
void SetScalarComponentFromDouble (int x, int y, int z, int component,double v)
87
CHAPTER 11. IMAGES IN VTK Draft Sep. 2009
In the example above to get the intensity at voxel (5,3,1) and component 0 we can call:
set v [ $img GetScalarComponentAsDouble 5 3 1 0 ]
To set the value at voxel (4,6,2) and component 0 to 15.0 we can call:
$img SetScalarComponentFromDouble 4 6 2 0 15.0
Note: Obviously if the underlying image is short, setting an intensity value to, for example, 15.2 will result in the
intensity being set to 15! While the intensity is specied using a double variable (i.e. one that stores numbers
in double precision oating type) this is appropriately (or inappropriately) converted to the specic type of the
image prior to it being set!
The Somewhat Faster Method: In this method, we rst get a pointer to the underlying data array in
which the data is stored and manipulate this array directly. This is accomplished by:
# The two-step method
set pdata [ $img GetPointData ]
set data [ $pdata GetScalars ]
# The direct-method
set data [ [ $img GetPointData ] GetScalars ]
Once the pointer to the underlying array is obtained, we can manipulate intensities using the GetCompo-
nent/SetComponent methods of the data array, e.g.
set v [ $data GetComponent 115 0 ]
$data SetComponent 224 0 15.0
where 115 is raster-index of voxel (5,3,1) and 224 is the raster-index of voxel (4,6,2) from before.
Direct Access, C++ only: The fastest method involves realizing that things like two and three-dimensional
arrays are simply ctions of our imagination (or really convenience devices) and that all array data is ultimately
stored in one-dimensional arrays (since memory is inherently one dimensional). To operate on an image at
maximum speed one needs to, in C++, access and manipulate the underlying data pointer.
88
CHAPTER 11. IMAGES IN VTK Draft Sep. 2009
11.4 Additional VTK Classes for Image Manipulation
11.4.1 Image Input and Output
VTK supports by default a number of standard image le formats for read/write:
Binary vtkImageReader
JPEG vtkJPEGReader, vtkJPEGWriter
PNG vtkPNGReader, vtkPNGWriter (ppm,pgm)
TIFF vtkTIFFReader, vtkTIFFWriter
BMP vtkBMPReader, vtkBMPWriter
DICOM vtkDICOMImageReader (not complete)
There are BioImage Suite extensions for reading Analyze, Signa LX/SPR, Prism (SPECT) etc.
11.4.2 Image to Image Filters
VTK has a rich set of lters for manipulating images. These are derived from vtkImageToImageFilter. Some
common examples are:
Smoothing vtkImageGaussianSmooth, vtkImageMedian3D
Computing Gradients/Laplacians vtkImageGradient, vtkImageLaplacian
Fourier Operations vtkImageFFT,vtkImageRFT
Resampling/Reslicing vtkImageResample, vtkImageReslice (vtkImageReslice on its own is reason enough
to learn VTK, it implements enough operations that would take more than a year to code from scratch!)
Flipping, Permuting vtkImageFlip,vtkImagePermute
11.4.3 A Simple Example
The following simple example (script11-1.tcl) generates an 20 20 image which is a smoothed x. (Some lines
are omitted for clarity.)
First we load the package newname and hide the default window since this is a command-line script:
lappend auto_path [ file dirname [ info script ]]; package require newname
wm withdraw .
We then create the image and ll it with zeros:
set img [ vtkImageData New ]
$img SetDimensions 20 20 1
$img SetNumberOfScalarComponents 1
$img SetScalarTypeToUnsignedChar
$img AllocateScalars
set data [ [ $img GetPointData ] GetScalars ]
$data FillComponent 0 0
89
CHAPTER 11. IMAGES IN VTK Draft Sep. 2009
Figure 11.5: The Output JPEG File.
Next we draw two diagonal lines (i == j) and (i +j = 19):
for { set i 0 } { $i <= 19 } { incr i } {
for { set j 0 } { $j <= 19 } { incr j } {
if { $i == $j || [ expr $i + $j ] == 19 } {
# $i, $j, 0 are the x, y, z-coordinates for the point.
# The second 0 in the following line specifies the first component of the image.
$img SetScalarComponentFromDouble $i $j 0 0 200
}
}
}
The smoothing lter is used to slightly blur the image:
set smooth [ vtkImageGaussianSmooth New ]
$smooth SetInput $img
$smooth SetStandardDeviations 1.0 1.0 0.0
$smooth SetDimensionality 2
$smooth Update
The output is saved in a jpeg le:
set writer [ vtkJPEGWriter New ]
$writer SetFileDimensionality 2
$writer SetFileName "smoothedx.jpeg"
$writer SetInput [ $smooth GetOutput]
$writer Write
We clean up the objects and exit.
90
CHAPTER 11. IMAGES IN VTK Draft Sep. 2009
$writer Delete;$smooth Delete;$img Delete
exit
11.5 Visualizing Images
In displaying three-dimensional images, we face an additional complexity when we compare this task to the task
of displaying polygonal surfaces. Surface display is predicated on the fact that surfaces are thin structures with
no inside. Images are, on the other hand, thick structures, where the inside is often more important than
the bounding voxels. Hence direct display of a three-dimensional image is problematic, it will simply look like a
cube.
The most common ways to render 3D images are:
Orthogonal Slices e.g. example in Figure 11.1 (right).
Oblique Slices slices obtained by resampling the image along a plane that is not aligned with one of the
coordinate axes.
Volume Rendering a 3d-rendering of the whole image using a combination of intensity mapping and
transparency mapping tricks.
While image slices can be displayed as surfaces with the intensity of each voxel mapped to patches of a surface,
they are most commonly displayed as textures mapped on rectangles of appropriate dimensions. As stated on
the Wikipedia page (paraphrasing) texture mapping is a method of adding detail, surface texture, or color to a
computer-generated graphic or 3D model. A texture map is applied (mapped) to the surface of a shape. This
process is akin to applying gift wrapping paper to a plain white box. This allows the realization of detail that
would take very many additional polygons to realize otherwise. This kind of coloration is the most common
application of texture mapping.
As an aside, the underlying rendering engine used in VTK, OpenGL, renders all textures in dimensions that are
powers of two. The images are interpolated before display, hence some (small) loss of sharpness takes place (only
visible in small images) E.g. an 100 50 image will be resampled to 128 64 before display. In cases of large
images (especially volume rendering) this is worth keeping in mind. A 257 257 image has the same eective
dimensions as a 512 512 image as far as display-performance is concerned. In such cases, small amounts of
judicious cropping can result in substantial improvements in rendering time.
We will discuss image rendering in more detail in the next chapter.
Assignment
Read the Handout!
Read the vtkImageData Man page http://noodle.med.yale.edu/vtk/classvtkImageData.html to
get fuller sense of the functionality in vtkImageData.
Modify, in three steps, script11-1.tcl to :
1. Change the code to generate a square, rather than an x shape.
2. Replace the Smoothing lter with a GradientMagnitude lter (vtkImageGradientMagnitude).
3. Replace the JPEG Writer with a PNM Writer.
A more complete image-manipulation assignment will be given out as part of the next chapter.
91
Draft Sep. 2009
Chapter 12
Displaying Images in VTK
Following a brief discussion on colormaps, we covers three dierent methods for displaying images in VTK, (i)
using the dedicated vtkImageViewer/vtkTkImageViewerWidget classes, (ii) displaying slices using texture mapping
on planes and (iii) volume rendering.
12.1 Introduction
There a number of ways to display images depending on the need to display additional elements in conjuction
with the images. In general, images are best displayed in a 3D rendering environment which enables polygonal
objects to be jointly displayed with them. This used to be the holy grail of medical image visualization when the
rst machines with 3D rendering capabilities appeared about 10-15 years ago. This enables the visual inspection
of, for example, segmentation results, by co-visualizing the underlying image and the surface of the segmented
structure.
There are three basic ways of displaying images in VTK. The rst, is to use the dedicated vtkImageViewer class
which renders images using bitmaps and is only really suitable for 2D image display with little additional elements.
The second method is to display selected slices through the 3D image as textures mapped on planes in 3D space.
The nal method is volume rendering which attempts to directly visualize the whole 3D volume using a set of
transparency tricks.
I have learned, the hard way, that it is dangerous to modify the original image for display purposes, in any way.
Visualization techniques, such as camera manipulation and zooming, can be used to ip and enlarge images.
Appropriate color mapping schemes can be used to change the image appearance to enhance the contrast. Hence
it is rarely the case that the original image needs to be modied.
My ve rules of image display are as follows:
1. Never modify the original image
2. If the image needs to be displayed in a dierent orientation, move the camera never modify the original
image.
3. If the image needs to be contrast-adjusted, use a colormap, never modify the original image
4. If an image needs to be magnied, use the zoom control on the camera, never modify the original image.
5. If in doubt, see rule 1.
92
CHAPTER 12. DISPLAYING IMAGES IN VTK Draft Sep. 2009
Figure 12.1: The Window/Level Colormap.
12.2 Colormaps
Colormaps (or lookup tables) are essentially functions that map image intensity to display color. The most
common colormap (which is often implicitly used) simply maps the smallest value in the image to black and the
highest value to white. For example, in the case of an image with range 0:255, 0 is mapped to black, 255 to white
and everything in between to progressively lighter shades of gray. The most common medical image colormap is
the so called Level/Window colormap. This colormap is dened by the variables, the level l and the window size
w. The mapping x y, where x is the input intensity and y the output color (from black = 0 to white = 1) is
then specied as:
y =
_
_
_
0 ifx l
w
2
1 ifx l +
w
2
x(l
w
2
)
w
otherwise
Colormaps in VTK are explicitly dened using an instance of the vtkLookupTable class. This denes the colormap
as a lookup table, consisting of N rows, each row having 4 elements (RGBA=Red,Green,Blue,Alpha). In addition,
we dene the table range L : H. The mapping is then performed in two steps: First given x we compute i the
index into the lookup table. Then. the output array y is simply row i of the lookuptable.
i =
_
_
_
0 ifx L
N 1 ifx H
xL
HL
(N 1) otherwise
The lookup table setup can be used to construct very complex Colormaps. In code this takes the form:
set cmap [ vtkLookupTable New ]
$cmap SetNumberOfColors 256
$cmap SetTableRange 0 255
for { set i 0 } { $i < 256 } { incr i } {
set v [ expr $i /255.0 ]
$cmap SetTableValue $i $v $v $v 1.0
}
See the man page of vtkLookupTable for more details.
93
CHAPTER 12. DISPLAYING IMAGES IN VTK Draft Sep. 2009
12.3 An Aside Using the Dedicated Image Viewer
When all one needs is the ability to display simple 2D Image Slices, then the vtkImageViewer class (embeded in
the vtkTkImageViewerWidget) is ofen the most convenient option. This takes four inputs namely: (i) Input an
image of type vtkImageData, (ii) ZSlice the slice to be displayed in the case of 3D Images, (iii) ColorLevel
the Level of the Colormap and (iv) ColorWindow the Window of the Colormap.
The following simple script (script12-1.tcl) illustrates its use. The rst part is fairly standard:
wm geometry . 300x150
Next we create a simple GUI with an empty frame (top) for the viewer and a slider/button combination on the
bottom for selecting the slice and exiting.
set top [ frame .top ]; set bot [ frame .bottom ]
pack $bot -side bottom -expand false -fill x -pady 2 -padx 20
pack $top -side top -expand true -fill both
set slice 0
label $bot.a -text "Smoothness:"
set slicescale [ scale $bot.b -variable slice -from 0 -to 10 -digits 3 \
-orient horizontal -resolution 1 ]
pack $bot.a $bot.b -side left -expand true -fill x
button $bot.exit -text "Exit!" -command { destroy . ; exit }
pack $bot.exit -side right
Next we load the image, which is stored, in this case, in the native VTK format. Upon loading, we print out some
diagnostics and get the number of slices, which is used to congure the range of the slider
# Load the Image
set reader [ vtkStructuredPointsReader New ]
$reader SetFileName brain.vt; $reader Update
set img [ $reader GetOutput ]
puts stdout "dimensions [ $img GetDimensions ] spacing [ $img GetSpacing ]"
set numslices [ lindex [ $img GetDimensions ] 2 ]
$slicescale configure -to [ expr $numslices - 1 ]
Now we are ready to create the viewer widget. This is created and packed much like a usual Tk Widget. We call
the GetImageViewer method to get the actual image viewer (a special type of renderer, of type vtkImageWindow,
that is embedded in the viewer widget).
set v1 [ vtkTkImageViewerWidget $top.1 -height 170 -width 170 ]
94
CHAPTER 12. DISPLAYING IMAGES IN VTK Draft Sep. 2009
pack $v1 -side top -expand true -fill both -pady 2
# Force GUI to be drawn
update
# Get the Image Viewers
set viewer [ $v1 GetImageViewer ]
Next we set 4 inputs for the viewer, as described previously:
$viewer SetInput $img
$viewer SetZSlice 0
$viewer SetColorLevel 128; $viewer SetColorWindow 255
$viewer Render
Finally a little bit of magic to set the slice from the GUI and to process expose events. GUI-elements need
to re-draw themselves each time they are exposed, i.e. a window moves from being above them etc. VTK
Render widgets, unlike say simple widgets like buttons, do not automatically do this unless they are explicitly
controlled by an interactor. They need be asked politely to re-draw themselves using their Render method. This is
accomplished using the Tcl bind command which captures GUI-events and sends them to appropriate procedures
(callbacks). A similar technique is used to detect when the value of the slider has changed.
eval "bind $v1 <Expose> { $viewer Render }"
bind $slicescale <ButtonRelease> { SetSlice }
Finally the SetSlice callback (which in the script is near the top). We use global variables here for simplicity.
proc SetSlice { } {
global slice;
global viewer
$viewer SetZSlice [ expr int($slice) ];
$viewer Render
}
The Image Viewer has a number of weaknesses. Images are displayed using simple bitmaps in actual-size i.e. an
100x100 images will use 100x100 pixels with no possibility for zooming etc. These limitations can be overcome
by magnifying the image, but this requires modifying the actual data directly which should be avoided (see rule
1). Texture mapping and 3D rendering oer alternatives.
12.4 Displaying Texture Mapped Slices
The pipeline for displaying slices as textures mapped onto planes is shown in Figure 12.2. In general there are
three parts: (i) We draw the geometry by placing a rectangle at the appropriate position. (ii) We generate the
95
CHAPTER 12. DISPLAYING IMAGES IN VTK Draft Sep. 2009
Figure 12.2: The
pipeline for displaying an
image slice as a tex-
ture mapped onto a rect-
angular plane. The
top row is the geome-
try pipeline which sim-
ply displays a bounded
plane. The middle row
is the texture pipeline
for generating the ap-
pearance of the rectan-
gle. A colormap (bot-
tom) is added to perform
the mapping from image
intensity to colors.
source image by extracting the appropriate 2D slice from a potentially 3D image. (iii) We create the texture by
combining this 2D image slice with a Colormap and apply it to the geometry.
The st part of the script (script12-2.tcl) is fairly standard:
wm geometry . 200x200
Next we load the image using a StructuredPointsReader, which reads in images stored in the native vtk format.
Following this we extract a single slice from the image using the vtkExtractVOI lter (VOI= Volume of Interest).
Incidentally, if we wanted to extract a coronal slice we could specify $voi SetVOI 0 41 25 25 0 55, to
extract a slice that has constant y-axis.
set tr [ vtkStructuredPointsReader New ]
$tr SetFileName brain.vt
$tr Update
set voi [ vtkExtractVOI New ]
$voi SetInput [ $tr GetOutput ]
$voi SetVOI 0 41 0 47 30 30
The next step is to generate a simple colormap that maps 0 to black and 1 to white. Then a texture (vtkTexture)
is created which takes two inputs: (i) the image slice from the output of vtkExtractVOI, and (ii) the colormap.
set colormap [ vtkLookupTable New ]
$colormap SetNumberOfColors 256
$colormap SetTableRange 0 255
for { set i 0 } { $i < 256 } { incr i } {
set v [ expr $i /255.0 ]
$colormap SetTableValue $i $v $v $v 1.0
96
CHAPTER 12. DISPLAYING IMAGES IN VTK Draft Sep. 2009
}
set texture [ vtkTexture New ]
$texture SetInput [ $voi GetOutput ]
$texture InterpolateOn
$texture SetLookupTable $colormap
Once the texture is ready we shift focus to the geometry. This consists of a plane (created using vtkPlaneSource).
This specic image slice has dimensions 42 48 and voxel spacing (1.0,1.0,1.0), with the origin (the centroid of
the 0-th voxel) at (0.0,0.0,0.0). Hence the plane must extend from the bottom-left corner of the bottom-left
voxel (-0.5,-0.5,30.0) to the top-right corner of the top-right voxel (40.5,46.5,30.0), as opposed to simply to voxel
centroids. The plane source is dened by three points, the origin, and two points which dene the two axis-lines,
Point1 and Point2. The output of the plane source is fed to the Mapper.
set imageplane [ vtkPlaneSource New ]
$imageplane SetXResolution 1
$imageplane SetYResolution 1
$imageplane SetOrigin -0.5 -0.5 30
$imageplane SetPoint1 40.5 -0.5 30
$imageplane SetPoint2 -0.5 46.5 30
set map [ vtkPolyDataMapper New ]
$map SetInput [ $imageplane GetOutput ]
Geometry and Texture meet at the Actor. This takes two inputs, the mapper which denes the geometry and the
texture which denes the appearance:
set imactor [ vtkActor New ]
$imactor SetMapper $map
$imactor SetTexture $texture
Finally, the usual infrastructure for the display. We create a Renderer/RenderWindow and a RenderWindowInter-
actor.
set ren [ vtkRenderer New ]
$ren AddActor $imactor
set renwin [ vtkRenderWindow New ]
$renwin AddRenderer $ren
$renwin SetSize 300 300
$ren ResetCamera
$renwin Render
# Interactor
set iren [ vtkRenderWindowInteractor New ]
97
CHAPTER 12. DISPLAYING IMAGES IN VTK Draft Sep. 2009
Figure 12.3: In
volume rendering, a
2D view of the whole
3D image is generated
by a process known as
ray-casting. At each
point in the camera
plane a ray is projected
through the 3D image
where an integral of the
form
_
l
F(intensity)
G(transparency)dl is
computed.
Figure 12.4: The
pipeline for volume ren-
dering. This is not as
complicated as it rst
appears and is described
in the text.
$iren SetRenderWindow $renwin
$iren Initialize
$iren AddObserver SetExitMethod { exit }
wm withdraw .; vwait forever
12.5 Volume Rendering
Volume rendering is the art of generating a 2D view of an object that is, in some way, an integral of the whole
3D volume. The whole process involves, in most cases, integrating a function of the image along parallel rays,
as illustrated in Figure 12.3, to generate a single 2D representation of the image. For example, in maximum-
intensity projection which is often used in angiography, the output of the ray-casting is simply the maximum value
encountered in the image along each ray. For more volumetric data, a more complex representation is used where
the integral becomes a function of the intensity mapping assigned to each voxel intensity and the transparency
mapping assigned to each intensity. Dening these mappings such that some structures become transparent
whereas others are highlighted is an art which, in some respects, resembles a segmentation process using simple
thresholds.
VTK provides a number of dierent volume rendering techniques, which are covered in detail in the VTK Users
guide. We discuss two of them here, which both involve the same type of ray casting described in Figure 12.3.
98
CHAPTER 12. DISPLAYING IMAGES IN VTK Draft Sep. 2009
In the rst method, the ray casting is done in software, whereas in the second method, we render a set of 2D
texture plane images and let the graphics card simulate the ray-casting process in hardware. This second method
uses hardware acceleration to give signicantly better performance at the expense of a slight decrease in quality
and exibility.
The basic structure of both methods is similar, see Figure 12.4 for the pipeline. The pipelines involves: (i) En-
suring that the input image is either UnsignedChar or Short using the vtkImageCast or vtkImageShiftScale lters.
(ii) Creating the two mapping functions, the opacity transfer function (G) and the intensity transfer function (F)
and attaching them to vtkVolumeProperty object. (iii) Creating the ray casting structure using a combination of
a vtkVolumeRayCastFunction which is then attached to a vtkVolumeRayCastMapper (vtkFixedPointVolumeRay-
CastMapper is another type of Volume Mapper), to which the image generated in (i) is fed. Finally the Mapper
and vtkVolumeProperty meet at an output class called vtkVolume which is to volume rendered data what vtkActor
is to polygonal data.
We note that for the texture mapping approach, the only changes are in the top row, where there is no longer a
vtkVolumeRayCastFunction and we replace the FixedPointVolumeRayCastMapper with a VolumeTextureMapper.
Software Volume Rendering by Ray Casting: This is illustrated in the following script (script12-3.tcl).
The rst part of the script is the standard header:
wm geometry . 200x200
Next we read the image in, and cast it to Unsigned Char using an instance of vtkImageCast:
# Create the reader for the data
# First Load the Image
set reader [ vtkStructuredPointsReader ]
$reader SetFileName brain.vt
$reader Update
set cast [ vtkImageCast New ]
$cast SetInput [ $reader GetOutput ]
$cast SetOutputScalarTypeToUnsignedChar
Next, we create the two transfer functions. For the opacity, this is a scalar function that maps the input intensity
to a scalar range (0.0:1.0). This is a piecewise linear function (as implied by the class name vtkPiecewiseFunction)
which generates a mapping based on a number of knot points. Functions of arbitrary complexity can be used.
The basic principle of most volume rendering is that some dark structures (typically outside the part of the image
we are interested in) are made completely transparent by mapping a low intensity value (e.g. 40.0 in this case)
to an opacity of 0.0 (completely transparent). We often saturate the upper range as well.
The intensity transfer function is generated using an instance of vtkColorTransferFunction. This takes the intensity
value and generates a vector output. This vector has three components corresponding to the red, green and blue
output color. For gray-scale like output we set the three components to equal values. This particular example, is
simply a linear ramp from black to white.
Finally the two transfer functions are attached to an instance of vtkVolumeProperty.
99
CHAPTER 12. DISPLAYING IMAGES IN VTK Draft Sep. 2009
# Create transfer function mapping scalar value to opacity
set opacityTransferFunction [ vtkPiecewiseFunction New ]
$opacityTransferFunction AddPoint 40 0.0
$opacityTransferFunction AddPoint 255 1.0
# Create transfer function mapping scalar value to color
set colorTransferFunction [ vtkColorTransferFunction New ]
$colorTransferFunction AddRGBPoint 0.0 0.0 0.0 0.0
$colorTransferFunction AddRGBPoint 255.0 1.0 1.0 1.0
# The property describes how the data will look
set volumeProperty [ vtkVolumeProperty New ]
$volumeProperty SetColor $colorTransferFunction
$volumeProperty SetScalarOpacity $opacityTransferFunction
$volumeProperty SetInterpolationTypeToLinear
The next step is to generate the ray-casting itself (top-row of Figure 12.4). This consists of the Volume Mapper
which is the replacement for vtkPolyDataMapper for volume rendered data. The Mapper takes two inputs, the ray-
cast function and an input image. The Mapper with all input data types and up to four components. It performs
composite or MIP rendering, and can be intermixed with geometric data. Calculation is performed in 15 bit xed
point precision. The Mapper has some convenience functions that allow you to specify the sampling distance along
the ray - SetMinimumImageSampleDistance. Alternatively, you can also use SetAutoAdjustSampleDistances that
adjusts the distance based on the interactivity desired by the user.
set volumeMapper [ vtkFixedPointVolumeRayCastMapper New ]
$volumeMapper SetInput [ $cast GetOutput]
Finally we create the Volume itself. This is of type vtkVolume and is the replacement for vtkActor for volume-
rendered data. It takes as inputs, the VolumeMapper and the VolumeProperty created earlier.
# The volume holds the mapper and the property and
# can be used to position/orient the volume
set volume [ vtkVolume New ]
$volume SetMapper $volumeMapper
$volume SetProperty $volumeProperty
The rest should be familiar by now. The volume is attached to a renderer in a similar fashion the way we attach
vtkActors. Then the usual RenderWindow and RenderWindowInteractor are created. The only additional element
is that we also set the desired update rate of the render window (SetDesiredUpdateRate) which denes the quality
of the volume rendering the lower the value the better. This is often valuable on slower computers (e.g. laptops):
set ren [ vtkRenderer New ]
$ren AddVolume $volume
100
CHAPTER 12. DISPLAYING IMAGES IN VTK Draft Sep. 2009
set renwin [ vtkRenderWindow New ]
$renwin AddRenderer $ren
$renwin SetSize 300 300
$ren ResetCamera
$renwin Render
$renwin SetDesiredUpdateRate 1.0
set iren [ vtkRenderWindowInteractor New ]
$iren SetRenderWindow $renwin
$iren Initialize
$iren AddObserver SetExitMethod { exit }
wm withdraw .; vwait forever
Hardware Accelerated Volume Rendering by Texture Mapping: Hardware accelerated volume render-
ing can yield signicant performance improvements over software-based ray-casting. This was a feature available
ten years ago on only really expensive graphics boards (typical medium-high end SGI graphics boards run close to
$10,000!). Now, the average sub $50 graphics board is fully capable of handling decent sized volumes! A major
revolution happened, when NVIDIA released Open-GL accelerated graphics drivers for Linux, which enabled the
transition from SGI-only setups to Linux about 7-8 years ago (2000 or so). This capability is now practically
universal.
To leverage hardware accelerated texture mapping for volume rendering we replace (new script script12-4.tcl) the
following lines from the previous script (script12-3.tcl):
set volumeMapper [ vtkFixedPointVolumeRayCastMapper New ]
$volumeMapper SetInput [ $cast GetOutput]
with:
# The mapper
# THIS IS THE ONLY CHANGE FROM script12-3.tcl !!
set volumeMapper [ vtkVolumeTextureMapper2D New ]
$volumeMapper SetInput [ $cast GetOutput]
# END OF CHANGE
This switches the mapper to texture mapping and eliminates the ray cast composite function.
12.6 A Big Example
A complete image viewer application is presented in script12-5.tcl. This uses the [ Incr Tcl ] extensions to build
a more-or-less functional 3D image viewer. The script consists of about 380 lines of code, and it is too long to
reproduce here. Rather I will highlight some key features, and expect you to look at it yourselves.
The viewer application consists of (i) an [ Incr ] Tcl class MyViewer and (ii) a little bit of code at the end to
101
CHAPTER 12. DISPLAYING IMAGES IN VTK Draft Sep. 2009
Figure 12.5: A Complete Viewer Application.
initialize this. MyViewer has the following components:
Methods
Constructor/Destructor: Create and Destroy the object.
InitializeParameters: Set all member variables to default values
PermuteImage: Since we are using the vtkImageViewer class, this is used to set to change the axis of the
image around to permit dierent slices to be displayed.
FlipScaleAndDisplayImage: As the names suggests, this performs left-right and top-bottom ips of the
image and sends it to the display.
ResetParametersAndGUI: Upon Image Load reset the ranges of the scales etc.
LoadImage : Loads an Image
BuildControlFrameGUI : Builds the Left frame with all the controls
GenerateMenuUI : Create the menu
CreateViewer: Creates the viewer
GenerateUI : Master routine for building the UI
Member Variables: There a number of these for storing (i) key widgets, (ii) the two images, currentimage=the
original load image and currentresults=the modied image for display and (iii) the params static array for storing
variables bound to widgets.
Breaking the Pipeline A key technique illustrated in this example is breaking the pipeline. Most VTK
example code assumes that the visualization will be run once through and that the pipeline goes from data source
to display. In real programs, we have a number of small pipelines that do piecemeal tasks that are more manually
connected. For example consider the PermuteImage method shown below:
itcl::body MyViewer::PermuteImage { } {
set perm [ vtkImagePermute New ]
$perm SetInput $currentimage
switch $params($this,slice) {
"XY" { $perm SetFilteredAxes 0 1 2 }
"YX" { $perm SetFilteredAxes 1 0 2 }
"XZ" { $perm SetFilteredAxes 0 2 1 }
"ZX" { $perm SetFilteredAxes 2 0 1 }
"YZ" { $perm SetFilteredAxes 1 2 0 }
"ZY" { $perm SetFilteredAxes 2 1 0 }
}
$perm Update
102
CHAPTER 12. DISPLAYING IMAGES IN VTK Draft Sep. 2009
$currentresults ShallowCopy [ $perm GetOutput ]
$perm Delete
}
Here the image input is specied in currentimage. The axes of the image are permuted using vtkImagePermute
and the output stored in another instance of vtkImageData currentresults using the ShallowCopy command of
vtkImageData. Then the vtkImagePermute object perm is deleted thus destroying the pipeline, yet maintaining
the output for setting as input to the viewer at a later stage.
1
Assignment
Read the Handout!
Read the vtkImageData Man page to get fuller sense of the functionality in vtkImageData.
Modify, in three steps, script11-1.tcl to :
1. Change the code to change the colorlevel and window to generate a binary looking image.
2. Modify script12-2.tcl to show BOTH an XY and an XZ slice (i.e. two separate Actors with associate
pipelines!)
3. Modify script12-4.tcl to adjust the transparency/opacity to fade out all gray matter (the outer layer
of the brain!)
1
The ShallowCopy function is broken in VTK 5.2 a bug x is applied to the version of VTK 5.2 that comes with
BioImage Suite 3.
103
Draft Sep. 2009
Chapter 13
Transforming Surfaces and Reslicing
Images
Medical image analysis programming involves, for the most part, three types of objects: (i) images, (ii) surfaces
and (iii) transformations. Transformations are often computed by a registration algorithm and can be used to
transform a surface (vtkTransformPolyDataFilter) and to reslice an image (vtkImageReslice). VTK has support
for both linear (i.e. transformations that are essentially 4 4 matrices) and non-linear warping transformations.
13.1 Introduction
The transformation class hierarchies in VTK are shown in Figures 13.1, 13.2 and 13.3. Transformations are maps
which implement the following equation:
T : x y, or y = T(x) (13.1)
where x is the input point and y is the output point. Transformations can be concatenated to yield complex
maps. We can describe a compound transformation in which a point is transformed rst by transformation T
1
and next by T
2
as:
T
1
: x x

, T
2
: x

y, or y = T
2
(T
1
(x)) (13.2)
This style of concatenation is termed post-multiply in VTK because of the order in which, in the case of linear
transformations, the transformation matrices are multiplied. In the two classes which allow for concatenation of
multiple transformations, namely vtkTransform and vtkGeneralTransform, the default order of concatenation is
pre-multiply. In my own work, I have never found this to be useful, and I switch these to post-multiply mode
using their PostMultiply methods.
13.2 Homogeneous Linear Transformations
Representation: Linear transformations in VTK are represented internally as 44 matrices. This enables the
use of a single operation to capture both a translation as well as a combination of rotation/shear/scale. Ordinarily,
we would write such a transformation in two parts as:
y = Ax +b (13.3)
where A is a 3 3 matrix that performs a combinations of rotation, scale and shear and b is a 3 1 vector
specifying the translation. A more compact representation is to use homogeneous coordinates. To accomplish
104
CHAPTER 13. TRANSFORMING SURFACES AND RESLICING IMAGES Draft Sep. 2009
this, we write each point as a 4-vector (x
1
, x
2
, x
3
, 1), and apply the transformation as follows:
_

_
y
1
y
2
y
3
1
_

_
=
_

_
A
11
A
12
A
13
b
1
A
21
A
22
A
13
b
2
A
31
A
32
A
13
b
3
0 0 0 1
_

_
x
1
x
2
x
3
1
_

_
(13.4)
This method can be used to transform all linear transformations into linear algebra operations on 4 4 matrices.
This enables easy concatenation (matrix multiplication) and inversion (matrix inversion). Note also that a linear
transformation can have at most 12-free parameters. There are 3 general types of linear transformations as follows:
1. Rigid these have six parameters (3 rotations and 3 translations)
2. Similarity these have seven parameters, rigid + overall scale factor
3. Ane this is the general linear transformation group and has 12 parameters.
The vtkMatrix4x4 Class: The class hierarchy for all linear transformations is shown in Figure 13.2. A key
helper class is the vtkMatrix4x4 class which is used to store the 4 4 matrices. This class has a number of
methods, the most important of which are:
1. void SetElement (int i, int j, double value)
2. double GetElement (int i, int j) const
3. void Zero ()
4. void Identity ()
5. void DeepCopy (vtkMatrix4x4 *source)
6. double Determinant ()
7. void MultiplyPoint (const oat in[4], oat out[4])
An example of directly using vtkMatrix4x4 is presented in the script below (script13-1.tcl)
wm withdraw .
set mat [ vtkMatrix4x4 New ]
$mat Identity
$mat SetElement 0 0 0
$mat SetElement 0 1 1.0
$mat SetElement 1 1 0.0
$mat SetElement 1 0 1.0
for { set i 0 } { $i <= 3 } { incr i } {
puts -nonewline stdout "\["
for { set j 0 } { $j <= 3 } { incr j } {
puts -nonewline stdout "[ $mat GetElement $i $j] "
}
puts stdout "\]"
}
This script creates the matrix, modies a small part and prints its contents.
The vtkTransform Class: This complex class has a variety of functionality for implementing linear trans-
formations. It has a vtkMatrix4x4 member which stores the current matrix, and allows for the concatenation of
various operations, either in pre-multiply or post-multiply order. The most important of its methods are:
105
CHAPTER 13. TRANSFORMING SURFACES AND RESLICING IMAGES Draft Sep. 2009
void Identity ()
void PostMultiply ()
void Inverse ()
vtkLinearTransform * GetLinearInverse ()
void SetMatrix (vtkMatrix4x4 *matrix)
void Concatenate (vtkMatrix4x4 *matrix)
void Concatenate (vtkLinearTransform *transform)
void Translate (double x, double y, double z)
void RotateWXYZ (double angle, double x, double y, double z)
void RotateX (double angle)
void RotateY (double angle)
void RotateZ (double angle)
void Scale (double x, double y, double z)
Typically, when using vtkTransform, we rst set it to Identity and then invoke the PostMultiply method. Then a
series of transformations can be concatenated to yield a single compound linear transformations. This concate-
nations can be either explicit (using the Concatenate methods) or implicit (using the Rotate, Translate and Scale
methods). An example is shown in the script below (script13-2.tcl):
set tr [ vtkTransform New ]
$tr Identity
$tr PostMultiply
$tr Scale 2.0 2.0 1.0
$tr Translate 1.0 0.0 -3
set out [ $tr TransformPoint 4 2 1 ]
puts stdout "(4 2 1) --> ($out)"
The TransformPoint method is the most basic method of all transformation classes. It is originally dened in
vtkAbstactTransform and overridden as needed by derived classes.
The vtkLandmarkTransform class: This extremely useful class contains functionality for the least squares
estimation of a linear transformation from two sets of corresponding points. Given two point sets, each having n
points: X = (x
1
, x
2
, . . . , x
n
) and Y = (y
1
, y
2
, . . . , y
n
) the operation implemented by this class can be written
mathematically as:

T =
arg min
T
n

i=1
[x
i
y
i
[
2
(13.5)
where T is a linear transformation. The key methods of vtkLandmarkTransform are:
void SetSourceLandmarks (vtkPoints *points)
void SetTargetLandmarks (vtkPoints *points)
void SetModeToRigidBody ()
void SetModeToSimilarity ()
void SetModeToAne ()
The last three specify the exact form of T which can be restricted to be rigid, similarity or the full ane
transformation. Its usage is illustrated in the following script (script13-3.tcl). First we dene the two sets of
points:
set pts1 [ vtkPoints New ]
106
CHAPTER 13. TRANSFORMING SURFACES AND RESLICING IMAGES Draft Sep. 2009
$pts1 SetNumberOfPoints 5
$pts1 SetPoint 0 0.0 0.0 0.0; $pts1 SetPoint 1 1.0 0.0 0.0
$pts1 SetPoint 2 0.0 1.0 0.0; $pts1 SetPoint 3 0.0 0.0 1.0
$pts1 SetPoint 4 1.0 1.0 1.0
set pts2 [ vtkPoints New ]
$pts2 SetNumberOfPoints 5
$pts2 SetPoint 0 0.0 0.1 0.0; $pts2 SetPoint 1 2.0 0.1 0.0
$pts2 SetPoint 2 0.0 1.1 0.0; $pts2 SetPoint 3 0.0 0.1 3.0
$pts2 SetPoint 4 2.0 1.1 3.0
Next we create a vtkLandmarkTransform object, set the two point sets and the transformation type and invoke
the Update method to compute the transformation:
set land [ vtkLandmarkTransform New ]
$land SetSourceLandmarks $pts1
$land SetTargetLandmarks $pts2
$land SetModeToAffine
$land Update
Once this is done, we print the output, as before:
set mat [ $land GetMatrix ]
puts stderr "Fitting Output is:"
for { set i 0 } { $i <= 3 } { incr i } {
puts -nonewline stdout "\["
for { set j 0 } { $j <= 3 } { incr j } {
puts -nonewline stdout "[ format "%+5.2f" [ $mat GetElement $i $j]] "
}
puts stdout "\]"
}
13.3 Non-Linear Transformations
Non-linear transformations are loosely dened as those transformations which can not be expressed as a 4 4
matrix. The family of nonlinear transformations dened in VTK is shown in Figure 13.3. The most useful ones
are vtkGridTransform and vtkThinPlateSplineTransform.
vtkThinPlateSplineTransform: Kernel-based transformations, such as the thin-plate spline transform, are
based on (i) two sets of corresponding points (landmarks) which the transformation maps exactly, and (ii) an
interpolation rule (kernel) which is used to sensibly interpolate between these landmark points. The thin-plate
spline transform, popularized by Bookstein in the early 90s, is one of the most common forms of this type of
transformation. Its most useful methods are:
void SetBasisToR ()
107
CHAPTER 13. TRANSFORMING SURFACES AND RESLICING IMAGES Draft Sep. 2009
void SetBasisToR2LogR ()
void SetSourceLandmarks (vtkPoints *source)
void SetTargetLandmarks (vtkPoints *target)
The general usage the syntax is very similar to vtkLandmarkTransform takes the form:
set tps [ vtkThinPlateSplineTransform New ]
$tps SetSourceLandmarks $pts1
$tps SetTargetLandmarks $pts2
$tps SetBasisToR
$tps Update
For 2D transformations use the SetBasisToR2LogR method instead, to select the appropriate basis function
(kernel).
vtkGridTransform: The most general type of non-linear transformation is one explicitly specied by a dense
displacement eld. VTK provides the vtkGridTransform class for this purpose. Its key methods are:
virtual void SetDisplacementGrid (vtkImageData *)
virtual void SetDisplacementScale (double)
virtual void SetDisplacementShift (double)
void SetInterpolationModeToNearestNeighbor ()
void SetInterpolationModeToLinear ()
void SetInterpolationModeToCubic ()
The displacements are specied on a grid stored in a vtkImageData structure this must have three components
(frames) which store the displacements in x, y, and z respectively. These displacements can be, optionally, scaled by
a scale factor (specied using SetDisplacementScale) and shifted by a translation term (SetDisplacementShift).
The only additional parameter is the interpolation mode which species how this displacement eld is to be
interpreted. (An internal extension also provides for a B-Spline tensor grid interpolation method). The following
snippet describes the basic usage:
set img [ vtkImageData New ]
$img SetDimensions 10 10 10
$img SetOrigin 0.0 0.0 0.0
$img SetSpacing 10.0 10.0 10.0
$img SetNumberOfScalarComponents 3
$img SetScalarTypeToDouble
$img AllocateScalars
set data [ [ $img GetPointData ] GetScalars ]
$data FillComponent 0 0.0
$data FillComponent 1 0.0
$data FillComponent 2 0.0
# Set some displacements much like image intensities
# Create Grid
set grid [ vtkGridTransform New ]
$grid SetInterpolationModeToLinear
108
CHAPTER 13. TRANSFORMING SURFACES AND RESLICING IMAGES Draft Sep. 2009
$grid SetDisplacementGrid $img
$img Delete
The image is created and manipulated as usual. Each voxel contains three components (its x,y and z displacement).
Next, the grid transform is created and the displacement eld image is attached to it.
13.4 The vtkGeneralTransform
The vtkGeneralTransform class allows for the concatenation of a number of transformations into a single transfor-
mation. These transformations (unlike similar functionality in vtkTransform) can be both linear and/or non-linear.
The syntax is very simple as illustrated in the following code snippet:
set gen [ vtkGeneralTransform New ]
$gen PostMultiply
$gen Concatenate $t1
$gen Concatenate $t2
$gen Concatenate $t3
where $t1,$t2 and $t3 are existing transformations.
13.5 Transforming Surfaces
Surfaces are most easily transformed using the vtkPolyDataFilter. A short code snippet illustrates its use:
set tr [ vtkTransform New ]
$tr Translate 10 5 0
set tf vtkTransformPolyDataFilter New ]
$tf SetInput $poly
$tf SetTransform $tr
$tf Update
set output [ vtkPolyData New ]
$output ShallowCopy [ $tf GetOutput ]
$tf Delete
The vtkTransformPolyDataFilter class takes two inputs: (i) an input surface and (ii) a transformation. In this
case, the script assumes an input surface stored in the variable $poly. While a linear transformation is created,
a non-linear transformation can also be specied. The output is also of type vtkPolyData. The lter essentially
transforms the points of the input surface one-by-one and stores the result in the output surface.
109
CHAPTER 13. TRANSFORMING SURFACES AND RESLICING IMAGES Draft Sep. 2009
13.6 Reslicing Images
Reslicing images is at the heart of most image registration procedures. While transforming surfaces is intuitive,
and can be summarized in the three steps (i) take point, (ii) transform point and (iii) store point, image reslicing
is somewhat counter-intuitive.
We will explain the process with reference to gure 13.4. In a typical registration process we have a Reference
image and a Target image. The registration estimates the transformation FROM the Reference image TO the
target image. This transformation can then be used in an image-reslicing operation to warp the Target image
BACK to the Reference image, i.e. make the target look like the reference. In this way, while the transformation
is forward the image moves BACKWARDS.
The process can be described by the following recipe, once the transformation exists:
Create an empty image (often having the same geometry as a Reference Image).
For each point (voxel) r in the empty image:
1. Compute its corresponding point in the Target image r

, T : r r

.
2. Interpolate the target image to nd the image intensity I at position r

which rarely corresponds

to an exact voxel.
3. Set the voxel r in the empty reference image to have intensity I.
Repeat for all voxels.
VTK has the very powerful vtkImageReslice class for performing various image-reslicing operations. An example
is shown in the script (script13-4.tcl) below: First we load an image:
set reader [ vtkTIFFReader New ]
$reader SetFileName brain.tif
$reader Update
Next we specify a rotation. All rotations are by default centered at the origin. In order to get a rotation centered
at the center of the image (in this case 62.0,80.0,0.0) we rst translate the center of the image to the origin,
apply the rotation, and translate back:
set xform [ vtkTransform New ]
$xform PostMultiply
$xform Translate -62.0 -80.0 0
$xform RotateWXYZ 20 0 0 1
$xform Translate 62.0 80.0 0.0
Next the heart of the script, the reslicing. vtkImageReslice has a ton of options; we present here one of the
two most common ways of using it. We specify the size of the empty image using the three statements
SetOutputExtent, SetOutputOrigin and SetOutputSpacing. The last two should be obvious: they specify the
position of the 0-th voxel and the distance between voxel centroids. The SetOutputExtent method species the
size of the empty image. The most common usage has the form:
SetOutputExtent 0 width-1 0 height-1 0 depth-1
Next we specify the input image (the moving image, the image to be resliced), the transformation and the
interpolation mode:
110
CHAPTER 13. TRANSFORMING SURFACES AND RESLICING IMAGES Draft Sep. 2009
set resl [ vtkImageReslice New ]
$resl SetOutputExtent 0 199 0 199 0 0
$resl SetOutputOrigin -20.0 -20.0 0.0
$resl SetOutputSpacing 1.0 1.0 1.0
$resl SetInput [ $reader GetOutput ]
$resl SetResliceTransform $xform
$resl SetInterpolationModeToLinear
$resl Update
Finally, we save the output to a jpeg le:
set writer [ vtkJPEGWriter New ]
$writer SetInput [ $resl GetOutput ]
$writer SetFileName rotated.jpg
$writer Write
An alternative way to specify the size of the empty image is to clone an existing image. This uses the
SetInformationInput method of vtkImageReslice for an example see script13-5.tcl. The heart of the script is:
set resl [ vtkImageReslice New ]
$resl SetInformationInput [ $reader GetOutput ]
$resl SetInput [ $reader GetOutput ]
$resl SetResliceTransform $xform
$resl SetInterpolationModeToLinear
$resl Update
where in this case the output image is set to have the same dimensions, origin and spacing as the input image,
for convenience. In a typical registration application, the reference image is used as the InformationInput and the
target image as the Input, with the registration transformation being used to set the ResliceTransform.
Assignment
Read the Handout!
Create a sphere surface (see script14-2.tcl for an example). Transform it using a linear transform (e.g.
scale + translate). Display both the original and the transformed surfaces as two separate actors in the
same scene. For good measure set the color of the transformed actor to red.
Hint: To set the color of an actor to red: do [ $act GetProperty ] SetColor 1.0 0.0 0.0
Create a ThinPlateSplineTransform. Use as source landmarks the 6 corners of a cube with corners (0,0,0)
and (100,100,100). Use the same points as targets but move one corner (e.g. 100,100,100) to (75,75,75).
Replace the linear transform in the script above with this non-linear transform and display both the original
and transformed surfaces.
Using script13-4.tcl as a guide, try to add a scaling to the transformation (hint do it before the rotation)
and save the output image as a pnm le (use the vtkPNMWriter).
111
CHAPTER 13. TRANSFORMING SURFACES AND RESLICING IMAGES Draft Sep. 2009
Figure 13.1: The top-level transformation hierarchy in VTK 5.2. The basic interface to all transformations is captured
in the abstract superclass vtkAbstractTransform. The transformations are divided into three branches. The rst group
is the linear transformations which are derived from vtkHomogeneousTransform. The second group is the nonlinear
transformations derived from vtkWarpTransform. The nal group is the vtkGeneralTransform.
Figure 13.2: The linear trans-
formation hierarchy derived from
vtkHomogeneousTransform. The
useful ones are on the bottom row.
vtkTransform can be used to manu-
ally specify a combination of trans-
lations/rotations and scales. vtk-
MatrixToLinearTransform is used to
create a linear transformation from
a vtkMatrix4x4 object. vtkLand-
markTransform creates a transforma-
tion by computing a least squares t
between two sets of corresponding
points. Finally, the vtkIterativeClos-
estPointTransform is an implementa-
tion of ICP transform.
Figure 13.3: The non-linear trans-
formation hierarchy derived from vtk-
WarpTransform. vtkGridTransform
is a transformation dened as a dis-
placement eld, whereas vtkThin-
PlateSpline transform implements a
transformation dened by two sets of
corresponding points and an interpo-
lation function.
112
CHAPTER 13. TRANSFORMING SURFACES AND RESLICING IMAGES Draft Sep. 2009
Figure 13.4: The Image Reslicing Process as implemented in vtkImageReslice
113
Draft Sep. 2009
Part IV
Interfacing To BioImage Suite using Tcl
114
Draft Sep. 2009
Chapter 14
Leveraging BioImage Suite Components
and Classes
In this chapter and the one following we describe how to leverage code already in the Yale BioImage Suite package
for your own projects. In this chapter, we focus on individual components, whereas in the next chapter we will
talk about how to use the BioImage Suite application framework to write your own BioImage Suite application.
More information about BioImage Suite can be found at its web-page www.bioimagesuite.org. These notes
pertain to BioImage Suite 3.0 which is still in beta. A version of this will be made available for download at the
end of this class.
14.1 Introduction
BioImage Suite is our home grown Medical Image Analysis Utility. It uses a combination of Tcl/[Incr Tcl] and
C++ and leverages both VTK and ITK fairly substantially. For more information see www.bioimagesuite.org.
BioImage Suite has around 300,000 lines of code and in it are a number of key classes and components which
can be usefully leveraged to simplify the tasks of implementing your own software. This can be accomplished by
loading the core BioImage Suite libraries as extensions into the vtk interpreter and using classes dened in the
BioImage Suite source tree directly from Tcl.
The package loadbioimagesuite.tcl loads the BioImage Suite extensions into your current vtk shell. The pack-
age assumes that BioImage Suite is installed in the default locations which are: /usr/local/bioimagesuite30
on Linux/Mac OS X and c:/yale/bioimagesuite30 on Windows. These can be overridden by either directly
editing the loadbioimagesuite.tcl le or by setting the environment variable BIOIMAGESUITE.
14.2 Loading and Saving Analyze & NIFTI Formatted Images
The rst example demonstrates the use of the class vtkpxAnalyzeImageSource and vtkpxAnalyzeImageWriter
which can be used to read and write analyze images into the old Analyze image format that is in common use.
It also transparently supports the newer NIFTI (.nii.gz) le format, which should be used instead of this. The
following script (script14-1.tcl) demonstrates its use:
lappend auto_path [ file dirname [ info script ]]
wm withdraw .; package require loadbioimagesuite 1.0
115
CHAPTER 14. LEVERAGING BIOIMAGE SUITE COMPONENTS AND CLASSES Draft Sep. 2009
Figure 14.1: . The
four specialized dis-
play objects in BioIm-
age Suite.
set ana [ vtkpxAnalyzeImageSource New ]
$ana Load axial.hdr
set img [ $ana GetOutput ]
set orient [ $ana GetOrientation ]
set shift [ vtkImageShiftScale New ]
$shift SetInput $img
$shift SetScale -1; $shift SetShift 0
$shift Update
set anaw [ vtkpxAnalyzeImageWriter New ]
$anaw SetInput [ $shift GetOutput ]
$anaw SetOrientation $orient
$anaw Save inverted.hdr
The image is rst loaded in using the vtkpxAnalyzeImageSource object. The interface to this is very simple as
shown in the script. It has two outputs (i) the usual Output (GetOutput) which is the image and an integer
(0:3) which is the orientation of the image (0=Axial,1=Coronal,2=Sagittal,3=Other).
1
Next this is contrast
inverted using an instance of vtkImageShiftScale. The inverted image is saved using an instance of vtkpxAna-
lyzeImageWriter. This last class takes three inputs: the input image, the input orientation and the lename which
is specied as an argument to the Save method.
14.3 Some Specialized Display Objects
BioImage Suite denes the following classes for specialized and streamlined image display:
vtkpxImageSlice this takes as inputs, an image, an orientation, a slice number and a frame and displays
a texture-mapped slice on a plane.
vtkpxOrthoImageSlice this is a combination of three vtkpxImageSlice objecs that display the three or-
thogonal slices simultaneously.
vtkpxObliqueSlice this class displays an oblique slice through an image
vtkpxVolume this is a derived class from vtkVolume and it streamlines the volume rendering pipeline.
These are shown in Figure 14.1.
The vtkpxImageSlice class: This displays a single image slice appropriately oriented in space. Its use is
illustrated in the following script (script14-4.tcl). Compare this to script12-2.tcl to appreciate the encapsulation
of the whole image slice pipeline into a single class.
1
The Orientation eld represents old-style notation here we are replacing this with a more complex version that
not-only species general orientation but axial orientation this is useful for DICOM/NIFTI images which have richer
headers. We will cover this in more detail in the next chapter, together with loading DICOM images etc.
116
CHAPTER 14. LEVERAGING BIOIMAGE SUITE COMPONENTS AND CLASSES Draft Sep. 2009
We rst load the image.
set tr [ vtkpxAnalyzeImageSource New ]
$tr Load axial.hdr
Then we create the imageslice object and specify: (i) the input image, (ii) the frame in the case of 4D images,
(iii) the current plane (2=XY,1=XZ,0=YZ), (iv) the level which is the slice number in the orientation selected
by the plane and (v) the display mode (0=nothing,1=image only,2 = rectangle only, 3= image +rectangle). set
imageslice [ vtkpxImageSlice New ]
$imageslice SetInput [ $tr GetOutput ]
$imageslice SetFrame 0
$imageslice SetCurrentPlane 2
$imageslice SetLevel 100
$imageslice SetDisplayMode 3
The rest is pretty boring stu, by now, and included in outline form. Note that vtkpxImageSlice functions like an
actor and can be added to a renderer directly.
set ren [ vtkRenderer New ]
$ren AddActor $imageslice
set renwin [ vtkRenderWindow New ]
...
set iren [ vtkRenderWindowInteractor New ]
....
The output of this script is shown in Figure 14.1(left). In addition the lookup table can be customized using the
SetLookupTable method which takes a vtkLookupTable as its argument.
The vtkpxOrthoImageSlice class: This displays three orthogonal slices in space. It consists of 3 vtkpxIm-
ageSlice objects and its use is illustrated by script14-5.tcl. The key code is:
set dim [ [ $tr GetOutput ] GetDimensions ]
set x [ expr round([ lindex $dim 0 ] /2)]
set y [ expr round([ lindex $dim 1 ] /2)]
set z [ expr round([ lindex $dim 2 ] /2)]
set colormap [ vtkLookupTable New ]
$colormap SetNumberOfColors 256
$colormap SetTableRange 0 255
for { set i 0 } { $i < 256 } { incr i } {
$colormap SetTableValue $i [ expr $i/255.0] 0.2 0.2 1.0
}
117
CHAPTER 14. LEVERAGING BIOIMAGE SUITE COMPONENTS AND CLASSES Draft Sep. 2009
set orthoslice [ vtkpxOrthoImageSlice New ]
$orthoslice SetInput [ $tr GetOutput ]
$orthoslice SetFrame 0
$orthoslice SetLevels $x $y $z
$orthoslice SetDisplayMode 3
$orthoslice SetLookupTable $colormap
set ren [ vtkRenderer New ]
$ren AddActor $orthoslice
Note that, in this case, we specify 3 levels on each for the YZ, XZ and XY planes respectively. In addition, we
set a custom red/black lookup. The output of this script is shown in Figure 14.1(left middle).
The vtkpxObliqueImageSlice class: This displays an oblique cut through a 3D image. Its use is illustrated
by script14-6.tcl. In addition to this slice, we also display a box object that consists of the outline bounds of the
image to improve the visualization. The key code is:
# First the Oblique Slice
set obliqueslice [ vtkpxObliqueImageSlice New ]
$obliqueslice SetInput [ $tr GetOutput ]
$obliqueslice SetFrame 0
$obliqueslice SetDisplayMode 3
$obliqueslice UpdateImagePlane 20.0 0.2 0.4 0.1
# Now the Outline
set outline [ vtkOutlineFilter New ]
$outline SetInput [ $tr GetOutput ]
set map [ vtkPolyDataMapper New ]
$map SetInput [ $outline GetOutput ]
set actor [ vtkActor New ]
$actor SetMapper $map
set ren [ vtkRenderer New ]
$ren AddActor $obliqueslice
$ren AddActor $actor
The position and orientation of the oblique slice is specied using the UpdateImagePlane command which takes
four inputs: (i) the oset from the image center and (ii-iv) the x,y and z components of the normal to the
image plane. The output from this script is shown in Figure 14.1(right middle). The lookup table can also be
customized, as in vtkpxImageSlice.
The vtkpxVolume class: This class encapsulates the volume rendering pipeline in VTK. It takes as input
an image, the volume resolution, and volume rendering mode and displays a volume rendered image. Its use is
illustrated by script14-7.tcl. In addition to this slice, we also display a box object that consists of the outline
bounds of the image to improve the visualization. The key code is:
118
CHAPTER 14. LEVERAGING BIOIMAGE SUITE COMPONENTS AND CLASSES Draft Sep. 2009
set vol [ vtkpxVolume New ]
$vol SetInput [ $tr GetOutput ]
$vol SetFrame 0
$vol SetResliceModeToHalfOriginal
$vol SetVolumeRenderingModeToTextureMap
set ren [ vtkRenderer New ]
$ren AddVolume $vol
The key settings are:
ResliceMode ( SetResliceModeToOriginal, SetResliceModeToHalfOriginal, SetResliceModeToQuarterO-
riginal ) which determines the size of the displayed volume and sets the tradeo between rendering perfor-
mance and quality.
SetVolumeRenderingModeToTextureMap selects the type of volume other choices are:
1. SetVolumeRenderingModeToRayCast selects raycasting mode.
2. SetVolumeRenderingModeToMIP selects Maximum Intensity Projection.
3. SetVolumeRenderingModeToTextureMap selects 2D Texture Mapping.
4. SetVolumeRenderingModeToTextureMap3D selects 3D Texture Mapping.
In addition the intensity/transparency set up can be specied using a standard vtkLookupTable which can be set
using the SetFromLookupTable method. The output of this script is shown in Figure 14.1(right).
14.4 Some Useful Filters
BioImage Suite has a number of useful lters. We discuss a small selection in this section. In particular we will
describe:
vtkpxImageExtract extracts a single slice from a 4D image.
vtkpxMatrix a simple class to make matrix operations easier. It is a wrapper around CLapack.
vtkpxAverageImages a class for computing the average of many images.
The vtkpxImageExtract class: This is an image-to-image lter which can be used to a extract a single 2D
slice from a 4D image. It is used extensively by vtkpxImageSlice and vtkpxOrthoImageSlice. The most common
methods are:
SetCurrentPlane sets the orientation of the slice: (0=YZ,1=XZ,2=XY).
GetCurrentPlane returns the orientation of the slice
SetSliceNo sets the slice number (beginning at 0) which corresponds to the X, Y or Z coordinates if the
CurrentPlane is 0,1 or 2 respectively.
GetSliceNo returns the current slice number
SetFrame sets the current frame (beginning at 0) for 4D images.
GetFrame returns the current plane number.
The following code snippet illustrates its use:
set extr [ vtkpxImageExtract New ]
$extr SetInput $img
$extr SetCurrentPlane 2
$extr SetSliceNo 25; $extr SetFrame 0
119
CHAPTER 14. LEVERAGING BIOIMAGE SUITE COMPONENTS AND CLASSES Draft Sep. 2009
$extr Update
The vtkpxMatrix class: This class is meant as a helper class for converting MATLAB code to C++/VTK,
and as such it is more usefully employed at the C++ level. However, it can also be used with Tcl. It has a
number of methods, most of whose names are meant to be similar to MATLAB commands. The most common
methods are:
Methods for Allocating space and lling with default values:
Allocate with 2 args sets the size of the matrix
Zeros with 2 args sets the size of the matrix and lls it with zeros.
Ones with 2 args sets the size of the matrix and lls it with ones.
Eye with 1 arg creates a square identity matrix
Identity with no arguments sets the current matrix to the identity (only if it is square!)
Zero with no arguments lls the current matrix with zeros.
Fill with 1 arg lls the current matrix with the value provide.d
Copy with 1 arg copies another matrix
Copy with 5 args (other,row1,row2,col1,col2) copies a part of another matrix to this one.
Transpose transpose the matrix.
Scale with 1 arg scales the values with a constant.
ScaleAdd with 2 args (scale shift) scales the values and adds a shift
Element Manipulation
GetElement with 2 args returns the value of the element arg1 arg2
SetElement with 3 args (row,col,val) sets the value of the element(row,col)=val.
AddToElement with 3 args (row,col,val) adds val to the value of the element (row,col)
Methods for Printing/Loading/Saving:
Print with no arguments prints the matrix.
Print with 1 arg prints the matrix with a name i.e. a=[ ].
PrintRange with 4 args (row1,row2,col1,col2) prints a portion of the matrix row1:row2, col1:col2.
PrintRange with 5 args (name,row1,row2,col1,col2) as above but with a name.
Print with 2 args (name,format) prints the matrix with name and format (e.g. %.2f)
Load with 1 arg loads the matrix from a le (custom le format)
Save with 1 arg saves the matrix from a le.
Information About the Matrix:
GetSize returns a list of the size of the matrix
Max returns the maximum value in the matrix
Median returns the median value of the matrix entries.
Sum returns the sum of all the values in the matrix
SumSquares returns the sum of the squares of all the entries in the matrix.
SumMagnitude returns the some of the absolute value of all the entries in the matrix.
MaxColumn with 1 arg returns the maximum value of column arg.
MaxRow with 1 arg returns the maximum value of row arg
RowSums returns a vtkpxMatrix object with the sums of all the rows
ColumnSums returns a vtkpxMatrix object with the sums of all the columns.
Simple Matrix Operations these are static methods, so the current value of the invoking object is not
used!
Add with 3 args (A , B , C), where A,B,C are vtkpxMatrix objects results in C=A+B;
Add with 5 args (wa, A, wb, B,C), where wa,wb are constants results in C = wa*A + wb*B
Multiply with 3 args (A,B,C) results in C=A*B
Multiply3 with 4 args (A,B,C,D) results in D=A*B*C
MultiplyTripleProduct with 3 args (A,B,C) results in C=A*B*A
Linear Algebra Operations these take the current object as input:
Diagonalize with 2 args (D,U) returns the eigenvalues in a row matrix D, and the eigenvectors in
U. Input must be square!
120
CHAPTER 14. LEVERAGING BIOIMAGE SUITE COMPONENTS AND CLASSES Draft Sep. 2009
Figure 14.2: . The
vtkpxGUIOrthogonalViewer
(left, script14-2.tcl) and the
vtkpxGUIMoisaicViewer (right,
script14-3.tcl).
Eigenvalues with 1 arg (D) returns the eigenvalues in a row matrix D.
QRDecomposition with 2 args (Q,R) performs QR decomposition with outputs in Q and R
Invert inverts the current matrix
SolveLeastSquares with 2 args (B,X) Solves this*X=B , output in X (this=current matrix)
SolveWeightedLeastSquares with 3 args (W,B,X) weighted version of above.
Some of these operations are illustrated in script14-8.tcl.
The vtkpxAverageImages Class: This class takes a number of inputs (vtkImageData objects) of the SAME
size and produces, either the mean, the median, the sum, and optionally the standard deviation. Its most common
methods are:
AddInput with 1 arg adds an image to the input list
SetMedian with 1 arg if 1 then we compute the median if 0 the mean or the sum
SetSumOnly with 1 arg if 1 then we compute the sum (not the mean)
SetComputeStandardDeviation with 1 arg if 1, in addition to the mean compute the standard deviation
GetOutputStandardDeviation returns a vtkImageData object with the standard deviation (if computed)
GetOutput returns the sum,mean or median depneding on the options.
14.5 The BioImage Suite Viewers
BioImage Suite has four dierent viewers, all of which have a similar interface. These are:
1. vtkpxGUIMosaicViewer can display multiple parallel slices of the same orientation.
2. vtkpxGUIOrthogonalViewer can display orthogonal slices with linked cursors, as well volume renderings
and oblique slices.
3. vtkpxGUIObjectmapOrthogonalViewer an extension to the OrthogonalViewer which can transparently
overlay a second image (commonly the output of a segmentation, hence the designation objectmap) over
the standard image.
4. vtkpxGUI4DOrthogonalViewer a 4D Version of the Orthogonal viewer which can be used to play movies
of time-varying data.
These viewers are writen in C++ and make use of Tk widgets through a custom C++ wrapper around Tcl/Tk.
The basic interface of all viewers is the same, as illustrated by the following scripts (script14-2.tcl,script14-3.tcl),
see also Figure 14.2. These have the same rst few lines, in which we load an image and create and pack an
exit button, as follows. Note that the callback of the exit button has a second command to ensure proper exiting
under windows.
lappend auto_path [ file dirname [ info script ]]
package require loadbioimagesuite 1.0
121
CHAPTER 14. LEVERAGING BIOIMAGE SUITE COMPONENTS AND CLASSES Draft Sep. 2009
set ana [ vtkpxAnalyzeImageSource New ]
$ana Load axial.hdr
set img [ $ana GetOutput ]
set orient [ $ana GetOrientation ]
puts stderr "Image Dimensions = [ $img GetDimensions ], Orientation = $orient"
button .bot -text "Exit" -command { destroy .; exit}
pack .bot -side bottom -expand false -fill x -pady 2
To create a viewer, we follow the following three steps:
Create it, much like any VTK-object
Optionally set any other ags to customize its functionality.
Initialize it given a parent widget and an inside ag. If the inside ag is set to zero the viewer is created
inside a new toplevel widget, whose value is returned by the Initialize method.
In the case of the Orthogonal Viewer this takes, for example, the form:
set ortho [ vtkpxGUIOrthogonalViewer New ]
$ortho DisableTalairachButtons
$ortho Initialize . 1
In the case of the MosaicViewer, this code, similarly, has the form:
set mos [ vtkpxGUIMosaicViewer New ]
$mos Initialize . 1
Finally an image can be set as the input to the viewers with the command:
$ortho SetImage $img $orient
The syntax is identical for the Mosaic Viewer. SetImage can also take an optional third argument specifying the
colormap (vtkLookupTable). If this is omitted, a default is used.
Getting access to the raw vtkRenderers: To add additional actors/volumes to the viewers one needs to
get access to the raw vtkRenderer objects contained in them. In the case of the MosaicViewer, this is accomplished
using the GetRenderer method, i.e. set ren [ $mos GetRenderer 0 ]. One must be careful not to ask for a
renderer that does not yet exist!
In the case of vtkpxGUIOrthogonalViewer, there are 4 renderers. These correspond to the XY-slice, the XZ-slice,
the YZ-slice and the 3D view respectively. They are accessed by a slightly more complex setup as follows:
set ren3d [ [ $ortho GetRenderer 3 ] GetRenderer ]
set renyz [ [ $ortho GetRenderer 0 ] GetRenderer ]
122
CHAPTER 14. LEVERAGING BIOIMAGE SUITE COMPONENTS AND CLASSES Draft Sep. 2009
set renxz [ [ $ortho GetRenderer 1 ] GetRenderer ]
set renxy [ [ $ortho GetRenderer 2 ] GetRenderer ]
BioImage Suite encapsulates vtkRenderer inside a new class vtkpxGUIRenderer which adds some additional func-
tionality. The rst GetRenderer statement returns an instance of vtkpxGUIRenderer, which in turn when asked
nicely (using GetRenderer) returns the raw underlying vtkRenderer object.
vtkpxGUIOrthogonalViewer: Additional Methods
GetLastClickedPoint/ GetLastClickedPointScaled. These return the last point clicked in the viewer in voxels
(GetLastClickedPoint) or mm (GetLastClickedPointScaled). An example is shown below:
set lv [ $vtk_viewer GetLastClickedPointScaled ]
set px [ lindex $lv 0 ]
set py [ lindex $lv 1 ]
set pz [ lindex $lv 2 ]
SetCoordinates/SetScaledCoordinates $px $py $pz The cross hairs can be set using the SetCoordinates
(in voxels) or the SetScaledCoordinates (mm) methods.
SetDisplayMode3D switches to 3D only view
SetDisplayMode2D switches to the 3-slice linked cursor mode
vtkpxGUIOrthogonalObjectmapViewer: This viewer allows for the display of a second image that is
transparently overlaid onto the main image. The level of transparency is set by the Mask control (0=completely
transparent, 100=completely opaque).
The following two methods are used to set the object map image, and its associated color map (vtkLookuptable).
int SetObjectMapImage(vtkImageData* img);
int SetObjectLookupTable(vtkLookupTable* lkp);
The Objectmap Image must have the same dimensions as the primary image of the viewer (the one set using the
SetImage method).
vtkpxGUIOrthogonal4DViewer: This class expands on vtkpxGUIOrthogonalViewer (is derived from it) and has
4 renderers for each frame, which are switched in and out as the frame is changed. To obtain access to these
renderers one needs to modify the example, above, for vtkpxGUIOrthogonalViewer, in the following fashion:
set ren3d [ [ $ortho GetMultiRenderer 3 ] GetRendererForFrame $frame ]
set renyz [ [ $ortho GetMultiRenderer 0 ] GetRendererForFrame $frame ]
set renxz [ [ $ortho GetMultiRenderer 1 ] GetRendererForFrame $frame ]
set renxy [ [ $ortho GetMultiRenderer 2 ] GetRendererForFrame $frame ]
where frame is the desired frame (beginning at 0).
A Final Comment: At this point one can begin to think in terms of writing a complete application around
one of these viewers. In practice, however, it is better to leverage higher-level Bioimage Suite functionality as
desribed in the next chapter.
123
CHAPTER 14. LEVERAGING BIOIMAGE SUITE COMPONENTS AND CLASSES Draft Sep. 2009
Assignment
(Perhaps using script14-8.tcl) write code that uses vtkpxMatrix to get the eigenvalues of a 4x4 symmetric
positive denite matrix of your choice.
Using script14-4.tcl as a base, modify it to add a scale widget that automatically adjusts the displayed slice
of $imageslice along the z-axis.
Using script14-2.tcl as a base, modify this to add a large sphere (centered at 80,80,62, radius 50) to the
3D renderer of OrthogonalViewer.
Hint: You can get this renderer using: set ren [ [ $ortho GetRenderer 3 ] GetRenderer ]
Hint2: You may want to use the SetDisplayMode3D command to switch the rendering to 3D mode.
124
Draft Sep. 2009
Chapter 15
Writing your own BioImage Suite Tcl
Application
In this chapter we describe how one goes about writing a BioImage Suite application. This enables the use of many
of the standard BioImage Suite components (e.g. collections of classes) for a new application. We rst describe
some of the [Incr] Tcl classes used by BioImage Suite to enhance the functionality of bare VTK classes. Then we
desribe how to use the BioImage Suite application framework to write your own BioImage Suite application.
More information about BioImage Suite can be found at its web-page www.bioimagesuite.org.
15.1 Introduction
BioImage Suite is our home grown Medical Image Analysis Utility. It uses a combination of Tcl/[Incr Tcl] and
C++ and leverages both VTK and ITK fairly substantially. For more information see www.bioimagesuite.org.
The driving philosophy behind BioImage Suite is that it is a collection of customized applications sharing the
same components and look and feel. As such it is ideally suited (not by accident) for use by students/researchers
to create their own applications, by adding their own modules to the considerable functionality already in place.
15.2 Some Key [ Incr ] Tcl Classes
In this section we describe some key classes dened and used in BioImage Suite. A basic understanding of these
classes is important in interfacing to this software.
15.2.1 The BioImage Suite Wrapper Classes
Most objects such as images, transformations and surfaces often require the storage of additional information that
is allowed for in the standard vtk data structures. An example is the storage of lenames. While such information
can be stored in the FieldData of a vtkDataObject, this is inconvenient at best.
The solution adopted in BioImage Suite is to create wrapper [Incr] Tcl objects around the basic vtk data structures
to captures this additional functionality. The most useful of these objects are:
pxitclobject the basic parent class (bioimagesuite/base/pxitclobject.tcl), with the following derived classes:
1. pxitclimage a wrapper around vtkImageData
2. pxticlsurface a wrapper around vtkPolyData
3. pxitcltransform a wrapper around any transformation derived from vtkAbstractTransform
125
CHAPTER 15. WRITING YOUR OWN BIOIMAGE SUITE TCL APPLICATION Draft Sep. 2009
These type of wrapper objects are what is typically passed around BioImage Suite modules, so it is important to
understand what they do.
The pxitclobject class: This does very little. It has one member variable (lename) and it denes a number
of methods that derived classes are expected to override.
itcl::class pxitclobject {
# --- begin variable def -----
public variable filename ""
# --- end of variable def -----
constructor { } { }
destructor { }
# --- begin method def -----
public method GetThisPointer { } { return $this }
# Things that must be over-ridden by derived classes
# --------------------------------------------------
public method GetObject { } { return 0 }
public method GetType { } { return "vtkObject" }
public method GetDescription { } { return [ $this GetType ]}
public method UpdateStatus { } { }
public method Copy { obj } { }
public method Clear { } { }
# Procedures to Load/Save
public method Load { args } { }
public method Save { args } { }
}
The GetObject method provides a generic method for getting the VTK Data object that is wrapped inside this.
The pxitclimage class: This is the rst concrete implementation. It adds four new member variables namely:
(i) orientation, (ii) the vtkImageData object img, (iii) a lookuptable for storing an associated colormap and (iv)
a status variable.
The most useful routine, in terms of interfacing to standard VTK code is the GetImage method which returns the
underlying vtkImageData data structure. The CopyImage method can be used to copy the contents of an existing
vtkImageData object into the wrapped object inside pxitclimage. This takes two arguments i.e. ShallowCopy obj
mode, where obj is the incoming vtkImageData object and mode = 0 (shallow copy i.e. simply link the pointer)
or 1 (deep copy, actually copying the data.)
The Load function takes an optional argument which is the lename. Depending on the lename this will read
the image in as an Analyze or NIFTI or DICOM or a vtkStructuredPoints (.vtk) le. In a manner transparent to
the user.
The other interesting method of this class is GetImageHeader which returns a pointer to the class vtkbisImageHeader.
This stores all header information about the image, which includes critically the image orientation (e.g. axial,
coronal etc.), any processing history, image type etc.
For most lters you will need to copy the image header from another pxitclimage class. For example one would
use the construct:
126
CHAPTER 15. WRITING YOUR OWN BIOIMAGE SUITE TCL APPLICATION Draft Sep. 2009
Figure 15.1: An example of two of the GUI controls: Top:
The pxitclimageGUI control for manipulating an image. The
lename is at the top (empty) with the dimensions (1 1 1) at
the top right. The textbox below gives a longer description
of the image. Common functionality (Load, Save, Browse,
Clear) is available from the buttonbar below. Note that the
two buttons Display Ref and Display Trn represent addi-
tional application specic functionality. Bottom: A transfor-
mation control for manipulating transformation objects (both
linear and non-linear). The buttons Check, Go to Com-
pute and Grab represent additional application specic
functionality.
$newimg CopyImageHeader [ $oldimage GetImageHeader ]
See bioimagesuite/base/pxitclimage.tcl for more details.
The pxitcltransform class: This provides a wrapper around any vtkAbstractTransform object. It denes
one new data member transformation which stores the VTK transformation. It also provides, similarly to
pxitclimage, the methods GetTransformation and CopyTransformation.
See bioimagesuite/base/pxitcltransform.tcl for more details.
The pxitclsurface class: This provides a wrapper around any vtkPolyData object. It denes two new data
members, (i) sur the vtkPolyData that is being wrapped and (ii) a status variable. It also provides, similarly to
pxitclimage, the methods GetSurface and CopySurface.
See bioimagesuite/base/pxitclsurface.tcl for more details.
15.2.2 The BioImage Suite Simple GUI Classes
Often when we have an object we need a small set of controls to quickly perform common operations (think of
right-clicking on an object in MS-WORD). For example, in the case of an image (pxitclimage), we need to be
able to Load/ Save/ Get Information/ Display the image etc.
BioImage Suite provides a set of ugly controls for this purpose. These are derived from a parent class pxit-
clobjectGUI (which is in the le as bioimagesuite/base/pxitclobject.tcl), whose header (somewhat abbreviated) is
described below. The rst part consists of the member variables:
itcl::class pxitclobjectGUI {
public variable itclobject 0
public variable callback 0
public variable buttonbar 0
public variable loadbutton 0
public variable savebutton 0
public variable clearbutton 0
127
CHAPTER 15. WRITING YOUR OWN BIOIMAGE SUITE TCL APPLICATION Draft Sep. 2009
The itclobject member contains a pointer to the object that is being managed by the GUI (a derived class of
pxitclobject). The callback variable contains the name of the command to call when the GUI is updated. The
buttonbar, loadbutton, savebutton and clearbutton variables store the values of key widgets in the GUI, so that
the user can manipulate its appearance, e.g. pack forget $savebutton will remove the Save button from the
GUI.
Next come the constructor and the destructor as well as some simple information methods. The GUI can either
own the underlying pxitclobject or simply manage it. If it owns it, then this must be deleted when this object
is deleted. The GetThisPointer method is a trick for returning a pointer to the object in global scope. The
constructor only returns the object variable in local scope, by default.
constructor { } { ... }
destructor { if { $owns_object == 1} {
# catch { } prevents error if the operation fails
catch { itcl::delete object $itclobject }
}
}
public method GetThisPointer { } { return $this }
public method GetObject { } { return $itclobject }
public method Reset { }
Next up, methods for creating the actual GUI and updating it. The GUI can be initialized either using the Initialize
method (where basewidg) is the name of the core widget (which must not exist!), or the InitializeLite method
which often creates a more compact look. The args variable can be omitted in this case.
public method Initialize { basewidg }
public method InitializeLite { basewidg args }
public method Update { }
These are the callback functions for the standard buttons (Load, Save, Clear). They are overridden in derived
classes.
public method Info { }
public method LoadObject { }
public method SaveObject { }
public method ClearObject { }
The add function command adds a new button to the gui whose name is specied by the name variable. The
callback function for the button takes the form: $command [ $this GetObject ] $post.
public method AddFunction { command name post }
128
CHAPTER 15. WRITING YOUR OWN BIOIMAGE SUITE TCL APPLICATION Draft Sep. 2009
Finally a couple of functions for creating a new object and setting it. These must be redened by derived classes
(e.g. pxitclimageGUI).
# Function that must be overriden
protected method CreateNewObject { } { puts stdout "Error!"}
public method SetObject { tr } { puts stdout "Error!" }
The most useful derived classes of pxitclobjectGUI are pxitclimageGUI and pxitcltransformGUI which are (probably
a bad idea) dened in the same les as pxitclimage and pxitcltransform respectively.
An extra layer of functionality is provided by the pxitclmultiObjectGUI (and derived classes). This can be used to
manage a list of objects. It will not be described here in any detail.
15.2.3 The BioImage Suite Control Classes
Large GUI controls in BioImage Suite are derived from pxitclbasecontrol. The key chain in the hierarchy is:
pxitclbasecontrol --> pxitclbaseimagecontrol --> pxitclbaseimageviewer
In creating a your own BioImage Suite application, you will have to:
Instantiate a new pxitclbaseimageviewer object to manage the core viewer.
Construct your additional controls as derived classes from pxitclbaseimagecontrol and attach them to the
viewer.
We describe these classes briey next.
pxitclbasecontrol: This class provides the core functionality for a BioImage Suite control. It provides some
useful methods which encapsulate complex functionality. The following is an abbreviated list:
constructor parent the parent refers to another pxitclbasecontrol (or derived class) which acts as a
master for this control. Key events in this (slave) control are also passed to the parent (e.g. see WatchOn
below).
ShowWindow shows the main control window.
HideWindow hides the main control window.
SetTitle title sets the title of the control window to title
EnableUI widgetlist enables all widgets (and their children) in the list widget list.
DisableUI widgetlist as above .
WatchOn puts a watch cursor on the window to indicate that something is going on. Also passes this to
the parent (via $parent WatchOn).
WatchO removes the watch cursor.
AddLogoLabel w Adds the BioImage Suite in a new label created inside frame $w. The new label is
returned by this method.
CreateProgressBar Creates a progressbar inside an existing frame $w. This can be used to monitor
progress of an algorithm.
SetFilterCallbacks lter comment Takes a lter (a vtkObject e.g. vtkImageShiftScale) and attaches it to
the progressbar so that its progress can be monitored.
129
CHAPTER 15. WRITING YOUR OWN BIOIMAGE SUITE TCL APPLICATION Draft Sep. 2009
AboutCommand brings up a dialog box whose contents are controlled by the member variables appname,
version and aboutstring to provide basic information about the current application.
pxitclbaseimagecontrol: This provides additional functionality in the form of three pxitclimage objects and
methods to manipulate them. In general, although this need not be followed in all cases. the assumption is
that the imagecontrol performs some processing on the currentimage and puts the output in currentresults. The
currentimage variable will often be set by supervising imageviewer classes to which the control is attached as the
image in the viewer is changed. Derived classes are naturally free to add more images if there is a need.
The most commonly used methods in this class are:
SetImage img sets the currentimage (img is a pxitclimage object)
SetResults img sets the currentresults object.
SetImageFromObject img obj similar to SetImage, obj is the source object for the change. This is often
used to notify parent objects of a new image.
SetResultsFromObject img obj as above.
AddPrex oldname prex takes a lename and adds a prex to it to reect some processing.
SendResultsToParent passes the currentresults image to the parent object.
SendImageToParent passes the currentimage image to the parent object.
CopyResultsToImage make the results of the processing permanent, i.e. undoimage=currentimage, cur-
rentimage = currentresults.
RevertToPreviousImage undo the last CopyResultsToImage call
pxitclbaseimageviewer: This is a fairly complex (perhaps too complex) class for creating a viewer (one of
the four viewers described in the last chapter), and attaching some default functionality to it, see Figure 15.2
for an example. In the next Section we describe how to interface to this. A detailed description of this class is
beyond the scope of the current class.
15.3 The Basic Application
Most BioImage Suite applications either derive from or have an instance of the pxitclbaseimageviewer.tcl class.
This is an [ Incr ] Tcl class and can be found under bioimagesuite/main.
This basic application framework see also http://bioimagesuite.org/public/AppStructure.html is
shown in Figure 15.2.
The main application rst initializes the framework and then adds any custom controls. This is best illustrated
by means of an example (mytool.tcl).
The rst few lines are straightforward. MyUtility is the package containing the new control to be added to the
core framework.
lappend auto_path [ file dirname [ info script ]]
package require loadbioimagesuite 1.0
package require pxitclbaseimageviewer 1.0
package require myutility 1.0
# Eliminate the default tk window
wm withdraw .
The next step is to initialize the core application, by selecting which of the common components should appear
130
CHAPTER 15. WRITING YOUR OWN BIOIMAGE SUITE TCL APPLICATION Draft Sep. 2009
in the main menu.
These are the defaults and should serve most people well:
set baseviewer [ pxitclbaseimageviewer \#auto 0 ]
$baseviewer configure -appname "BioImage Suite::My Tool"
$baseviewer configure -show_standard_images 1
$baseviewer configure -enable_helpmenu 0
$baseviewer configure -enable_multisubjectcontrol 0
$baseviewer configure -enable_overlaytool 0
$baseviewer configure -enable_vvlinkgadget 0
$baseviewer configure -enable_talaraichoption 0
$baseviewer configure -enable_rendering_on_startup 0
The following is list of ags for enabling/disabling standard components (along with their default state: 1=ON,
0=OFF):
1. enable displaymenu 1 if this is set to 0 the display menu is not added.
2. enable helpmenu 1 if this is set to 0 the help menu is not added. This is useful for adding your own
menu before adding the help menu as in the standard example.
3. enable histcontrol 1 controls whether the histogram tool is added.
4. enable overlaytool 1 controls whether the overlay/registration tool is added.
5. enable brainstriptool 1 controls whether the segmentation tool is added.
6. enable headercontrol 1 controls whether the analyze header editor tool will appear under the le menu.
7. enable importcontrol 1 controls whether the image import tool will appear under the le menu.
8. enable imageutility 1 controls whether the basic image processing tool will be available.
9. enable landmarkcontrol 1 controls whether the landmark tool is included.
10. enable polydatacontrol 1 controls whether the surface tool is included.
11. enable vvlinkgadget 0 controls whether the VectorVision Link tool is included (if available)
12. enable talaraichoption 0 shows/hides the option for displaying Talairach Coordinates.
The next step is to create a viewer. BioImage Suite has four viewers (described in the previous chapter). One
(and only one) can be initialized in this application, using one of the commands below. Make sure that 3/4 are
commented out or deleted.
1
# Create a Viewer
# Uncomment one of the next four lines to select the viewer of your choice
$baseviewer InitializeOrthogonalViewer .[ pxvtable::vnewobj ] 1
#$baseviewer InitializeObjectmapViewer .[ pxvtable::vnewobj ] 1
#$baseviewer InitializeMosaicViewer .[ pxvtable::vnewobj ] 1
#$baseviewer InitializeOrthogonal4DViewer .[ pxvtable::vnewobj ] 1
The next step is to create our menu for adding new tools
1
The pxvtable::vnewobj command simply returns a unique name. It is often good practise to generate widget names
automatically rather than giving them xed names so that when you copy code from other programs it still works, i.e.
there are no name conicts!
131
CHAPTER 15. WRITING YOUR OWN BIOIMAGE SUITE TCL APPLICATION Draft Sep. 2009
Figure 15.2: The framework for a BioImage
Suite application. This consists of a menu bar,
the viewer and a status bar. All these are created
for you by invoking the pxitclbaseimageviewer
class. The new components are attached to this
framework using a customized menu entry.
Figure 15.3: A customized ap-
plication based on the core BioIm-
age Suite framework. Note the new
menu entry My Menu that pro-
vides access to the new functionality.
# Add a submenu for your own tools
set menubase [ $baseviewer cget -menubase ]
set mb [ menu $menubase.vesselm -tearoff 0 ]
$menubase add cascade -label "MyMenu" -menu $menubase.vesselm -underline 0
Each tool must be packaged into a class derived from pxitclbaseimagecontrol.tcl (more later). Here we initialize
a control of type myutility. We next add it to the menu (using the AddToMenuButton method), and register it
to the main application (using the AddControl method).
# Create your own tool (or tools), add it to the menu, and register it with baseviewer
set myutil [ myutility \#auto $baseviewer ]
$myutil Initialize [ $baseviewer GetBaseWidget ].[ pxvtable::vnewobj ]
$myutil AddToMenuButton $mb
$baseviewer AddControl $myutil
The rest is standard. Now that our menu is added, we create the help menu. Next we show the main window
and if an argument is specied attempt to load an image from it into the main viewer.
# Finally create the help menu on the far right
$baseviewer CreateHelpMenu
# Show the Main Window
$baseviewer ShowWindow
132
CHAPTER 15. WRITING YOUR OWN BIOIMAGE SUITE TCL APPLICATION Draft Sep. 2009
Figure 15.4: A typical BioImage Suite control.
This consists of a (optionally) a menu not used
in this case, a tabbed notebook widget with mul-
tiple tabs that dene the workow through the
control, potentially a button bar at the bottom
and a status bar. Each tab may be directly ac-
cessed from the main application menu, if this is
desired.
# If an argument is specified, attempt to Load an Image from it
set argc [llength $argv]
if { $argc > 0 } { $baseviewer LoadImage [lindex $argv 0] }
# Enable rendering
update idletasks
[ $baseviewer GetViewer ] SetEnableRendering 1
The customized application is shown in Figure 15.3.
15.4 A Custom Control
Custom controls in BioImage Suite are most easily constructed by creating a class derived from pxitclbaseimage-
control.tcl. This is also an [Incr] Tcl object. The typical control consists of multi-tab notebook window with
dierent tabs providing for, for example, user interface elements for specifying input data, setting parameters and
manipulating/saving output data. An example is shown in Figure 15.4.
The rest of this section describes a custom shell control (myutility.tcl) which can be used as a boilerplate for your
own controls. The rst part simply species the requirements etc.
package provide myutility 1.0
package require Itcl 3.2
package require Iwidgets 4.0
package require pxitclbaseimagecontrol 1.0
Next we dene the class header. This derives from pxitclbaseimagecontrol (to be found in bioimagesuite/main).
The common array thisparam is use to store variables which are linked to widgets directly, in the standard fashion
of object-oriented GUIs.
The constructor simply calls the parent class constructor and then the InitializeControl method this methods
sets all the default parameters.
133
CHAPTER 15. WRITING YOUR OWN BIOIMAGE SUITE TCL APPLICATION Draft Sep. 2009
The Initialize method is used to build the graphical user interface. Since this is a tabbed-notebook design, it
makes sense to have individual methods for each tab (CreateImageControl and CreateComputeControl). The
AddToMenuButton method is used to add this object to a menubar.
Next there are two methods for interfacing to the main application: (i) GetPointsFromLandmarkControl which
gets the current set of landmarks from the Landmark Tool. This is useful for setting seeds etc. (ii) GrabImage,
which grabs the current image from the viewer and stores it into one of the image-gui controls (more on this
later) in the tool.
Finally, there are two methods for actually executing our algorithm (Compute) and storing the output image
(StoreResult).
itcl::class myutility {
# derives from pxitclbaseimagecontrol
inherit pxitclbaseimagecontrol
protected common thisparam
#-----------------------------------
# construction and descruction
#-----------------------------------
constructor { par } { pxitclbaseimagecontrol::constructor $par } { InitializeControl }
destructor { }
# initialization methods
#-----------------------------------
public method InitializeControl { }
public method Initialize { inpwidg }
# interface creation methods
#-----------------------------------
protected method CreateImagesControl { name }
protected method CreateComputeControl { name }
# Add this control to a Menu Button
# ----------------------------------
public method AddToMenuButton { mb args }
# Computational Utility Stuff
# ----------------------------------
public method GetPointsFromLandmarkControl { }
public method GrabImage { image control }
# Do Something
# ----------------------------------
public method Compute { }
public method StoreResult { image { name "" } { orientation 0 } }
}
The Initialize Control method is straightforwards and simply sets some default parameter values.
134
CHAPTER 15. WRITING YOUR OWN BIOIMAGE SUITE TCL APPLICATION Draft Sep. 2009
itcl::body myutility::InitializeControl { } {
set appname "My Tool v1.0"
set aboutstring "(c) 2006"
# All Parameters that need to be appear in the GUI need to be part of this array
# ------------------------------------------------------------------------------
set thisparam($this,enablesmoothing) 1
set thisparam($this,initialdistance) 2.0
# We need three image controls
set thisparam($this,inputgui) 0
set thisparam($this,initialphigui) 0
set thisparam($this,outputgui) 0
}
The Initialize method creates the control window. This sits inside a new toplevel widget specied using the
widget parameter. Note that the initialized ag is set when this is done, multiple calls of the Initialize method
only create one set of GUI-elements.
Once the toplevel is created, we create a notebook widget using iwidgets::tabnotebook. Each tab in the notebook
is added using the notebook add command and passed to a helper method for creating the appropriate GUI inside
it.
itcl::body myutility::Initialize { widget } {
if { $initialized == 1 } { return $basewidget }
# -------------------------------------
# Create User Interface
# -------------------------------------
set basewidget [toplevel $widget ]
wm geometry $basewidget 610x450
wm withdraw $basewidget
set notebook $basewidget.notebook
iwidgets::tabnotebook $notebook -tabpos w
set widget_list(notebook) $notebook
set mb [ frame $basewidget.mb ]
pack $mb -side top -fill x -expand false
CreateImagesControl [ $notebook add -label "Images" ]
CreateComputeControl [ $notebook add -label "Compute" ]
pack $notebook -side top -fill both -expand t -padx 5
set initialized 1
SetTitle "My Utility"
# this is critical
eval "wm protocol $basewidget WM_DELETE_WINDOW { wm withdraw $basewidget }"
return $basewidget
135
CHAPTER 15. WRITING YOUR OWN BIOIMAGE SUITE TCL APPLICATION Draft Sep. 2009
}
The CreateImagesControl is used to create the GUI inside the Images tab. BioImage Suite provides a small
additional number of classes, in particular pxitclimage an [ Incr] Tcl wrapper around vtkImageData which
provides additional storage for a lename and an orientation ag (see bioimagesuite/base/pxitclimage.tcl) and
pxitclimageGUI (stored in the same le) which is a basic GUI for handling images.
Three pxitclimageGUIs are created. To each we add two new methods (buttons), a Display method which sends
the image stored in the imageGUI to the viewer and (optionally) a Grab method which takes the image from the
viewer and puts it in the image GUI.
itcl::body myutility::CreateImagesControl { base } {
set names [ list "inputgui" "initialphigui" "outputgui" ]
set titles [ list "Input Image" "Initial LevelSet" "Output" ]
for { set i 0 } { $i < [ llength $names ] } { incr i } {
set guiname [ lindex $names $i ]
set title [ lindex $titles $i ]
set thisparam($this,$guiname) [ [ pxitclimageGUI \#auto ] GetThisPointer ]
$thisparam($this,$guiname) configure -description $title
# $base.$i is a new frame that gets created and the control put in it
$thisparam($this,$guiname) Initialize $base.$i
pack $base.$i -side top -expand f -fill x
set bbut [ $thisparam($this,$guiname) cget -browsebutton ]
pack forget $bbut
$thisparam($this,$guiname) AddFunction "$parent SetImageFromObject" "Display" "$this"
if { $guiname != "Output" } {
$thisparam($this,$guiname) AddFunction "$this GrabImage" "Grab" \
"$thisparam($this,$guiname)"
}
}
}
The Compute control is straight forward, standard [ Incr ] Tcl.
itcl::body myutility::CreateComputeControl { base } {
iwidgets::labeledframe $base.frame0 -labelpos -labeltext "Parameters"
pack $base.frame0 -fill both -expand f -pady 5
set frame0 [ $base.frame0 childsite ]
checkbutton $frame0.c -variable [ itcl::scope thisparam($this,enablesmoothing) ] \
-text "Enable Smoothing"
136
CHAPTER 15. WRITING YOUR OWN BIOIMAGE SUITE TCL APPLICATION Draft Sep. 2009
iwidgets::entryfield $frame0.e -labeltext "Initial Distance:" \
-textvariable [ itcl::scope thisparam($this,initialdistance) ] \
-relief sunken -width 6 -validate real
pack $frame0.c $frame0.e -side left -padx 2 -fill x -expand false
eval "button $base.but -text \"Compute\" -command { $this Compute }"
pack $base.but -side bottom -expand t -fill x
}
The control can be added to the menu, using the AddToMenuButton command. This creates one menu entry for
each tab of the notebook as shown below:
itcl::body myutility::AddToMenuButton { mb args } {
eval "$mb add command -label \"Images\" -command {$this ShowWindow \"Images\"}"
eval "$mb add command -label \"Compute\" -command {$this ShowWindow \"Compute\"}"
}
The GrabImage function takes the current image in the viewer and puts it into one of the image guis. The main
application object is stored in the parent member variable. The pxitclbaseimageviewer GetDisplayedImage method
gets the currently displayed image in the viewer and returns it as a pxitclimage object. (From this we can get the
raw vtkImageData object using its GetImage method, i.e. [ pxitclimg GetImage ]).
This image is copied into the image stored into the image-GUI (obtained using [ $gui GetObject ]) and then the
gui is updated.
itcl::body myutility::GrabImage { image control } {
# image is a pxitclimage of the control
# control is the image gui
set gui $control
set img [ $parent GetDisplayedImage ]
if { [ $img GetImageSize ] > 0 } {
[ $gui GetObject ] ShallowCopy $img
$gui Update
}
}
The landmarks in the landmark control are stored in a special internal class called vtkpxBaseCurve. The vtkpxBase-
Curve class is too complex to describe here in any detail. The code in the next function takes care of converting
a vtkpxBaseCurve object to a naked vtkPoints class.
itcl::body myutility::GetPointsFromLandmarkControl{ } {
set landmarkcontrol [ [ $parent GetLandmarkControl ] GetLandmarkControl ]
set tempc_lv [ vtkpxBaseCurve New ]
$tempc_lv Copy [ $landmarkcontrol GetCollection -1 ] $tempc_lv Compact set
137
CHAPTER 15. WRITING YOUR OWN BIOIMAGE SUITE TCL APPLICATION Draft Sep. 2009
pts [ $tempc_lv GetPoints ]
set points [ vtkPoints New ]
$points DeepCopy $pts
$tempc_lv Delete
return $points
}
The Compute function is just a placeholder for invoking your own algorithm! The pair pxtkconsole, pxtkprint use
the BioImage Suite console (found under the help menu) to print the output.
itcl::body myutility::Compute { } {
set smoothing $thisparam($this,enablesmoothing)
set distance $thisparam($this,initialdistance)
# This opens the console
pxtkconsole
# This prints to the console
pxtkprint "Parameters\n-----------------\n"
pxtkprint "Use Smoothing : $smoothing\n"
pxtkprint "Initial Dist : $distance\n"
set points [ $this GetPointsFromLandmarkControl ]
if { $points != 0 } {
set np [ $points GetNumberOfPoints ]
pxtkprint "Number of Points : $np\n"
for { set i 0 } { $i < $np } { incr i } {
pxtkprint "Point $i : [ $points GetPoint $i ]\n"
}
pxtkprint "\n"
}
if { $points !=0 } { $points Delete }
}
The StoreResult method can be used to put an image (vtkImageData), with an associated lename (name) and
orientation into the outputgui. This is for storing the output of your procedure.
itcl::body myutility::StoreResult { image { name "" } { orientation 0 } } {
set gui $thisparam($this,outputgui)
[ $gui GetObject ] ShallowCopy $img
[ $gui GetObject ] configure -filename $name
[ $gui GetObject ] configure -orientation $orientation
$gui Update
}
138
CHAPTER 15. WRITING YOUR OWN BIOIMAGE SUITE TCL APPLICATION Draft Sep. 2009
Finally after the class is dened, the following piece of code informs anyone who tries to execute this le directly
that this is not a stand-alone program.
if { [ file rootname $argv0 ] == [ file rootname [ info script ] ] } {
puts "\n[ file rootname $argv0 ] is not a stand-alone program.\n"
exit
}
15.5 Appendix: The Object-Oriented Algorithm Framework
15.5.1 Introduction
In previous versions of BioImage Suite (up to version 2.6 released in Nov 2006), all algorithms are implemented in
C++ and invoked from either command line scripts or GUI modules both written in the Tcl scripting language.
However, the command line scripts and GUI modules are based on dierent sets of code and can sometimes
diverge in the setting of default parameters without extensive coordination. In addition, automated testing is
performed on the command line scripts and this process does not account for any potential errors introduced in
the GUI modules as the C++ interfaces for these algorithms evolve.
In the new framework, all BioImage Suite algorithms are being packaged into a single set of Incr-Tcl classes.
Incr-Tcl is a set of object-oriented extensions to the Tcl language. These classes can be used to (i) invoke the
algorithm (using an Execute method), (ii) as components of other algorithms (e.g. the image smoothing algorithm
is invoked by the edge detection algorithm), (iii) as standalone applets, i.e. dedicated single algorithm applications
with an image viewer and a graphical user interface and nally (iv) as components of a larger application. Each
algorithm explicitly denes three sets: (i) inputs, which are objects such as images, surfaces, landmarks etc.,
(ii) parameters, which are single values such as integers, strings, lenames and (iii) outputs, which are also
objects just like the inputs. Based on the denition of these sets, the base abstract classes of the framework
have functionality (which need not be touched by more concrete implementations) to (i) parse command line
arguments if the algorithm class is invoked as an application; (ii) automatically create a graphical user interface
using the CreateGUI method (this method can be overridden by some algorithms to generate a more customized
interface); (iii) to perform testing by parsing a test le.
One of the great advantages of this setup is that users can learn how to invoke these algorithms at the Graphical
User Interface level and then using the Show Command Line option embedded in the GUI to get a detailed
command line specication for performing exactly the same task by invoking exactly the same code at the
command line which enables the creation of customized batch jobs/pipelines.
Core Classes: The new framework will have at its core the following eight Incr-Tcl classes: (1) bis_option
encapsulates an option value (e.g. smoothness factor, etc.). An option can have a type of: listofvalues, boolean,
real, integer, string or lename. Within this class there is functionality for creating an appropriate graphical
user interface for each option. (2) bis_object encapsulates the input and output objects of the algorithms.
The core objects supported are: image, transform (both linear and non-linear), polygonal surface, landmark set
and electrode grid. (3) bis_basealgorithm is the core algorithm class from which all algorithms are derived.
It has all the functionality for manipulating options, inputs and outputs. (4) bis_algorithm is derived from
base_algorithm and adds the functionality needed for taking an algorithm and making it into a component or
an executable. More specialized classes are derived from bis_algorithm such as bis_imagetoimagealgorithm
which serves as a base for algorithms which take a single image as an input and produce a single image as
an output. (5) bis_guicontainer is a derived class of bis_algorithm and serves as a parent class for cre-
139
CHAPTER 15. WRITING YOUR OWN BIOIMAGE SUITE TCL APPLICATION Draft Sep. 2009
ating multi-algorithm containers (e.g. a tabbed-notebook style GUI where each tab is a separate algorithm).
(6) bis_datamodel is a core class for managing all internal data. This class will encapsulate a database interface
to either mySQL or SQLite (using the database interface from VTK 5.2). (7) bis_viewer will be a wrapper
around the existing C++ viewer classes and will also contain tightly bound Tcl code e.g. animation tool, change
background color, atlas tools, landmark controls etc. (8) bis_baseapplication will be a reimplementation of
the current baseapplication class in BioImage Suite 2.x that will take full advantage of the new algorithms. In
essence, the old baseapplication class will be split into the new bis_viewer and bis_application classes for
better maintainability.
We have practically complete working versions of the rst ve core classes (and a large number of their derived
classes for specic algorithms). Prototypes of the other three are in development.
15.5.2 An Example
Figure 15.5 shows the non-rigid registration algorithm in action in its three major incarnations, namely (i)
command-line application, (ii) standalone GUI applet and (iii) GUI component of a larger application. We
emphasize that all three incarnations are generated using a single script/class depending on how it is called.
Figure 15.6 shows a snippet of the source code from an algorithm for image smoothing. This example illustrates
the relative ease with which the packaging comes together. Two key methods need to be dened (Initialize
and Execute see Figure 15.6) and the rest of the infrastructure for turning this into a GUI component, applet
or command-line application is inherited from the base class. We have found this framework to be extremely
benecial for new programmers who can simply focus on the task at-hand (i.e. the algorithm to be implemented,
most likely in C++, and then called by the Execute method) without having to worry about parameter parsing,
GUI generation etc. A typical C++ algorithm can be packaged in this format in less than half-an-hour. Two
other methods commonly overridden are: (i) CreateGUI which allows for the customization of the automatically
generated user interface and (ii) ProgressCallback which can be used to provide display updates during algorithm
execution.
15.5.3 Writing your own Algorithm Class
The rst thing to identify (much like in all object-oriented programming) is if there is a class that does more or
less what you want it to do. In that case, simply copy that class over and modify it! Failing this, the next best
thing is to identify a parent class that will be most helpful.
The key parent classes are:
|bis algorithm.tcl this is the most general base class. Use this only if you have to.
|bis imagetoimagealgorithm.tcl this is the class is to use if your algorithm takes a single image as an
input and outputs a single image.
|bis surfacetosurfacealgorithm.tcl this is a surface to surface base class.
|bis imagetosurfacealgorithm.tcl
--this is a base class for image to surface type processes.
| bis dualimagealgorithm.tcl this is a bass class for one input, two image output type methods.
| bis dualimagetransformationalgorithm.tcl this is a base class where the inputs are two images and
a transformation and the output is an image (e.g. an image reslice operation).
| bis imageandsurfacetosurfacealgorithm.tcl this takes an image and a surface and returns a surface
| bis baseintensityregistration.tcl this is a basis for registration algorithms.
15.5.3.1 Dening Inputs, Outputs and Options
Inputs are input objects e.g. images, surfaces etc. Outputs are output objects. Parameters are small things (e.g.
numbers, strings etc.) which are used to modulate the performance of the algorithm. Some outputs (e.g. text
140
CHAPTER 15. WRITING YOUR OWN BIOIMAGE SUITE TCL APPLICATION Draft Sep. 2009
Figure 15.5: The new algorithm framework. Each algorithm is packaged within an Incr-Tcl class. Depending on its
invocation, this can function as: (i) a standalone application (or applet) as in A, (ii) a command line tool see B or
(iii) a component of a larger application C1,C2. The options of the algorithm are grouped into Standard (see A,C1)
or Advanced C2. Finally, all graphical user interfaces have a Show Command Line option which produces the text
or the command (shown in D) that can be used to run the identical algorithm using the identical script as the GUI
applet/component from the command line.
Figure 15.6: Sample algorithm implementation. A new algorithm usually requires the implementation of two methods.
The rst is Initialize (shown in detail in this gure) where the inputs, outputs and parameters are dened. The second is
Execute (not shown) which simply takes the specied inputs and parameters and runs the actual algorithm to generate the
desired output. Derived classes can have customized graphical user interfaces by overriding the CreateGUI method.
141
CHAPTER 15. WRITING YOUR OWN BIOIMAGE SUITE TCL APPLICATION Draft Sep. 2009
lenames for debug
All inputs and outputs of an algorithm are dened as lists-of-lists.
To dene a list in tcl use something like
|set lst 2 3 4
Similarly a list of lists is dened s
|set lst 1 2 3 a b c
The inputs, outputs and options for each algorithm are dened using three list-of-lists of the same names. The
syntax for each is as follows.
Inputs
set inputs { $list1 $list2 $list3 .... }
where each component (e.g. list1) is also a list containing either 4 or 5 items (the last one is optional and is set
to zero by default). These are
the name of the input. For imagetoimage algorithms which have only one input the name is input image.
The name is used to access the object from the list array using the GetInputObject method from bis base
algorithm. Examples:
set inp [ $this GetInputObject input_image ]
Note that img is of type pxitclimage in this case, e.g. |thisSetInputObjectinput imageimg
description a longer description of the purpose of the input, used for help messages and documentation
objectype the type of the object. Currently this is one of pxitclimage, pxitcltransform, pxitclsurface,
pxitcllandmarks, pxitclelectrodemultigrid .
lename the initial lename for this object
priority the priority value of the object the absolute value of which determines the order in which gui
elements are created etc. Lower value is higher priority this is intuitive if you are used to UNIX! If the
priority greater or equal to 100 then this is marked as an optional input. If the object is not properly dened
then an error is given then algorithm execution is aborted. An example is if the image dimensions are 1x1x1
suggesting that the image has not been loaded etc.
See the denitions in bis object.tcl for more details. An example is as follows:
set inputs {
{ reference_image "Reference image" pxitclimage "" 2 }
{ target_image "Target Image" pxitclimage "" 4 }
{ initial "Initial Transformation" pxitcltransform "" 120 }
}
}
Here we have three inputs, the rst two are images (and must exist as the priority 100 ) where as the third one
is the intitial transformation which is optional (priority = 100)
142
CHAPTER 15. WRITING YOUR OWN BIOIMAGE SUITE TCL APPLICATION Draft Sep. 2009
Outputs These are stored in a list-of-lists called outputs. The syntax is identical to the inputs array. The user
can access this using the convenience method GetOutputObject e.g.
set out [ $this GetOutputObject "ouput_image" ]
Options These are used to speicify the parameters of the object. They are stored (as inputs and outputs) in a
list-of-lists, in this case by the name of options. Each option is specied using the following components.
name this is the name of the option on the commandline. E.g. if the commandswitch is smooth the
user can set the value on the command line using, for example, -smooth 2.0. The value of the name is also
used to access the option value using convenience methods dened in bis basealgorithm.
thisSetOptionV aluesmooth2.0setsm[this GetOptionValue smooth ]
See the denitions in bis option.tcl for more details.
Description this is used for help messages/documentation
Short Description this is used to identify the option on a graphical user interface.
OptionType this can have up to three components and species (i) the option type, (ii) the option type
hint and (iii) the width of the option on the GUI. If a single value is specied (i.e. a string) then this is
used to set the option type and the hing and width are set to default values (most common case). The
types (with associated optionhints) are:
1. listofvalues this option can take one value from a list of permitted ones. The allowed optionhints
are default and radiobuttons these hints are used to create the appropriate graphical user element
if needed.
2. boolean either 1 or 0. The only allowed optionhints is default.
3. real a real valued number. The allowed optionhints are default, scale and triplescale (a scale with
a text box). these hints are used to create the appropriate graphical user element if needed.
4. integer an integer number same hints as for real.
5. string an abitrary string The only allowed optio hints is default.
6. lename a string that represents either a le, or a directory name. The allowed optionhints are
default (used for les to read), writele and directory./OL
7. comment a sting that gives the user a comment on the algorithm, typically used to describe the
inputs for multi-input algorithm a string that gives more detailed comment.
Should a bad optiion hint be specied, the default value is used.
Default Value the default value of the option
Value Range this list determines the allowed range for the parameter. For boolean options this is either 0
or 1. For real and integer options this list is set to determine the minimum and maximum allowed values.
For listofvalues options, this is a list of all allowed values. Finally for lenames this contains restrictions on
the lename type [Need to provide more details on this one ]
Priority this determines the order of appearance of the option on the graphical user interface. If it is
negative then this option is labeled as Advanced.
Here is an example
set options {
{ startx "Start Value of Crop Region in X" "Start X"
{ integer default } 0 { -10 100 } 0 }
{ blank "If set to On perform blank instead of crop" "Blank"
{ listofvalues radiobuttons } Off { On Off } 12 }
{ orientation "orientation of image axial,coronal,sagittal"
"Orientation"
{ listofvalues radiobuttons } axial { axial coronal sagittal } 0 }
143
CHAPTER 15. WRITING YOUR OWN BIOIMAGE SUITE TCL APPLICATION Draft Sep. 2009
{ detail "description length for header" "Detail"
{listofvalues radiobuttons} short {short long full } 0 }
{ filename "specify the list of images to print the header" "Image File Name"
{ filename default 40 } "" { image } 1 }
{ guicmt "Order of Inputs" "Note" comment "Takes in one image as input"
"The input image will be cropped" 400 }
}
15.5.4 A Full Example bis smoothimage.tcl
Boring Boilerplate start stu
#!/bin/sh
# the next line restarts using wish \
exec vtk "$0" "$@"
lappend auto_path [ file dirname [ info script ]]
lappend auto_path [file join [file join [ file dirname [ info script ]] ".." ] base]
lappend auto_path [file join [file join [ file dirname [ info script ]] ".." ] apps]
Class Denition Each class needs at least three methods (in addition to the constructor). The Initialize
methods is used to dene the lists of inputs, outputs and options. This ends by calling the initialize method of
its parent class which will append to these lists and then go on to initialize everything
The GetGUIName method simply gives the English name for the class
The Execute method is where all the action is!
package provide bis_smoothimage 1.0
package require bis_imagetoimagealgorithm 1.0
itcl::class bis_smoothimage {
inherit bis_imagetoimagealgorithm
constructor { } { $this Initialize }
public method Initialize { }
public method Execute { }
public method GetGUIName { } { return "Smooth Image" }
}
The Initialize Method
144
CHAPTER 15. WRITING YOUR OWN BIOIMAGE SUITE TCL APPLICATION Draft Sep. 2009
itcl::body bis_smoothimage::Initialize { } {
#commandswitch,description,shortdescription,optiontype,defaultvalue,valuerange,priority
set options {
{ blursigma "kernel size [mm/voxel] of FWHM filter size"
"Filter Size" { real triplescale 100 } 2.0 { 0.0 20.0 } 0 }
{ unit "kernel size unit mm or voxels " "Units"
{ listofvalues radiobuttons } mm { mm voxels } 1}
{ radius "radius factor of the gaussian in voxel units "
"Filter Radius" real 1.5 { 0.0 5.0 } -1 }
{ dimension "2 or 3 to to do smoothing in 2D or 3D"
"Dimensionality" { listofvalues radiobuttons } 3 { 2 3 } -999 }
}
set defaultsuffix { "_sm" }
set scriptname bis_smoothimage
set completionstatus "Done"
#
#document
#
set description "Smoothes an image with a specific gaussian kernel."
set description2 "Smoothing kernel size blursigma (in mm by default ) represents the FWHM filter size."
set backwardcompatibility "Reimplemented from pxmat_smoothimage.tcl. unit and radius options are added.\
Multiple image processing eliminated,which will be recovered upon request. "
$this InitializeImageToImageAlgorithm
}
The Execute Method This is where the actual code is.
itcl::body bis_smoothimage::Execute { } {
Part 1 get the parameters and inputs
set blursigma [ $OptionsArray(blursigma) GetValue ]
set unit [ $OptionsArray(unit) GetValue ]
set radius [ $OptionsArray(radius) GetValue ]
set dimension [ $OptionsArray(dimension) GetValue ]
Next get the actual input image. This is of type pxitclimage
set image_in [ $this GetInput ]
# To get the spacing first we need a pointer to the
# encapsulated vtkImageData obtained using the
# GetImage method of pxitclimage
set spacing [[ $image_in GetImage ] GetSpacing ]
145
CHAPTER 15. WRITING YOUR OWN BIOIMAGE SUITE TCL APPLICATION Draft Sep. 2009
# Stuff to compute proper smoothness kernels if unit is voxels or mm
if { $unit == "voxels"} {
for { set j 0 } { $j <=2 } { incr j } {
set sigma($j) [ expr $blursigma * 0.4247 / [ lindex $spacing $j ]]
}
} else {
for { set j 0 } { $j <=2 } { incr j } {
set sigma($j) [ expr $blursigma * 0.4247 ]
}
}
set radiusz $radius
if { $dimension == 2 } {
set radiusz 0
set sigma(2) 0.0
}
This is the actual VTK pipeline code:
# Actual vtk code
set smooth [ vtkImageGaussianSmooth [ pxvtable::vnewobj ] ]
$smooth SetStandardDeviations $sigma(0) $sigma(1) $sigma(2)
$smooth SetRadiusFactors $radius $radius $radiusz
$smooth SetInput [ $image_in GetObject ]
$this SetFilterCallbacks $smooth "Smoothing Image"
$smooth Update
Next we store the output.
# When done store the output of the vtk pipeline in the Output Object
set outimage [ $OutputsArray(output_image) GetObject ]
$outimage ShallowCopyImage [ $smooth GetOutput ]
$outimage CopyImageHeader [ $image_in GetImageHeader ]
# Add a comment to the image header (if NIFTI!) for later reference
set comment [ format " [ $this GetCommandLine full ]" ]
[ $outimage GetImageHeader ] AddComment "$comment $Log" 0
# Clean up
$smooth Delete
return 1
}
Magic Stu at the bottom This checks if executable is called (in this case bis smoothimage.tcl) if it is
execute
if { [ file rootname $argv0 ] == [ file rootname [ info script ] ] } {
146
CHAPTER 15. WRITING YOUR OWN BIOIMAGE SUITE TCL APPLICATION Draft Sep. 2009
# this is essentially the main function
set alg [bis_smoothimage [pxvtable::vnewobj]]
$alg MainFunction
}
15.5.5 Execution Model
An algorithm can be execute in one of four ways:
1. Directly from other code
2. From the Command Line as a command line process.
3. From the commandline as a graphical applet.
4. From the Graphical user interface as a managed tool
Consider bis smoothimage as an example, which is derived from bis imagetoimagealgorithm.
15.5.5.1 Directly from other code
The following example is from bis edgeimage. This is probably the least interesting.
set smooth_alg [bis_smoothimage [pxvtable::vnewobj]]
$smooth_alg InitializeFromContainer $this
$smooth_alg SetInput $image_in
$smooth_alg SetOptionValue blursigma $blursigma
$smooth_alg SetOptionValue unit $unit
$smooth_alg Execute
set outimage [ [ $smooth_alg GetOutput ] GetImage ]
Here image in and outimage are of type pxitclimage.
The SetInput and GetOutput methods are convenience methods dened in bis imagetoimagealgorithm. In the
generic case one would have to use
$smooth_alg SetInputObject input_image $image_in
set outimage [ $smooth_alg GetOutputObject output_image $image_out
The parameters of the algorithm can be set using the SetOptionValue method.
The Execute method is where all the action is for any given algorithm, i.e. this is where the actual code is used
to take the inputs and options and produce the outputs. This is the key method in writing a new bis algorithm
derived class.
The key to understanding the rest of the bis algorithm architecture is to appreciate how the Execute method is
called indirectly in the other scenaria. More on this next.
Note: If you dont have a container, i.e. the code is being invoked from other TCL code use 0 as the value of
containerobject i.e.
147
CHAPTER 15. WRITING YOUR OWN BIOIMAGE SUITE TCL APPLICATION Draft Sep. 2009
$smooth_alg InitializeFromContainer 0
15.5.5.2 From the Command Line as a command line process
If the -dogui ag is set to zero (default) the bis algorithm class becomes a command line application without a
graphical user interface.
In this case the chain of command is as follows.
1. bis algorithm::MainFunction this acts (as the name implies) as the input point to the algorithm
2. bis algorithm::RunCommandLine this is the main function for running on a command line we will look
at this in more detail.
itcl::body bis_algorithm::RunCommandLine { parameterlist commandline } {
$this SetGUIMode commandline
set ret [ $this SetArguments $parameterlist $commandline ]
if { $ret == 1 } { set ret [ $this LoadObjects ] }
if { $ret == 1 } { set ret [ $this CheckInputObjects ] }
if { $ret == 1 } { set ret [ $this Execute ] }
if { $ret == 1 } { set ret [ $this SaveObjects ] }
if { $ret == 0 } { PrintUsage }
$this CleanAll
}
The rst call is to SetArguments which does a number of critical things
1. Calls SetParameters this initializes the OptionsArray where all options are stored, sets them to default
values and then takes any values set on the command line and stores them in the appropriate places.
2. Calls SetFileLists this creates the default input le names (from the commandline) and output le names
(using default rules) (input lelist, output lelist)
3. Calls SetInputArray this initializes the InputsArray which is an associative array of all input objects and
sets the lenames from the input lelist
4. Calls SetOutputArray this initializes the OutputsArray (the mirror of InputsArray) using lenames from
output lelist.
The rest of the sequence is obvious
1. All Input Objects are loaded using LoadObjects
2. The Inputs (with priority 100) are checked for existence ....
3. The algorithm is invoked using Execute
4. The Outputs are saved using SaveObjects
15.5.5.3 From the commandline as a graphical applet
If the -dogui at is set to 1 (on the command line), then each bis algorithm can be used to create a standalone
graphical application with menus/viewers etc. In this case the execution order is as follows.
Part 1 Initilize algorithm
148
CHAPTER 15. WRITING YOUR OWN BIOIMAGE SUITE TCL APPLICATION Draft Sep. 2009
set cmdline $argv
$this InitializeStandalone $cmdline
set ret [ $this SetParameters $parameterlist ]
This part is functionally the equivalent of the SetArguments part of run from command line. The dierence here
is the when running from the command line, if an insucient number of arguments is specied (e.g. no input
image lename) then the program gives an error and returns. In this case, the user may with to specify the
lename on the GUI hence no error is given, instead, the InitializeStandalone method adds empty lenames to
the end of the command line to prevent this from resulting an error.
Part 2 Load the Objects
$this LoadObjects
Part 3 Create the Graphical User Interface
set tp [ toplevel .[ pxvtable::vnewobj ] ]
$this CreateGUI $tp
$this UpdateInputsFromContainer
wm title $tp "BioImage Suite::[ $this GetGUIName ] Applet"
wm geometry $tp 1000x600
update idletasks
15.5.5.4 From the Graphical user interface as a managed tool
Container: When the algorithm appears inside a larger user interface it has a container which may also be
derived from bis algorithm (most likely bis guicontainer). The container takes care of feeding inputs to the
algorithm and managing its outputs (i.e. sends them to the common viewer etc.).
The key initial method is InitializeFromContainer
set alg [ [ bis_smoothimage \#auto ] GetThisPointer ]
$alg SetShowOutputsMode 0
$alg InitializeFromContainer $containerobject
InitializeFromContainer is almost identical to InitializeStandalone other than for setting the guimode ag to
managed as opposed to standalone
The next step is also fairly simple it is simply creating the GUI inside an existing widget $w
$alg SetGUIMode managed
149
CHAPTER 15. WRITING YOUR OWN BIOIMAGE SUITE TCL APPLICATION Draft Sep. 2009
$alg CreateGUI $w
Invoking from a GUI either standalone or as a managed tool At this point the program sits until
the algorithm is invoked using a button which calls InvokeAlgorithmFromGUI.
This is worth looking at in detail. It has four phases.
1. If callbackbefore exists call it this is a way for the user to add stu.
2. Update Values from GUI
3. Execute the algorithm using the Execute Method
4. Update the output lenames using UpdateOuputFilenames the default does nothing!
5. 5. Update the Container class (if it exists) (in the case that this is a managed class)
itcl::body bis_algorithm::InvokeAlgorithmFromGUI { } {
if { $callbackbefore !=0 } {
# this is a hook from master object to set filenames eval $callbackbefore
}
# Update Values from GUI set names [array names OptionsArray]
foreach name $names {
set opt $OptionsArray($name)
$opt UpdateFromGUI
}
set ret [ $this CheckInputObjects ]
if { $ret == 0 } {
tk_messageBox -type ok -message "Failed to execute [ $this GetGUIName ] ...\
bad inputs" -title "Error!" -icon error
return 0
}
set ret [ $this Execute ]
if { $ret == 0 } {
tk_messageBox -type ok -message "Failed to execute [ $this GetGUIName ] ... \
something is wrong with the options" -title "Error!" -icon error
return 0
}
$this UpdateOutputFilenames
$this UpdateContainerWithOutput
if { $callbackafter !=0 } {
eval $callbackafter
}
if { $hasoutputguis == 1 } {
foreach name [ array names OutputsArray ] {
set gui [ $OutputsArray($name) GetObjectGUI ]
set obj [ $OutputsArray($name) GetObject ]
$gui SetObject $obj
$gui Update
}
}
}
150
CHAPTER 15. WRITING YOUR OWN BIOIMAGE SUITE TCL APPLICATION Draft Sep. 2009
Each of the following methods
1. Execute
2. UpdateOutputFilenames
3. UpdateContainerWithOutput
can be redened in a derived class to generate more specialized behavior.
Assignment
There is no explicit assignment for this chapter. Understanding the material presented in this lecture might enable
you to complete the nal assignment for easily (and elegantly!)
The nal assignment will consist of:
Implementing a medical image analysis algorithm as a collection of C++ classes.
Creating a graphical user interface for the algorithm in Tcl (perhaps using [Incr Tcl] and Iwidgets)
optimally as BioImage Suite component.
Creating a viewer for displaying the results and continually updating the viewer as the algorithm is running.
The ability to save both the results and the parameters used will be required.
151
Draft Sep. 2009
Chapter 16
BIS Algorithm Framework
This is an overview of the new framework dened to allow easy integration of novel algorithms into BioImage
Suite. The framework is called bis algorithm and was started by Hirohito Okuda and Xenios Papademetris to
repackage all BioImage Suite algorithms into a uniform architecture captured by a set of incr tcl classes.
16.1 Core Classes
There are the following core classes:
bis option this encapsulates an option value (e.g. smoothness factor etc.). An option can have a type of
: listofvalues, boolean, real, integer, or string. There is also support here for creating a GUI element for
this option.
bis object this encapsulates the input and output objects of the algorithms. A bis object encapsulates
an object of type: pxitclimage, pxitcltransform, pxitclsurface, pxitcllandmarks and pxitclelectrodemultigrid.
These objects in turn are wrappers around vtk data structures (sometimes more than one e.g. pxitcltrans-
form).
bis basealgorithm.tcl this is the core algorithm class from which all algorithms are derived. It has all
functionality for manipulating options, inputs and outputs.
bis algorithm this is derived from bis algorithm and adds the functionality needed for taking an algorithm
and making it into a component or an executable.
bis imagetoimagealgorithm a derived class of bis algorithm to serve as a base for algorithms which take
a single image as an input and produce a single image as an output
bis guicontainer a derived class of bis algorithm this serves as a parent class for creating multi-algorithm
containers (e.g. tabbed-notebook style guis where each tab is a separate algorithm).
bis guiimageutility a specialized container for image to image lters
bis workow To Come a serial operation of multiple steps.
bis baseapplication To Come. Application level class for managing the bis datamodel etc.
bis viewer To Come . Wrapper around C++ viewer classes to contain also tightly bound tcl code e.g.
animation tool, change background color, atlas tools, landmark controls etc
bis datamanager To Come . Application for managing lots of data (global sqllite data manager) with GUI
capability for customized queries.
152
CHAPTER 16. BIS ALGORITHM FRAMEWORK Draft Sep. 2009
Figure 16.1: This gure provides an overview of the BIS Algorithm framework. The inputs can be
any of image, surface or shapes. The options/parameters for the algorithm can be one of the options
mentioned. The algorithm can be invoked given the input and input parameters to get the output. The
output can be one of image, surfaces, transformations, shapes, matlab les and so on.
16.2 Dening Inputs/Outputs and Options
These are dened in the constructor phase of each algorithm class as, essentially, Tcl Lists-of-lists.
Tcl List
Remember that to dene a list in tcl use something like
set lst { 2 3 4 }
Similarly a list of lists is dened as
set lst { { 1 2 3 } { a b c } }
The inputs, outputs and options for each algorithm are dened using three list-of-lists of the same names. The
syntax for each is as follows. Figure 16.1 provides an overview of the BIS Algorithm framework.
Inputs
set inputs { $list1 $list2 $list3 .... }
153
CHAPTER 16. BIS ALGORITHM FRAMEWORK Draft Sep. 2009
where each component (e.g. list1) is also a list containing either 4 or 5 items (the last one is optional and is set
to zero by default). These are
name the name of the input. For imagetoimage algorithms which have only one input the name is
input image. The name is used to access the object from the list array using the GetInputObject method
from bis base algorithm. Examples:
set inp [ $this GetInputObject input_image ]
# (img is of type pxitclimage in this case)
$this SetInputObject "input_image" $img
description a longer description of the purpose of the input, used for help messages and documentation
objectype the type of the object. Currently this is one of pxitclimage, pxitcltransform, pxitclsurface,
pxitcllandmarks, pxitclelectrodemultigrid .
lename the initial lename for this object
priority the priority value of the object the absolute value of which determines the order in which gui
elements are created etc. If the priority is greater or equal to zero, this object is checked. If the object is
not properly dened then an error is given then algorithm execution is aborted. An example is if the image
dimensions are 1x1x1 suggesting that the image has not been loaded etc.
See the denitions in bis object.tcl for more details.
An example (from bis register.tcl) is as follows:
set inputs {
{ reference_image "Reference image" pxitclimage "" 2 }
{ target_image "Target Image" pxitclimage "" 4 }
{ initial "Initial Transformation" pxitcltransform "" -6 }
}
}
Here we have three inputs, the rst two are images (and must exist as the priority >= 0) where as the third one
is the intitial transformation which is optional (priority < 0)
Outputs
These are stored in a list-of-lists called outputs. The syntax is identical to the inputs array. The user can access
this using the convenience method GetOutputObject e.g.
set out [ $this GetOutputObject "ouput_image" ]
Dening the Output Filenames: There a couple of ways of doing this.
Essentially the output lename consists of:
path/stem_suffix1_suffix2.extension
154
CHAPTER 16. BIS ALGORITHM FRAMEWORK Draft Sep. 2009
1. .extension is same as input image 1
2. path is either the path of the key input or dened using the option outputpath (if this is not the empty
string which is the default).
3. stem is either derived from the input lename or dened using the option outputstem (if this is not the
empty string which is the default).
4. sux1 is dened from the option outputsux
5. sux2 comes from the defaultsux option of the algorithm and is a preset dened there. There are 2 cases
here
The algorithm has only one output in this case if outputsux is set then the defaultsux is ignored.
The algorithm has multiple ouputs and defaultsux is a list (e.g. stripped brain). In this case the
two are appended.
Options
These are used to speicify the parameters of the object. They are stored (as inputs and outputs) in a list-of-lists,
in this case by the name of options. Each option is specied using the following components.
1. name this is the name of the option on the commandline. E.g. if the commandswitch is smooth the
user can set the value on the command line using, for example, -smooth 2.0. The value of the name is also
used to access the option value using convenience methods dened in bis basealgorithm.
$this SetOptionValue "smooth" 2.0
set sm [ $this GetOptionValue "smooth" ]
See the denitions in bis option.tcl for more details.
2. Description this is used for help messages/documentation
3. Short Description this is used to identify the option on a graphical user interface.
4. OptionType this can have up to three components and species (i) the option type, (ii) the option type
hint and (iii) the width of the option on the GUI. If a single value is specied (i.e. a string) then this is
used to set the option type and the hing and width are set to default values (most common case). The
types (with associated optionhints) are:
listofvalues this option can take one value from a list of permitted ones. The allowed optionhints
are default and radiobuttons these hints are used to create the appropriate graphical user element
if needed.
boolean either 1 or 0. The only allowed optionhints is default.
real a real valued number. The allowed optionhints are default, scale and triplescale (a scale with
a text box). these hints are used to create the appropriate graphical user element if needed.
integer an integer number same hints as for real.
string an abitrary string The only allowed option hints is default.
lename a string that represents either a le, or a directory name. The allowd optionhints are
default (used for les to read), writele and directory.
Should a bad option hint be specied, the default value is used.
5. Default Value the default value of the option
6. Value Range this list determines the allowed range for the parameter. For boolean options this is either 0
or 1. For real and integer options this list is set to determine the minimum and maximum allowed values.
For listofvalues options, this is a list of all allowed values. Finally for lenames this contains restrictions on
the lename type [Need to provide more details on this one ]
155
CHAPTER 16. BIS ALGORITHM FRAMEWORK Draft Sep. 2009
7. Priority this determines the order of appearance of the option on the graphical user interface. If it is
negative then this option is labeled as Advanced.
For an example see
set options {
{ startx "Start Value of Crop Region in X" "Start X" { integer default } \\
0 { -10 100 } 0 }
{ blank "If set to On perform blank instead of crop" "Blank" \\
{ listofvalues radiobuttons } Off { On Off } 12 }
{ orientation "orientation of image axial,coronal,sagittal" "Orientation" \\
{ listofvalues radiobuttons } axial { axial coronal sagittal } 0 }
{ detail "description length for header" "Detail" {listofvalues radiobuttons} \\
short {short long full } 0 }
{ filename "specify the list of images to print the header" "Image File Name" \\
{ filename default 40 } "" { image } 1 }
}
}
16.3 How does invoking an algorithm work
An algorithm can be invoked in multiple ways using the BIS Algorithm framework. Here we discuss the four ways
in which an algorithm can be invoked. Figure 16.2 provides a schematic showing the four ways that an algorithm
can be invoked.
Consider bis smoothimage as an example.
This is derived from bis imagetoimagealgorithm
It can be invoked in four separate ways:
16.3.1 Directly from other code
The following example is from bis edge.
set smooth_alg [bis_smoothimage [pxvtable::vnewobj]]
$smooth_alg InitializeFromContainer $this
$smooth_alg SetInput $image_in
$smooth_alg SetOptionValue blursigma $blursigma
$smooth_alg SetOptionValue unit $unit
$smooth_alg Execute
set outimage [ [ $smooth_alg GetOutput ] GetImage ]
Here image in and outimage are of type pxitclimage.
The SetInput and GetOutput methods are convenience methods dened in bis imagetoimagealgorithm. In the
generic case on would have to use
156
CHAPTER 16. BIS ALGORITHM FRAMEWORK Draft Sep. 2009
Figure 16.2: An algorithm can be invoked in these four ways in the BIS Algorithm framework. The
simplest way to invoke an algorithm is to invoke it directly in some other code. The algorithm can be
executedly independently as a command line process or as a graphical applet (using the -dogui ag.
The algorithm can also be plugged into a larger container user interface such as the main BioImage
Suite application.
$smooth_alg SetInputObject input_image $image_in
set outimage [ $smooth_alg GetOutputObject output_image $image_out
The parameters of the algorithm can be set using the SetOptionValue method.
The Execute method is where all the action is for any given algorithm, i.e. this is where the actual code is used
to take the inputs and options and produce the outputs. This is the key method in writing a new bis algorithm
derived class.
The key to understanding the rest of the bis algorithm architecture is to appreciate how the Execute method is
called indirectly in the other scenaria.
16.3.2 From the Command Line as a command line process
If the -dogui ag is set to zero (default) the bis algorithm class becomes a command line application without a
graphical user interface.
In this case the chain of command is as follows.
bis algorithm::MainFunction this acts (as the name implies) as the input point to the algorithm
bis algorithm::RunCommandLine this is the main function for running on a command line we will look
at this in more detail.
157
CHAPTER 16. BIS ALGORITHM FRAMEWORK Draft Sep. 2009
itcl::body bis_algorithm::RunCommandLine { parameterlist commandline } {
$this SetGUIMode commandline
set ret 1
set ret [ $this SetArguments $parameterlist $commandline ]
if { $ret == 1 } { set ret [ $this LoadObjects ] }
if { $ret == 1 } { set ret [ $this CheckInputObjects ] }
if { $ret == 1 } { set ret [ $this Execute ] }
if { $ret == 1 } { set ret [ $this SaveObjects ] }
if { $ret == 0 } { PrintUsage }
$this CleanAll
}
The rst call is to SetArguments which does a number of critical things:
Calls SetParameters this initializes the OptionsArray where all options are stored, sets them to default
values and then takes any values set on the command line and stores them in the appropriate places.
Calls SetFileLists this creates the default input le names (from the commandline) and output le names
(using default rules) (input lelist, output lelist)
Calls SetInputArray this initializes the InputsArray which is an associative array of all input objects and
sets the lenames from the input lelist
Calls SetOutputArray this initializes the OutputsArray (the mirror of InputsArray) using lenames from
output lelist.
The rest of the sequence is as follows:
All Input Objects are loaded using LoadObjects
The Inputs (with priority >=0) are checked for existence.
The algorithm is invoked using Execute
The Outputs are saved using SaveObjects
16.3.3 From the commandline as a graphical applet
If the -dogui ag is set to 1 (on the command line), then each bis algorithm can be used to create a standalone
graphical application with menus/viewers etc.
In this case the execution order is as follows.
16.3.3.1 Part 1 Initialize algorithm
set cmdline $argv
$this InitializeStandalone $cmdline
set ret [ $this SetParameters $parameterlist ]
158
CHAPTER 16. BIS ALGORITHM FRAMEWORK Draft Sep. 2009
This part is functionally the equivalent of the SetArguments part of run from command line. The dierence here
is the when running from the command line, if an insucient number of arguments is specied (e.g. no input
image lename) then the program gives an error and returns. In this case, the user may with to specify the
lename on the GUI hence no error is given, instead, the InitializeStandalone method adds empty lenames to
the end of the command line to prevent this from resulting an error.
16.3.3.2 Part 2 Load the Objects
$this LoadObjects
16.3.3.3 Part 3 Create the Graphical User Interface
set tp [ toplevel .[ pxvtable::vnewobj ] ]
$this CreateGUI $tp
$this UpdateInputsFromContainer
wm title $tp "BioImage Suite::[ $this GetGUIName ] Applet"
wm geometry $tp 1000x600
update idletasks
16.3.4 From the Graphical user interface as a managed tool
Container: When the algorithm appears inside a larger user interface it has a container which may also be
derived from bis algorithm (most likely bis guicontainer). The container takes care of feeding inputs to the
algorithm and managing its outputs (i.e. sends them to the common viewer etc.)
The key initial method is InitializeFromContainer
set alg [ [ bis_smoothimage \#auto ] GetThisPointer ]
$alg SetShowOutputsMode 0
$alg InitializeFromContainer $containerobject
InitializeFromContainer is almost identical to InitializeStandalone other than for setting the guimode ag to
managed as opposed to standalone
The next step is also fairly simple it is simply creating the GUI inside an existing widget $w
$alg SetGUIMode managed
$alg CreateGUI $w
16.4 Invoking from a GUI either standalone or as a managed tool
At this point the program sits until the algorithm is invoked using a button which calls InvokeAlgorithmFromGUI.
159
CHAPTER 16. BIS ALGORITHM FRAMEWORK Draft Sep. 2009
This is worth looking at in detail. It has ve phases:
If callbackbefore exists call it this is a way for the user to add stu (callbackbefore should become a
list!)
Update Values from GUI
Execute the algorithm using the Execute Method
Update the output lenames using UpdateOuputFilenames the default does nothing!
Update the Container class (if it exists) (in the case that this is a managed class)
itcl::body bis_algorithm::InvokeAlgorithmFromGUI { } {
if { $callbackbefore !=0 } {
# this is a hook from master object to set filenames
eval $callbackbefore
}
# Update Values from GUI
set names [array names OptionsArray]
foreach name $names {
set opt $OptionsArray($name)
$opt UpdateFromGUI
}
set ret [ $this CheckInputObjects ]
if { $ret == 0 } {
tk_messageBox -type ok -message "Failed to execute [ $this GetGUIName ] ... bad inputs" \\
-title "Error!" -icon error
return 0
}
set ret [ $this Execute ]
if { $ret == 0 } {
tk_messageBox -type ok -message "Failed to execute [ $this GetGUIName ] ... something is wrong with \\
the options" -title "Error!" -icon error
return 0
}
$this UpdateOutputFilenames
$this UpdateContainerWithOutput
if { $callbackafter !=0 } {
eval $callbackafter
}
if { $hasoutputguis == 1 } {
foreach name [ array names OutputsArray ] {
set gui [ $OutputsArray($name) GetObjectGUI ]
set obj [ $OutputsArray($name) GetObject ]
$gui SetObject $obj
$gui Update
160
CHAPTER 16. BIS ALGORITHM FRAMEWORK Draft Sep. 2009
}
}
}
Each of the following methods
Execute
UpdateOutputFilenames
UpdateContainerWithOutput
can be redened in a derived class to generate more specialized behaviour.
16.5 Custom GUIS
Theroretically all that is needed is to ovverride the CreateGUI Method. However this is messy at best.
The key is to redene the components of this method which are:
CreateViewer this creates a viewer if needed the type is set by the variable vtk viewer type
CreateInputGUIInterface widget this creates the input interface inside a pre-existing widget
CreateOptionsGUIInterface widget this creates the interface for the parameters inside a pre-existing widget
this creates all the option guis in a giant list and then passes them to PackOptionsGUIInterfae for packing
or gridding (see bis cropimage.tcl for an example of overdiding this)
CreateOutputGUIInterface as above for the outputs
Note: 1, 2 and 4 are optional steps and may not be invoked depending on the mode.
We need to add a data tree mode here where the inputs are not input guis but tree nodes etc.
16.5.1 Two concrete examples
In this section, we provide two complete examples to demonstrate the use of the BIS Algorithm framework. The
Smoothing algorithm uses vtkImageGaussianSmooth, whereas the Smooth surface example uses the vtkSmooth-
PolyDataFilter or vtkWindowedSincPolyDataFilter class to smooth the polygonal input surface.
16.5.2 A Smoothing Algorithm
The complete code is below.
This is an image-to-image algorithm and hence derives from bis imagetoimagealgorithm.
It denes three methods
GetGUIName this is trivial
Initialize this sets the parameter structure
Execute this is where the algorithm is invoked
161
CHAPTER 16. BIS ALGORITHM FRAMEWORK Draft Sep. 2009
Top part of the le denitions etc.
#!/bin/sh
# the next line restarts using wish \
exec vtk "$0" "$@"
lappend auto_path [ file dirname [ info script ]]
lappend auto_path [file join [file join [ file dirname [ info script ]] ".." ] base]
lappend auto_path [file join [file join [ file dirname [ info script ]] ".." ] apps]
package provide bis_smoothimage 1.0
package require bis_imagetoimagealgorithm 1.0
#
# smooth image
#
itcl::class bis_smoothimage {
inherit bis_imagetoimagealgorithm
constructor { } {
$this Initialize
}
public method Initialize { }
public method Execute { }
public method GetGUIName { } { return "Smooth" }
}
Dening the Options
itcl::body bis_smoothimage::Initialize { } {
PrintDebug "bis_smoothimage::Initialize"
#commandswitch,description,shortdescription,optiontype,defaultvalue,valuerange,priority
set options {
{ blursigma "kernel size [mm/voxel] of FWHM filter size" "Filter Size" \\
{ real triplescale 100 } 2.0 { 0.0 20.0 } 0 }
{ unit "kernel size unit mm or voxels " "Units" { listofvalues radiobuttons } \\
mm { mm voxels } 1}
{ radius "radius factor of the gaussian in voxel units " "Filter Radius" real 1.5 \\
{ 0.0 5.0 } -1 }
{ dimension "2 or 3 to to do smoothin in 2D or 3D" "Dimensionality" \\
{ listofvalues radiobuttons } 3 { 2 3 } -999 }
{ testlog "0: no test log , 1: output test log " "testlog" boolean 0 \\
{ 0 1 } -1000 }
162
CHAPTER 16. BIS ALGORITHM FRAMEWORK Draft Sep. 2009
}
set defaultsuffix { "_sm" }
set defaultoutput 1
set scriptname bis_smoothimage
set completionstatus "Done"
#
#document
#
set description "Smoothes an image with a specific gaussian kernel."
set description2 "Smoothing kernel size blursigma (in mm by default ) \\
represents the FWHM filter size."
set backwardcompatibility "Reimplemented from pxmat_smoothimage.tcl. unit \\
and radius options are added. Multiple image processing eliminated, \\
which will be recovered upon request. "
$this InitializeImageToImageAlgorithm
}
The Execute Method
itcl::body bis_smoothimage::Execute { } {
PrintDebug "bis_smoothimage::Execute"
set blursigma [ $OptionsArray(blursigma) GetValue ]
set unit [ $OptionsArray(unit) GetValue ]
set radius [ $OptionsArray(radius) GetValue ]
set dimension [ $OptionsArray(dimension) GetValue ]
set image_in [ $this GetInput ]
set spacing [[ $image_in GetImage ] GetSpacing ]
if { $unit == "voxels"} {
for { set j 0 } { $j <=2 } { incr j } {
set sigma($j) [ expr $blursigma * 0.4247 / [ lindex $spacing $j ]]
}
} else {
for { set j 0 } { $j <=2 } { incr j } {
set sigma($j) [ expr $blursigma * 0.4247 ]
}
}
set radiusz $radius
if { $dimension == 2 } {
set radiusz 0
set sigma(2) 0.0
}
163
CHAPTER 16. BIS ALGORITHM FRAMEWORK Draft Sep. 2009
PrintDebug "radius = $radius,$radius,$radiusz, sigma = $sigma(0) $sigma(1) $sigma(2)"
set smooth [ vtkImageGaussianSmooth [ pxvtable::vnewobj ] ]
$smooth SetStandardDeviations $sigma(0) $sigma(1) $sigma(2)
$smooth SetRadiusFactors $radius $radius $radiusz
$smooth SetInput [ $image_in GetObject ]
$this SetFilterCallbacks $smooth "Smoothing Image"
$smooth Update
set image_out [ $smooth GetOutput ]
[ $OutputsArray(output_image) GetObject ] ShallowCopyImage $image_out
# Output orientation etc. is same as input
set outimage [ $OutputsArray(output_image) GetObject ]
$outimage CopyImageHeader [ $image_in GetImageHeader ]
set comment [ format "BIS::pxmat_smoothimage.tcl blursigma=%.2f in voxels \\
(%.2f,%.2f,%.2f)" $blursigma $sigma(0) $sigma(1) $sigma(2) $blursigma ]
[ $outimage GetImageHeader ] AddComment "$comment $Log" 0
$smooth Delete
return 1
}
Boilerplate at bottom to make executable
# -----------------------------------------------------------------------------------------
# This checks if executable is called (in this case bis_smoothimage.tcl) if it is execute
# ----------------------------------------------------------------------------------------
if { [ file rootname $argv0 ] == [ file rootname [ info script ] ] } {
# this is essentially the main function
set alg [bis_smoothimage [pxvtable::vnewobj]]
$alg MainFunction
}
16.5.3 Smoothing a surface
This is an example of a surface-to-surface algorithm (surface input, surface output) where a smoothing operation
is performed in the specied input surface.
Top part of the le denitions etc.
164
CHAPTER 16. BIS ALGORITHM FRAMEWORK Draft Sep. 2009
#!/bin/sh
# the next line restarts using wish \
exec vtk "$0" "$@"
lappend auto_path [ file dirname [ info script ]]
lappend auto_path [file join [file join [ file dirname [ info script ]] ".." ] base]
lappend auto_path [file join [file join [ file dirname [ info script ]] ".." ] apps]
package require bis_surfacetosurfacealgorithm 1.0
package provide bis_smoothsurface 1.0
#
# compute curvatures of polydata
#
itcl::class bis_smoothsurface {
inherit bis_surfacetosurfacealgorithm
constructor { } {
$this Initialize
}
public method Initialize { }
public method Execute { }
public method GetGUIName { } { return "Smooth Surface" }
}
Dening the Options
itcl::body bis_smoothsurface::Initialize { } {
PrintDebug "bis_smoothsurface::Initialize"
#commandswitch,description,shortdescription,optiontype,defaultvalue,valuerange,priority
set options {
{ iterations "Number of iterations for smoothing" "Number of iterations" \\
{ integer triplescale 100 } 5 { 0 200 } 0 }
{ relaxationfactor "Specify the relaxation factor for Laplacian smoothing" \\
"Relaxation factor" { real triplescale 100 } 0.2 { 0.0 1.0 } 1 }
{ smoothBoundary "Turn on/off the smoothing of vertices on the boundary of the mesh. " \\
"Smooth Boundaries (Sinc)" { boolean } 0 { 0 1} 2}
{ smoothMethod "Pick a smoothing method: Laplacian or Windowed Sinc. " \\
"Smooth Method" { listofvalues radiobutton} Laplacian { Laplacian Windowed_Sinc} 3}
}
set defaultsuffix { "_smooth" }
set scriptname bis_smoothsurface
#
165
CHAPTER 16. BIS ALGORITHM FRAMEWORK Draft Sep. 2009
#document
#
set category "surface_processing"
set description "smoothing a surface"
set description2 ""
set backwardcompaibitlity "Newly added."
set authors "[email protected]"
$this InitializeSurfacetoSurfaceAlgorithm
}
The Execute Method
itcl::body bis_smoothsurface::Execute { } {
PrintDebug "bis_smoothsurface::Execute"
# Get Inputs
set surface [ $this GetInput ]
# GetOptions
set iterations [ $OptionsArray(iterations) GetValue ]
set relaxationfactor [ $OptionsArray(relaxationfactor) GetValue ]
set smoothBoundary [ $OptionsArray(smoothBoundary) GetValue ]
set smoothMethod [ $OptionsArray(smoothMethod) GetValue ]
# Run Algorithm
if { $smoothMethod == "Laplacian"} {
set filter [ vtkSmoothPolyDataFilter [ pxvtable::vnewobj ] ]
$filter SetInput [ $surface GetObject ]
$filter SetNumberOfIterations $iterations
$filter SetRelaxationFactor $relaxationfactor
$filter Update
} else {
set filter [vtkWindowedSincPolyDataFilter [ pxvtable::vnewobj ] ]
$filter SetInput [ $surface GetObject ]
$filter SetNumberOfIterations $iterations
if { $smoothBoundary == "1"} {
$filter BoundarySmoothingOn
} else {
$filter BoundarySmoothingOff
}
$filter Update
}
# Store Results in output
set surface_out [$filter GetOutput ]
[ $OutputsArray(output_surface) GetObject ] DeepCopySurface $surface_out
166
CHAPTER 16. BIS ALGORITHM FRAMEWORK Draft Sep. 2009
# Clean up
$filter Delete
return 1
}
Boilerplate at bottom to make executable
# -----------------------------------------------------------------------------------------
# This checks if executable is called (in this case bis_smoothsurface.tcl) if it is execute
# ----------------------------------------------------------------------------------------
if { [ file rootname $argv0 ] == [ file rootname [ info script ] ] } {
# this is essentially the main function
set alg [bis_smoothsurface [pxvtable::vnewobj]]
$alg MainFunction
}
16.6 The New Data Model
In the older framework (in BioImage Suite 2.6), there was no explicit data model with the exception of the
datatree. Each component stored its own data in sets of variables and the main applications (derived from
pxitclbaseimageviewer) essentially also kept track of stu in terms of the variables (currentimage,currentresults)
etc.
As part of the redesign of BioImage Suite for version 3.0, we are replacing this with an explicit data model
based on an embedded database (SQLLite as wrapped in VTK). This will enable us to store data/parameter
settings/application status explicitly in a database and then make all data display tools (e.g. the list of bers in
mpjbertracking.tcl) essentially database queries.
16.6.1 Key advantages
This will allow us to save/load the complete (almost complete) state of the application, especially in the event of
a crash. This will also ensure ecient memory management. Additionally, all data is stored centrally and each
algorithms GUI simply stores its results on this (as opposed to keeping things in its private space).
16.6.2 Database Structure Design
The key is to have as many tables as we have data structures. The current list is
images
surfaces
167
CHAPTER 16. BIS ALGORITHM FRAMEWORK Draft Sep. 2009
landmarks
electrodes
electrodegrids
text output (logs) to be added
new objects (i.e. setup les) we may need an explicit setup le class with explicit load/save etc.
Also add a table for bis algorithm derived classes so that we can store their current parameters.
16.6.3 Operation
16.6.3.1 Input Side
All data in the main application is stored physically in the database, and propagated via ShallowCopy operations
as inputs to various algorithms. Danger: We need to ensure that if an object is the equivalent of current image
and is deleted in the database, we clean the private copies. This might be easy as when an image is deleted from
the memory model then something else will become the current image i.e. the next image or the rst image
and this selection should be propagated to all client images. Alternatively UpdateInputsFromContainer should
be replaced by two new methods essentially
UpdateInputsInformationFromContainer this is simply to update Guis e.g. range of threshold scale etc.
UpdateInputsFromContainer this does what it does now and is invoked as part of the InvokeAlgorithmFromGUI
Procedure
16.6.3.2 Output Side
Data on the output side (inputs too but not relevant) will be marked as keep or transient (particularly for images
which are large). The data will also be tagged with the name of the algorithm that produced them
On completion of the algorithm execution:
1. All transient data deriving from the same algorithm will be cleaned out.
2. The new outputs are stored on the database model (i.e. added to the tables)
3. The algorithms own ClearOutputs and ClearInputs methods (to be written and added to the new destructor
in bis basealgorithm) are then used to clean up the internal copies of the outputs and inputs.
168
Draft Sep. 2009
Part V
C++ Techniques
169
Draft Sep. 2009
Chapter 17
Cross-Platform Compiling with CMAKE
CMake, is a cross-platform, open-source make system. It takes as an input as set of, relatively simple, conguration
les and generates as an output native makeles (UNIX) or Visual Studio projects and MSYS makeles (MS-
WINDOWS) for the application. It was developed as part of the NLM Insight Segmentation and Registration
Toolkit (ITK), and it has become another de facto standard in open source medical image analysis software
development.
17.1 Introduction
In the dark old days of software development, writing a program, or a library, that could compile and run on dierent
operating systems was a nightmare even for the most experienced programmer. Large software development (i.e.
anything larger than 3-4 source les) uses some form of batch or automated compilation and build system. The
UNIX standard is something called a makele which is executed using the make command. In most MS-Windows
development, the standard currency is MS Visual Studio project les. The key contribution of CMake is that it
can take a single set of conguration les and generate, depending on the platform, either makeles or Visual
Studio project les, thus eliminating the need for the programmer to do both by hand (with the usual issues of
propagating changes from one to the other as the project evolves).
More information on CMake can be found on its web-page: www.cmake.org. The material in this chapter covers
CMake 2.6, the latest version is 2.7.
17.2 A Simple Introduction
CMake takes as an input a le called CMakeLists.txt. The syntax of this can be a little confusing at rst. For the
most part, however, people simply use existing les and edit them appropriately for their application, so getting
used to this should not take too long. Consider the simple C++ program hello which consists of a single source
le called hello.cpp:
#include <stdio.h>
int main(int argc,char *argv[])
{
fprintf(stderr,"Hello (CMAKE) World!");
return 0;
170
CHAPTER 17. CROSS-PLATFORM COMPILING WITH CMAKE Draft Sep. 2009
Figure 17.1: The CMake conguration program. (Left) Windows version cmakesetup.exe. (Right) Linux version ccmake.
}
To compile this program using CMake we rst need to generate a CMakeLists.txt le in the same directory, which
reads:
# First name the project
PROJECT(HELLOWORLD)
# Define the C++ Files that will go into the library, just one in this case
SET (PROGRAM_SRCS
hello.cpp
)
# Next add an executable
ADD_EXECUTABLE(hello ${PROGRAM_SRCS})
Once this is in place, it needs to be parsed to generate the appropriate development les (makele of Visual
Studio project les). This is accomplished using one of: (i) cmakesetup.exe (Windows), (ii) cmake-gui.exe
(Windows/msys) or (iii) ccmake (Unix), two of which are shown in Figure 17.1.
On Windows the rst step is to specify the source code directory and the build directory it is a good idea to
keep these separate using the two Browse buttons at the top. On Linux, the build directory is assumed to be
the current directory and source directory is specied as an argument. For example, if you have two directories
src and build, which contain the source code and the build les, we execute ccmake as follows:
cd build
ccmake ../src
Working with ccmake involves multiple parsings of the CMakeLists.txt le using the congure option (or c on
Unix). Each time the le is congured CMake may return requests for dierent pieces of information (e.g. where
is Tcl located etc.). Once the necessary information is specied then the OK button (Windows) or the g
(generate and exit) option (Unix) appears, which can be used to complete the process.
171
CHAPTER 17. CROSS-PLATFORM COMPILING WITH CMAKE Draft Sep. 2009
The result of this is either a Visual Studio project le a solution le with the sux .sln or a Unix makele.
1
Building on Linux: The project can be simply build using the make command. This is described in more
detailed in the last section.
Building on Windows: Visual Studio must be rst started. Then open the le HELLOWORLD.sln using
the File/Open Solution option in MS Visual Studio. Then Build this as usual.
17.3 A Second Example
In the programming style that I am attempting to sell in this class, C++ is primarily to be used to create
new Tcl commands which can then be invoked from Tcl scripts. VTK has a very sophisticated mechanism for
automatically generating the necessary wrapper code to accomplish this wrapping, which we will leverage later.
However, it is useful to once see the manual version of this process. The process involves creating a dynamic
library (.so or .dll) which will include our new code, which we will then load into the Tcl interpreter.
The new command to be added is a simple hello-world like statement. The C++ code takes the form:
#include "tcl.h"
#include "tclDecls.h"
#include <stdio.h>
#include <string.h>
int HelloTcl(ClientData, Tcl_Interp* interp,int objc, Tcl_Obj *CONST objv[])
{
fprintf(stderr,"Hello From C++\n");
return TCL_OK;
}
The rst part of the program denes the new command HelloTcl. It has a xed parameter structure. When called
it will simply print Hello from C++ and return the appropriate error code (TCL OK) signifying that all is OK.
The next step is to let the Tcl interpreter know that this command is available and to provide the mechanism
for calling it. We need to do two arcane pieces of coding rst: (i) we need to provide Tcl with an initialization
procedure which must be in C linkage to ensure portability (dont worry if you dont understand this!) and
(ii) we need to, in the case of MS-WINDOWS, declare the initialization procedure as dllexport so it is accessible
from outside the DLL.
// Initialization Code
#ifdef _WIN32
#define HELLOTCL_EXPORT __declspec( dllexport )
#else
#define HELLOTCL_EXPORT
#endif
1
On Windows if you are using the MSYS/MINGW combination, then the results look more like the Unix universe.
172
CHAPTER 17. CROSS-PLATFORM COMPILING WITH CMAKE Draft Sep. 2009
extern "C" {
int HELLOTCL_EXPORT Hellotcl_Init(Tcl_Interp* interp);
int HELLOTCL_EXPORT Hellotcl_SafeInit(Tcl_Interp* interp);
}
int Hellotcl_Init(Tcl_Interp* interp)
{
Tcl_CreateObjCommand(interp, "cpphello", HelloTcl,(ClientData) NULL,
(Tcl_CmdDeleteProc *) NULL);
return TCL_OK;
}
int Hellotcl_SafeInit(Tcl_Interp* interp)
{
return Hellotcl_Init(interp);
}
The initialization procedure is called Hellotcl Init, where hellotcl is the name of the dynamic library (hellotcl.dll or
libhellotcl.so). We need two of these procedures for both ordinary and safe-execution (another detail which you
can safely ignore) cases.
The key command is the invocation of the Tcl library function Tcl CreateObjCommand which tells the Tcl
intepreter that when a user invokes the cpphello command, the interpreter must call the HelloTcl function just
dened above.
A First Attempt at a CMakeLists.txt le
PROJECT(HELLOTCL)
# Set this flag variable first
SET (HELLO_TCL_CANBUILD 1)
# Define the C++ Files that will go into the library, just one in this case
SET (Library_SRCS
hellotclcpp.cpp
)
# Set the default location for outputting the library
SET (LIBRARY_OUTPUT_PATH ${HELLOTCL_SOURCE_DIR})
# Look for tcl.h
FIND_PATH(TCL_INCLUDE_PATH tcl.h PATHS)
IF (TCL_INCLUDE_PATH)
INCLUDE_DIRECTORIES(${TCL_INCLUDE_PATH})
FIND_LIBRARY(TCL_LIBRARY NAMES tcl tcl84 tcl8.4)
ELSE (TCL_INCLUDE_PATH)
SET (HELLO_TCL_CANBUILD 0)
ENDIF (TCL_INCLUDE_PATH)
# If tcl.h has been found and also libtcl then we can build
IF (HELLO_TCL_CANBUILD)
# Add a new shared library based on the source files above
ADD_LIBRARY(hellotcl SHARED ${Library_SRCS})
# Ensure that the TCL library is linked to it
TARGET_LINK_LIBRARIES(hellotcl ${TCL_LIBRARY})
173
CHAPTER 17. CROSS-PLATFORM COMPILING WITH CMAKE Draft Sep. 2009
ENDIF (HELLO_TCL_CANBUILD)
Our library has two dependencies: (i) the tcl.h le and (ii) the tcl library.
2
The rst step in the CMakeLists.txt is to
nd where tcl.h is located on the current machine. This is accomplished using the FIND PATH(TCL INCLUDE PATH
tcl.h PATHS) statement. Once tcl.h is specied, its location is added to the include directories used for
compilation using the INCLUDE DIRECTORIES command. There should be no white-space between directives,
e.g. INCLUDE DIRECTORIES, and the parenthesis sign that follows them!
Next, conditional upon nding tcl.h, we look for the Tcl library. This is accomplished using the FIND LIBRARY()
statement. If the Tcl library is found we can now proceed to building our library otherwise we set the CAN BUILD
variable to zero to indicate that a dependency is missing.
To build our library we need four parts: (i) We need to list the les that need to be compiled into the library, this
is accomplished using:
SET (Library_SRCS
hellotclcpp.cpp
)
Next, we, optionally, dene the default library location using:
SET (LIBRARY_OUTPUT_PATH ${HELLOTCL_SOURCE_DIR})
The HELLOTCL SOURCE DIR variable (where HELLOTCL is the project name) points to the current source
tree. On windows, MS Visual Studio, will add a Debug or Release subdirectory under this depending on the type
of build.
The third step involves dening the library itself: small
ADD_LIBRARY(hellotcl SHARED ${Library_SRCS})
The rst argument is the library name, The second is either SHARED or STATIC depending on the type of library,
and th third is the list of C++ les that will be used.
In the nal step we specify any libraries that need to be linked to it, in this case the Tcl library, using:
TARGET_LINK_LIBRARIES(hellotcl ${TCL_LIBRARY})
Then we compile our library using the usual ccmake and make steps.
2
When using BioImage Suite, these les are in /usr/local/bioimagesuite3 base/tcl84/include and
/usr/local/bioimagesuite3 base/tcl84/lib respectively. On MS-Windows replace /usr/local with c:/yale as usual.
174
CHAPTER 17. CROSS-PLATFORM COMPILING WITH CMAKE Draft Sep. 2009
Loading the Library: The use of this small library is demonstrated by the following short script (testhello.tcl):
lappend auto_path [ file dirname [ info script ]]
set fname libhellotcl
if { $tcl_platform(platform) == "windows" } {
set suffix .dll
} else {
set suffix .so
}
set tname [ file normalize ${fname}${suffix} ]
if { [ file exists $tname ] > 0 } {
load $tname
} else {
set tname [ file normalize [ file join debug ${fname}.dll ] ]
load $tname
}
cpphello
exit
The load command is used to load the DLL into the interpreter. This makes the new command cpphello
available which we can then invoke. The lename of the library and its location are slightly dierent on dierent
operating systems/compilers. A more sophisticated example of how to nd the library will be described in two
chapters.
A More Sophisticated CMakeLists.txt le CMake has special macros for nding commonly used packages
such as Tcl, VTK or ITK. In this case we search for the TCL package as follows:
INCLUDE (${CMAKE_ROOT}/Modules/FindTCL.cmake)
IF(TCL_LIBRARY)
INCLUDE_DIRECTORIES(${TCL_INCLUDE_PATH})
ADD_LIBRARY(hellotcl SHARED ${Library_SRCS})
TARGET_LINK_LIBRARIES(hellotcl ${TCL_LIBRARY})
ENDIF(TCL_LIBRARY)
17.4 Some Additional Comments
CMake has lots of other options which we can not cover here. A reference guide is available at
http://cmake.org/HTML/Documentation.html. The following commands are particularly important:
ADD EXECUTABLE: Add an executable to the project using the specied source les.
175
CHAPTER 17. CROSS-PLATFORM COMPILING WITH CMAKE Draft Sep. 2009
ADD LIBRARY: Add a library to the project using the specied source les.
IF..ELSE..ENDIF: These constructs enable conditional execution.
FILE: This enables outputting of information to text les
FIND FILE: Find the full path to a le.
FIND LIBRARY: Find a library.
FIND PACKAGE: Load settings for an external project.
FIND PATH: Find the directory containing a le.
INCLUDE: Read and include CMake code from the given le. This is a way of modularizing complex
CMakeLists.txt les.
LINK DIRECTORIES: Specify directories in which to search for libraries.
LINK LIBRARIES: Link libraries to all targets added later.
MESSAGE: Display a message to the user.
OPTION: Provides an option that the user can optionally select.
PROJECT: Set a name for the entire project.
SET SOURCE FILES PROPERTIES: Source les can have properties that aect how they are built. This
is useful for declaring classes as abstract etc.
TARGET LINK LIBRARIES: Link a target to given libraries.
17.5 Appendix: A Brief Overview of the Make Utility
On Unix systems, CMake results in a makele which needs to be processed with the Unix Make Utility (most
commonly GNU Make). Makeles are primarily used to compile large programs they are a great mechanism for
large system development because they:
Can be used to dene dependencies i.e. le B must be recompiled because le A changed.
Use of multiple processors at the same time i.e. compile two or more les at once.
The standard name for a makele is unsurprisingly makele. Given a makele, the build procedure is executed
using the command:
make
On Linux systems make=gmake i.e. typing gmake is equivalent to typing make this may appear in some
examples. Additional useful ags include:
1. -n - do a dry run i.e. simply print a list of commands to be executed without doing anything
2. -j - specify how many jobs to run at once typically equal to the number of processors available e.g. to
use 2 processors type make -j2.
In addition makeles contain a number of jobs which may be explicitly specied. For example make clean will
clean all results of previous compilations, so that the whole project can be compiled from scratch upon invocation
of the next make command.
Assignment
Ensure that you can compile and run the three examples in the repository (helloworld, cpphelloworld,
cpphelloworld2)
Modify helloworld to print some additional statements.
Add a second function in hellotcl.cpp (e.g. goodbye world) and make this also available as a Tcl command.
176
Draft Sep. 2009
Chapter 18
C++ Techniques and VTK
In this chapter we will quickly review basic C++ coding conventions, with respect to object orientated program-
ming and templates. We also present examples of the use of VTK with C++. This chapter is meant to serve as
a transition from the more Tcl oriented material presented in the past 17 chapters to C++-based programming
that will form the core of the following chapters. All examples are meant to be built using the CMake utility
described in the previous lecture.
18.1 Introduction
C++ has become the de facto choice of programming language for large projects. While other languages such
as Java and C# have gained popularity, especially for web-related programming, C++ is still for all intents and
purposes the premier programming language out there. One of the great strenghts and weaknesses of C++ is
that is a big language which allows for a multitude of programming styles. It can be used as simply a better
version of C. However, C++, when expanded by use of object-oriented and more recently generic programming
techniques, becomes a very dierent beast. We try to discuss some of these aspects in this lecture.
New and Delete: The most common problem in C/C++ that programmers face is memory allocation and
de-allocation. In VTK this is dealt with by the use of reference counted objects, but in many cases one needs to
understand how memory allocation works at the raw level. Memory in C++ is allocated using the new operator.
For example, a single object obj (see later examples) and an array of 10 integers can be allocated using:
obj* newobj= new obj; int *i = new int[10];
Freeing the allocated memory takes two forms depending on whether one is deleting a single object or an array of
objects (or variables). For a single object we use the delete command whereas for an array we use the delete []
command. This is illustrated below:
delete newobj; delete [] i;
177
CHAPTER 18. C++ TECHNIQUES AND VTK Draft Sep. 2009
18.2 Object-Oriented Programming with C++
The Infant Revisited: In chapter 7 we introduced the concept of object-oriented programming using the [
Incr ] Tcl extensions of Tcl. This was done on purpose, so as to introduce a fairly complex concept without
the need to use a heavy duty compiled language, like C++, which adds additional baggage related to compiling
and building programs. In this section we revisit the material of chapter 7 and convert it (mostly line-by-line)
to C++. (All code for this section is in the oopexample subdirectory). The infant class described in chapter 7,
when converted to C++ takes the form:
infant.tcl top portion infant.h
itcl::class Infant { class Infant {
# Class Variables protected:
protected variable myWeight 2.0 // Member Variables
protected variable myName "Anonymous" float myWeight;
char* myName;
# Constructor and Destructor
constructor { newname } {set myName $newname} public:
destructor {set myName "";set myWeight ""} Infant(char* newname);
virtual ~Infant();
# Interface, public methods
public method GetName { } // Interface
public method GetWeight { } virtual char* GetName();
public method SetWeight { wgt } virtual float GetWeight();
public method DailyRoutine { } virtual void SetWeight(float wgt);
public method PrintSelf { } virtual void DailyRoutine();
virtual void PrintSelf();
# Protected methods
protected method Cry { } protected:
}; virtual void Cry();
};
The virtual construct implies that this method may be overridden by a derived class in the future. The exact
behavior of this is harder to explain, but a simple rule of thumb is: make all object methods virtual unless you
have a really good reason not to (in which case you probably also understand exactly what virtual does!). The
class constructor has the same name as the class, e.g. Infant. The destructor takes the name ~classname. It
is a good idea to also make the destructor virtual. Also note that unlike [ Incr Tcl], we do not use a protected
or public attribute at the start of each denition, rather protected and public methods/attributes are declared in
separate sections beginning with the public: or protected: (or private: for that matter) declaration.
Finally in C++ the standard convention is to put the class denition in a header le (.h e.g. infant.h) and the
implementation in a separate le. The C++ implementation for the infant class is in a le infant.cpp which is
presented below:
#include "infant.h"
#include <string.h>
#include <stdio.h>
Infant::Infant(char* newname)
{
this->myWeight=2.0;
this->myName=new char[200];
strncpy(this->myName,newname,200);
178
CHAPTER 18. C++ TECHNIQUES AND VTK Draft Sep. 2009
}
Infant::~Infant() { delete [] this->myName;}
char* Infant::GetName() { return this->myName;}
float Infant::GetWeight() { return this->myWeight;}
void Infant::SetWeight(float wgt) { this->myWeight=wgt;}
void Infant::DailyRoutine()
{
fprintf(stderr,"infant ... Daily Routine Start\n");
this->Cry();
fprintf(stderr,"infant ... Daily Routine End\n");
}
void Infant::Cry() { fprintf(stderr, "infant ... WaWa!\n");}
void Infant::PrintSelf()
{
fprintf(stderr,"infant ... myName = %s\n",this->myName);
fprintf(stderr,"infant ... myWeight = %.2f\n",this->myWeight);
}
Note the use of this-> prex to access all member variables and methods. While this is not strictly necessary, it
is good programming style it is also the standard convention in VTK. The code itself has no complex constructs
worth discussing, make sure you understand what each line does!
The nal step is to instantiate a version of the infant class. We present below the original [ Incr ] Tcl code
side-by-side with the C++ code.
script7-1.tcl main1.cpp
lappend auto_path [ file dirname [ info script ]] #include <stdio.h>
package require Itcl 3.2 #include "infant.h"
package require Infant int main(int argc,char *argv[])
{
set leanboy [ Infant \#auto "A" ] Infant* leanboy=new Infant("A");
$leanboy SetWeight 7.0 leanboy->SetWeight(7.0);
set fatboy [ Infant \#auto "B" ] Infant* fatboy=new Infant("B");
$fatboy SetWeight 11.0 fatboy->SetWeight( 11.0);
puts stderr "\nLets see the details on $leanboy" fprintf(stderr,"\n ... details on leanboy\n");
$leanboy PrintSelf leanboy->PrintSelf();
$leanboy DailyRoutine leanboy->DailyRoutine();
puts stderr "\nLets see the details on $fatboy" fprintf(stderr,"\n ... details on fatboy\n");
$fatboy PrintSelf fatboy->PrintSelf();
itcl::delete object $leanboy delete leanboy;
itcl::delete object $fatboy delete fatboy;
return 0;
179
CHAPTER 18. C++ TECHNIQUES AND VTK Draft Sep. 2009
}
Note that the new object is instantiated by calling its constructor (the new Infant() construct) and deleted
using the delete command.
A Derived Class FourYearOld: [ Incr ] Tcl uses the inherit statement to declare that a class is derived
from another. In C++, by contrast, this is performed at the class denition itself i.e.
class FourYearOld : public Infant This is illustrated below:
fouryearold.tcl top portion fouryearold.h
itcl::class FourYearOld { class FourYearOld : public Infant {
inherit Infant
...
....
}; };
We will not discuss this in any more detail, the code is available in fouryearold.h. The only other point worth
touching upon is calling a method of the parent class explicitly. For example, while FourYearOld redenes the
PrintSelf method, it still needs to call the PrintSelf method from the Infant class. This is accomplished using
::Infant::PrintSelf() as shown below:
void FourYearOld::PrintSelf()
{
::Infant::PrintSelf();
fprintf(stderr, "fouryearold ... myFavoriteFood = %s\n",this->myFavoriteFood);
fprintf(stderr, "fouryearold ... myFavoriteColor = %s\n",this->myFavoriteColor);
}
Static Members and the IrsInfant class: In chapter 7 (script7-4.tcl) we also demonstrate the use of static
or common member functions and variables: functions and variables which belong to the collective as opposed to
any individual object. The syntax in C++ is similar, compare for example:
script10-4.tcl top portion irsinfant.h
itcl::class IrsInfant { class IrsInfant : public Infant
... {
# Common public:
protected common NumInfants 0 static int GetNumberOfInfants();
protected proc NewSSN { } ...
public proc GetNumInfants { }
protected:
// Static Stuff
.... static int NewSSN();
}; static int NumInfants;
};
One minor dierence is that in the C++ case, IrsInfant is derived from Infant, to save coding. Note that static
member variables in [ Incr ] Tcl are declared using the common statement and static functions are declared
using the proc statement (as opposed to method). In C++ both methods and member variable declarations
use the static qualier.
The code of irsinfant.cpp is worth looking at:
180
CHAPTER 18. C++ TECHNIQUES AND VTK Draft Sep. 2009
#include "irsinfant.h"
#include <stdio.h>
// This is critical, it sets the default value of the static variable and in
// some respects creates it.
int IrsInfant::NumInfants=0;
IrsInfant::IrsInfant(char* newname):Infant(newname) {
this->mySSN=IrsInfant::NewSSN();}
int IrsInfant::GetSSN() { return this->mySSN; }
void IrsInfant::PrintSelf() { ::Infant::PrintSelf();
fprintf(stderr, "irsinfant ... mySSN = %d\n",this->mySSN);
}
int IrsInfant::GetNumberOfInfants() { return IrsInfant::NumInfants;}
int IrsInfant::NewSSN() { ++IrsInfant::NumInfants;
return 1000+IrsInfant::NumInfants;}
Note a couple of things: (i) the static member variable NumInfants needs to be explicitly declared in the .cpp
le (otherwise it never exists!). Its value can also be initialized (although beware, if this code is compiled in a
Windows DLL function, this initialization will not occur!) (ii) All static member functions/variables are accessed
using the IrsInfant:: prex whereas ordinary member variables are accessed using this->.
Compiling these examples: The four classes: Infant, FourYearOld, Teenager and IrsInfant are dened in the
le pairs infant.h/infant.cpp, fouryearold.h/fouryearold.cpp etc and can be found in the oopexample directory.
You may notice that at the top and bottom of each header le there is a construct of the form:
#ifndef _IrsInfant
#define _IrsInfant
.... alll code here ...
#endif /* _IrsInfant */
This is a well-known trick for speeding up compilation. It essentially says: If the constant/macro _IrsInfant
is not dened (#ifndef), dene it and process the code following. This is terminated by the #endif statement.
On the other hand if _IrsInfant is dened, that means that the compiler has already parsed this code probably
because irsinfant.h was included from multiple les that are being compiled, hence the compiler should skip the
code below, it has already been compiled! If you use this (which I recommend) make sure that you use a unique
identier (_IrsInfant in this case) for each header le, otherwise you will get unexpected results (unexpected
= bad! when programming is concerned).
The compilation process is based on CMAKE (see last chapter). The CMakeLists.txt le for this set of examples
has the form:
181
CHAPTER 18. C++ TECHNIQUES AND VTK Draft Sep. 2009
PROJECT(DEMO)
# Define the C++ Files that will go into the library, just one in this case
SET (Library_SRCS
infant.cpp
fouryearold.cpp
teenager.cpp
irsinfant.cpp
)
INCLUDE_DIRECTORIES(${DEMO_SOURCE_DIR})
ADD_LIBRARY(mylib STATIC ${Library_SRCS})
ADD_EXECUTABLE(main1 main1.cpp)
TARGET_LINK_LIBRARIES(main1 mylib)
ADD_EXECUTABLE(main2 main2.cpp)
TARGET_LINK_LIBRARIES(main2 mylib)
ADD_EXECUTABLE(main3 main3.cpp)
TARGET_LINK_LIBRARIES(main3 mylib)
ADD_EXECUTABLE(main4 main4.cpp)
TARGET_LINK_LIBRARIES(main4 mylib)
We rst compile the four classes (infant, fouryearold, teenager and irsinfant) into a static library mylib. This is
common programming practice in which utility code is placed into libraries for reuse. There are two main kinds
of libraries, (i) static, which end in .a on UNIX or .lib on Windows, which must be explicitly added to the main
executable at build time, or (ii) dynamic/shared which end in .so on Unix, .dll on Windows (and .dylib on Mac
OS!) which are linked with the main program at run time. Small libraries are usually compiled as static, whereas
larger libraries, especially those shared by multiple programs, are compiled in dynamic format to save disk space
(among other benets). Note also that on Windows, you need static libraries in addition to dynamic libraries for
linking, dynamic libraries alone are insucient.
To create the library we rst put all the C++ les into a list Library_SRCS and then create the library using:
ADD_LIBRARY(mylib STATIC ${Library_SRCS})
The executables main1.cpp to main4.cpp (see their code) are created using the ADD_EXECUTABLE statement, and
the TARGET_LINK_LIBRARIES statement is used to ensure that mylib is linked to each of them.
18.3 VTK with C++
In the previous chapters we have invoked VTK objects from Tcl. In this section we directly instantiate such objects
in C++. (All code for this section is in the vtkexample subdirectory.) The syntax is very similar and it is often
trivial to go back and forth. One caveat, however, is that some functions of VTK objects are only accessible from
C++ as they require the passing of pointers to arrays. We will revisit some of the example scripts from chapter
10 and convert them to C++. The rst example we will look at is script10-1.tcl which maps to example1.cpp of
this section.
example10-1.tcl example1.cpp
lappend auto_path [ file dirname [ info script ]] #include <vtkFloatArray.h>
package require newname
182
CHAPTER 18. C++ TECHNIQUES AND VTK Draft Sep. 2009
wm withdraw . int main(int argc,char *argv[])
{
set arr [ vtkFloatArray [ newname::vnewobj ]] vtkFloatArray* arr=vtkFloatArray::New();
$arr SetNumberOfComponents 2 arr->SetNumberOfComponents(2);
$arr SetNumberOfTuples 12 arr->SetNumberOfTuples(12);
$arr FillComponent 0 0.0 arr->FillComponent(0,0.0);
$arr FillComponent 1 10.0 arr->FillComponent(1,10.0);
$arr SetComponent 10 0 3.0 arr->SetComponent(10, 0, 3.0);
$arr SetTuple2 11 9.0 2.0 arr->SetTuple2(11,9.0, 2.0);
$arr SetComponent 4 4 -2.1 arr->SetComponent(4, 1, -2.1);
puts stdout "The array has" fprintf(stderr,
puts stdout "[ $arr GetNumberOfTuples ] tuples" "Array has %d tuples and %d components\n\n",
puts stdout " and [ $arr GetNumberOfComponents ]" arr->GetNumberOfTuples(),
puts stdout "Here are its contents" arr->GetNumberOfComponents());
fprintf(stderr,"Here are its contents\n");
set nt [ $arr GetNumberOfTuples ] int nt=arr->GetNumberOfTuples();
set nc [ $arr GetNumberOfComponents ] int nc=arr->GetNumberOfComponents();
for { set i 0 } { $i < $nt } { incr i } { for (int i=0;i<nt;i++)
puts -nonewline stdout "Tuple $i : (" {
for { set j 0 } { $j < $nc } { incr j } { fprintf(stderr,"Tuple %d : (",i);
puts -nonewline stdout "\t [ $arr GetComponent $i $j ]" for (int j=0;j<nc;j++)
} fprintf(stderr,"\t %.2f ",arr->GetComponent(i,j));
puts stdout "\t)" fprintf(stderr,"\t)\n");
} }
exit return 0;
}
The rst key item is the creation of a new object. To enable proper use of reference counting in VTK, the
constructor is protected! The new object is constructed by calling a public static member New(). Hence the array
is created using:
vtkFloatArray* arr=vtkFloatArray::New();
Also note that VTK constructors (and the New methods) take no arguments, all parameters must be specied
subsequently. The object is deleted using its delete method (this is not shown in the code) which takes the form:
arr->Delete();
The rest of the code is very similar and it is a simple matter of translating Tcl Syntax to C++ syntax (which is
what I did in creating example1.cpp).
A Second Example: This example (example4.cpp) derives from (script10-5.tcl) in which we construct a cone
and render it on the scene. First we include all the header les for all the VTK classes we are going to invoke:
#include <vtkPoints.h>
#include <vtkCellArray.h>
#include <vtkPolyData.h>
#include <vtkMath.h>
#include <vtkRenderer.h>
#include <vtkRenderWindow.h>
#include <vtkPolyDataMapper.h>
183
CHAPTER 18. C++ TECHNIQUES AND VTK Draft Sep. 2009
#include <vtkProperty.h>
#include <vtkCamera.h>
#include <vtkRenderWindowInteractor.h>
The main function follows. As an aside, since the value of is not dened on many operating systems by default,
one way to get this is to use the static member function of the vtkMath class Pi(). Next we create the points for
the cone as before:
int main(int argc,char *argv[]) {
double pi=vtkMath::Pi().
vtkPoints* pts=vtkPoints::New();
pts->SetNumberOfPoints(9);
for (int i=0;i<=7;i++) {
double rad= 2.0*double(i)*pi/8.0;
double x = 10.0* sin(rad);
double y = 10.0* cos(rad);
pts->SetPoint(i,x,y,0.0);
double* p=pts->GetPoint(i);
fprintf(stdout,"Point %d = (%.2f,%.2f,%.2f)\n",i,p[0],p[1],p[2]);
}
pts->SetPoint(8,0.0,0.0,10.0);
In creating the cells, we can use a shorthand that enables the passing of an array of points to the cell using a
dierent version InsertNextCell command. Recall that in Tcl, each triangle was created using:
$triangles InsertNextCell 3
$triangles InsertCellPoint $p1;
$triangles InsertCellPoint $p2
$triangles InsertCellPoint 8
The C++ code is below:
vtkCellArray* triangles=vtkCellArray::New();
triangles->Allocate(10, 5);
int p[3];
for (int i=0;i<=7;i++) {
p[0]=i; p[1]=i+1; p[2]=8;
if ( p[1] > 7) p[1]=0;
triangles->InsertNextCell(3,p);
fprintf(stdout,"Set Triangle %d = (%d,%d,%d)\n",i,p[0],p[1],p[2]);
}
184
CHAPTER 18. C++ TECHNIQUES AND VTK Draft Sep. 2009
Next we create the surface and do some clean up note the use of the Delete method.
vtkPolyData* surface=vtkPolyData::New();
surface->SetPoints(pts);
surface->SetPolys(triangles);
// Since VTK uses reference counting delete pts and triangles
// as these are counted inside surface
pts->Delete();
triangles->Delete();
The rest of the code deals with the display. It is essentially a line-by-line translation of the original Tcl code and
there is nothing particular worth highlighting.
vtkPolyDataMapper* map=vtkPolyDataMapper::New(); map->SetInput(surface);
//Create the actor and set it to display wireframes
vtkActor* actor=vtkActor::New();
actor->SetMapper(map);
map->Delete();
vtkProperty* property=actor->GetProperty();
property->SetColor(1,1,1); property->SetAmbient(1.0);
property->SetDiffuse(0.0); property->SetSpecular(0.0);
property->SetRepresentationToWireframe();
// Create the renderer
vtkRenderer* ren= vtkRenderer::New(); ren->AddActor(actor);
// Set camera mode to Orthographic as opposed to Perspective
ren->GetActiveCamera()->ParallelProjectionOn();
// Rest is standard window/interactor etc.
// Render Window
vtkRenderWindow* renWin=vtkRenderWindow::New();
renWin->AddRenderer(ren); renWin->SetSize(300,300);
// Interactor
vtkRenderWindowInteractor* iren=vtkRenderWindowInteractor::New();
iren->SetRenderWindow(renWin);
iren->Initialize(); iren->Start();
}
Compiling the examples: As before use CMAKE. The CMakeLists.txt le is:
PROJECT(VTKEXAMPLES)
185
CHAPTER 18. C++ TECHNIQUES AND VTK Draft Sep. 2009
INCLUDE (${CMAKE_ROOT}/Modules/FindVTK.cmake)
FIND_PACKAGE(VTK)
IF (USE_VTK_FILE)
INCLUDE(${USE_VTK_FILE})
ELSE(USE_VTK_FILE)
MESSAGE(FATAL_ERROR,"Please specify VTK_DIR")
ENDIF (USE_VTK_FILE)
LINK_LIBRARIES(
vtkCommon
vtkGraphics
vtkIO
vtkRendering)
ADD_EXECUTABLE(example1 example1.cpp)
ADD_EXECUTABLE(example2 example2.cpp)
ADD_EXECUTABLE(example3 example3.cpp)
ADD_EXECUTABLE(example4 example4.cpp)
There are two interesting aspects of this le. First we need to link against the VTK libraries, which means we
need to rst nd them! This is accomplished using the FIND_PACKAGE(VTK) command. If VTK is not found we
ouput a fatal error message to the user and the process stops.
Once VTK is found, we dene a set of libraries that are to be linked to all executables (and libraries if any
were dened) using the LINK_LIBRARIES command. Finally we add the executables one-by-one. Since they are
not described in this text, we mention here that example2.cpp corresponds to script10-3.tcl and example3.cpp
corresponds to script10-4.tcl.
18.4 Templates
Templates are a means of creating many similar methods from one piece of code. For example consider a method
that returns the maximum value of two inputs. Strictly speaking we would have to write one such function for
each type of variable (e.g. oat, double, int etc.). Templates are a way to generate these functions automatically
from a simple piece of code.
1
The following example (templateexample/main.cpp) illustrates the concept:
#include <stdio.h>
// T stands for the arbitrary type
template<class T>
T maxvalue(T a,T b) {
if (a>=b)
return a;
return b;
}
1
One of the downsides of using templates, is that a syntax error in a templated function will result in multiple compiler
errors, one for each time the compiler created a new specic function from the templated function! Try writing i instead
of if in the code below and see what happens.
186
CHAPTER 18. C++ TECHNIQUES AND VTK Draft Sep. 2009
The function maxvalue is designed to accept values of the arbitrary type T. When invoked (see below) using a
specic type (e.g. double), a new function double maxvalue(double a,double b) is created (automatically)
and used. Hence if there is a syntax error in a templated function, the compiler will return the same error multiple
times, one for each new version of the templated function that it has to create.
The main program is below. We call the same function, once with double arguments and once with integer
arguments.
int main(int argc,char *argv[])
{
double a=1.0,b=2.0;
fprintf(stderr,"Maximum of %f, %f =%f\n",a,b,maxvalue(a,b));
int a1=44,b1=22;
fprintf(stderr,"Maximum of %d, %d =%d\n",a1,b1,maxvalue(a1,b1));
return 0;
}
The CMakeLists.txt le to compile this is trivial but it oers non-trivial functionality.
PROJECT(DEMO)
ADD_EXECUTABLE(main main.cpp)
Templates are used heavily by the C++ Standard Template Library to support many dierent kinds of objects
such as vectors, stacks etc. In addition to templated functions, it is also possible to have templated classes and
derived classes. The use of templates enables a style of programming known as generic programming, which is
heavily invoked in ITK (we may get to this in the nal chapter). VTK does not use templates in any of its
interface code, but templated functions are used as part of the implementation of many lters.
Templated code presents unusual problems to the C++ compiler and many of them are not capable of handling
all features of templated programming correctly. An easy solution to most of these problems is to include all
denitions of templated functions and implementation in a single le essentially put the code in the .h le!
18.5 The C++ String Class and C++ Streams
In many respects the code in this handout is old fashioned (i.e. more C-like than modern C++-line) in that
it still uses C style strings and IO. The standard template library in C++ denes a modern string class and an
alternative, more type-safe, stream based input/output setup (replacing fprintf etc.). We may return to this topic
later in the semester. For now, take a look at:
String class http://www.cprogramming.com/tutorial/string.html
and:
C++ File I/O http://www.cprogramming.com/tutorial/lesson10.html
In general, the tutorial at the Cprogramming.com webpage http://www.cprogramming.com/tutorial/ is a
187
CHAPTER 18. C++ TECHNIQUES AND VTK Draft Sep. 2009
great resource.
Assignment
This is a rehash of Assignments 7 and 10, but using C++ rather than [ Incr ] Tcl. You may simply translate
your previous Tcl scripts to C++. All solutions must be submitted with an appropriate CMakeLists.txt le for
compilation.
Create an additional class (collegestudent) derived from Teenager. Add at least one additional member
variable (e.g. Age) with appropriate Get/Set Methods for modifying and obtaining its value. Appropriately
modify the PrintSelf method to print this also.
Using example2.cpp as a base, change the code to create a helix, i.e. a curve where the z-coordinate
increase with each point as opposed to a constant 0.0. Save the output in a le helix.vtk
Using the last assignment as a base and the display code from example4.cpp (or your own version) write a
script that creates and displays a helix.
Modify the last assignment to display both a helix and a cone. Hint: you will need two actors.
188
Draft Sep. 2009
Part VI
VTK Programming with C++ and Tcl
189
Draft Sep. 2009
Chapter 19
VTK Code in C++: Procedural Code
and Derived Classes
In this section we give examples of some simple VTK classes. In particular we will rst demonstrate how to add
your own procedural code to VTK in this case there is no attempt to integrate into the VTK piepline. Next we
present two simple image-to-image lters that do integrate with the pipeline. The rst lter is one in which the
input image has the same size as the output, whereas in the second lter the size of the output image is dierent
from that of the input.
19.1 Introduction
In this chapter we describe how to write your own C++ classes derived from VTK classes. There are two key
advantages of deriving classes from VTK classes as opposed to simply writing classes from scratch:
The ability to write lters that will integrate into the VTK pipeline
The ability to have your C++ code (if the header is properly written) wrapped for use with Tcl with no
eort on your part.
We will discuss three examples in this class. The rst example shows how to add procedural code to VTK the
key trick being the addition of these procedures as static member functions of a class which serves simply as a
container. This is used in VTK for the math utility class vtkMath, and it is the only way to add procedural code in
pure OOP -languages such as Java. Next we will discuss two image-to-image lters. The rst performs a simple
thresholding operation, whereas the second resamples the image to extract a volume-of-interest (VOI).
19.2 Adding Procedural Code to VTK
We will create a class vtkMyUtility which will act as a container for dierent procedures/functions. These functions
will be added as static members of vtkMyUtility. The reason for going down this route, as opposed to simply
writing procedural code, is that we can leverage the VTK Tcl wrapping code to make our new functions available
from Tcl.
The Interface File: The header le (vtkMyUtility.h) is listed below:
190
CHAPTER 19. VTK CODE IN C++: PROCEDURAL CODE AND DERIVED CLASSESDraft Sep. 2009
#include <vtkObject.h>
class vtkImageData;
class vtkMyUtility : public vtkObject
{
public:
static vtkMyUtility *New();
vtkTypeMacro(vtkMyUtility,vtkObject);
// Description:
// A few simple functions
static double Average(double a,double b);
/* The BTX and ETX directives enclose code that should not be parsed
by vtkWrapTcl to generate Tcl code. Use these to hide stuff that
might confuse the wrappers.
*/
//BTX
static double Average(double* a,int n);
//ETX
static vtkImageData* ThresholdImage(vtkImageData* input,double minv,
double maxv,int binary);
protected:
vtkMyUtility() {};
virtual ~vtkMyUtility() {};
};
The rst thing to notice is that we use vtkObject as a parent class. This is one of the most generic classes in
VTK and does very little (other than some basic denitions of reference counting), but for our purposes thats
OK! The next two methods are the New method and the vtkTypeMacro, both of which are boiler-plate code and
are needed for all VTK classes.
Next we dene the three static methods. The rst two simply perform number averaging, whereas the last one is
a simple image thresholding procedure. All these methods are public as they are meant to be accessed from the
outside.
Finally, in the protected section, we dene the constructor and the destructor. These are both dummy functions
(i.e. they do nothing), but they must be dened as protected to prevent their default use from the outside as
this would break the reference counting scheme.
The Implementation File: The implementation le (vtkMyUtility.cpp) is listed below. The rst part simply
includes the appropriate header les:
#include "vtkMyUtility.h"
#include "vtkObjectFactory.h"
#include <stdio.h>
#include "vtkImageData.h"
#include "vtkDataArray.h"
#include "vtkPointData.h"
191
CHAPTER 19. VTK CODE IN C++: PROCEDURAL CODE AND DERIVED CLASSESDraft Sep. 2009
Next we use a standard VTK macro to generate the static member New function appropriately.
vtkStandardNewMacro(vtkMyUtility);
Next we dene the two average methods, nothing exciting here:
// This function is accessible both from C++ and Tcl
double vtkMyUtility::Average(double a,double b) {
return 0.5*(a+b);
}
// This function is only accessible from C++
double vtkMyUtility::Average(double* a,int n) {
if (a==NULL) return 0.0;
double sum=0.0;
for (int i=0;i<n;i++)
sum+=a[i];
return sum/double(n);
}
Finally the thresholding function. Here, a key step is to make sure that a valid image has been passed as an input.
This is key, our code has to be defensively oriented and assume that the user is out to get us. Next we create
an output image of the same dimensions and type as the input using the CopyStructure and AllocateScalars
methods.
After this, we simply get the two data arrays and do the thresholding voxel-by-voxel:
vtkImageData* vtkMyUtility::ThresholdImage(vtkImageData* input,double minv,
double maxv,int binary)
{
if (input==NULL) {
fprintf(stderr,"Bad Input Image\n"); return NULL;
}
vtkImageData* output=vtkImageData::New();
output->CopyStructure(input);
output->AllocateScalars();
vtkDataArray* inarray=input->GetPointData()->GetScalars();
vtkDataArray* outarray=output->GetPointData()->GetScalars();
int numcomp=inarray->GetNumberOfComponents();
int numvox =inarray->GetNumberOfTuples();
for (int component=0;component<numcomp;component++) {
outarray->FillComponent(component,0.0); // This method sets the specified component to the specified
// value for all tuples in the data array.
for (int i=0;i<numvox;i++) {
double out=0.0;
double v=inarray->GetComponent(i,component);
192
CHAPTER 19. VTK CODE IN C++: PROCEDURAL CODE AND DERIVED CLASSESDraft Sep. 2009
if (v>=minv && v<=maxv) {
if (binary)
out=1.0;
else
out=v;
outarray->SetComponent(i,component,out);
}
}
}
return output;
}
19.3 A Simple Thresholding Filter
While the thresholding function in the previous section is perfectly adequate for most tasks, such code can not
be directly integrated into the usual VTK pipeline as a regular lter.
Interface: To accomplish this we rewrite our thresholding lter as a derived class of vtkSimpleImageToImage-
Filter
1
The header le (vtkImageSimpleThreshold.h) is listed below. First the usual class header etc. Note that
the vtkTypeMacro takes two arguments, the current class names and the name of its parent class.
#include "vtkSimpleImageToImageFilter.h"
#include "vtkImageData.h"
class vtkImageSimpleThreshold : public vtkSimpleImageToImageFilter
{
public:
static vtkImageSimpleThreshold *New();
vtkTypeMacro(vtkImageSimpleThreshold,vtkSimpleImageToImageFilter);
Next we dene some interface members, using the vtkGetMacro and vtkSetMacro macros. These make the parsing
of the C++ code to create Tcl code easier. The vtkGetMacro(VariableName,type) creates the function:
type GetVariableName()
whereas the vtkSetMacro, creates the function:
void SetVariableName(type T) { this->VariableName=T;}
One additional complexity is provided by the vtkSetClampMacro which makes sure that the new variable set in the
valid range (in the case of BinaryOutput below, the range is 0 to 1). If the user species a value below the minimum
or above the maximum allowed values, the variable is set to the minimum or maximum allowed value respectively.
Finally the vtkBooleanMacro creates two functions of the form BinaryOutputOn(),and BinaryOutputOff()
that result in the variable BinaryOutput being set to 1 or zero respectively.
// Description:
1
This very useful teaching tool class has unfortunately been removed from VTK 5.0.
193
CHAPTER 19. VTK CODE IN C++: PROCEDURAL CODE AND DERIVED CLASSESDraft Sep. 2009
// Set the thresholds between which output is 1
vtkGetMacro(LowerThreshold,float);
vtkSetMacro(LowerThreshold,float);
vtkGetMacro(UpperThreshold,float);
vtkSetMacro(UpperThreshold,float);
// Description:
// If Binary = 1, then output is 0,1 else output is 0, original
vtkSetClampMacro(BinaryOutput,int,0,1);
vtkGetMacro(BinaryOutput,int);
vtkBooleanMacro(BinaryOutput,int);
In the protected part of the denition, we rst dene the constructor, which in this case is not a dummy function,
and make sure that things like the copy construrctor and the = operator are protected, to prevent their use
which will mess up the reference counting scheme.
protected:
vtkImageSimpleThreshold();
vtkImageSimpleThreshold(const vtkImageSimpleThreshold&) {};
void operator=(const vtkImageSimpleThreshold&) {};
Finally we dene the SimpleExecute function which will be called when this lter is updated (this is dened in
the parent class) and the three member variables.
// Description:
// Do Thresholding Operation
virtual void SimpleExecute(vtkImageData* inp,vtkImageData* out);
//BTX
float LowerThreshold,UpperThreshold;
int BinaryOutput;
//ETX
};
Implementation: The rst part is the usual, include the header les and call the vtkStandardNewMacro, so
we will skip it. Next we dene the constructor, whose job it is to set the default values for all member variables
and to perform any other initialization:
vtkImageSimpleThreshold::vtkImageSimpleThreshold()
{
this->LowerThreshold=-1.0e+9;
this->UpperThreshold= 1.0e+9;
this->BinaryOutput=1;
194
CHAPTER 19. VTK CODE IN C++: PROCEDURAL CODE AND DERIVED CLASSESDraft Sep. 2009
}
The work of the lter is performed by the SimpleExecute function. This is very similar to the function
vtkMyUtility::ThresholdImage, but here the code is placed in the context of the VTK pipeline. Also since
this is not a static member function, but a proper member function, we can use the vtkErrorMacro to report
errors (e.g. in this case bad input). The code is below:
void vtkImageSimpleThreshold::SimpleExecute(vtkImageData* input,vtkImageData* output )
{
if (input==NULL) {
vtkErrorMacro(<<"Bad Input to vtkImageSimpleThreshold");
return;
}
int dim[3]; input->GetDimensions(dim);
int numvox=dim[0]*dim[1]*dim[2];
int numcomp=input->GetNumberOfScalarComponents();
vtkDataArray* inarray=input->GetPointData()->GetScalars();
vtkDataArray* outarray=output->GetPointData()->GetScalars();
for (int component=0;component<numcomp;component++) {
outarray->FillComponent(component,0.0);
for (int i=0;i<numvox;i++) {
double out=0.0;
double v=inarray->GetComponent(i,component);
if (v>=this->LowerThreshold && v<=this->UpperThreshold) {
if (this->BinaryOutput)
out=1.0;
else
out=v;
outarray->SetComponent(i,component,out);
}
}
}
}
19.4 A Slightly More Complex Filter
In the previous example, we derived the simplest possible type of image-to-image lter, one in which the image
output has the same type and dimensions as the input image. If either of these last two conditions do not hold
(i.e. image dimensions or type are dierent), then we need to let the pipeline controlling processes know, by
overriding the RequestInformation method. As an example, consider the vtkImageExtractVOI class which takes
an image as an input and extracts a piece of it the volume of interest (VOI) as the output. In this case we,
potentially, have a change in image size.
The Interface: The interface (vtkImageExtractVOI.h) is listed below. The only two highlights here are: (i)
the use of vtkGetVectorMacro and vtkSetVectorMacro in the case of xed length arrays, in this case the volume
195
CHAPTER 19. VTK CODE IN C++: PROCEDURAL CODE AND DERIVED CLASSESDraft Sep. 2009
of interest VOI (int[6]) and the sampling rate SampleRate (int[3]) and (ii) the presence of the ExecuteInformation
method.
#include "vtkSimpleImageToImageFilter.h"
class vtkImageExtractVOI : public vtkSimpleImageToImageFilter
{
public:
vtkTypeMacro(vtkImageExtractVOI,vtkSimpleImageToImageFilter);
void PrintSelf(ostream& os, vtkIndent indent);
static vtkImageExtractVOI *New();
// Description:
// Specify i-j-k (min,max) pairs to extract. The resulting structured points
// dataset can be of any topological dimension (i.e. point, line, image or volume).
vtkSetVector6Macro(VOI,int);
vtkGetVectorMacro(VOI,int,6);
// Description:
// Set the sampling rate in the i, j, and k directions.
vtkSetVector3Macro(SampleRate, int);
vtkGetVectorMacro(SampleRate, int, 3);
protected:
vtkImageExtractVOI();
~vtkImageExtractVOI() {};
vtkImageExtractVOI(const vtkImageExtractVOI&) {};
void operator=(const vtkImageExtractVOI&) {};
// Description:
// Execute Stuff
virtual int RequestInformation(vtkInformation *vtkNotUsed(request),
vtkInformationVector **inputVector,
vtkInformationVector *outputVector);
virtual void SimpleExecute(vtkImageData* input, vtkImageData* output);
// Description:
// Data Members
int VOI[6];
int SampleRate[3];
};
Implementation: The listing for vtkImageExtractVOI.cpp is listed below. We skip the header insertion and
standard macro invocation. The constructor sets the initial values in the usual way:
vtkImageExtractVOI::vtkImageExtractVOI()
{
this->VOI[0] = this->VOI[2] = this->VOI[4] = 0;
this->VOI[1] = this->VOI[3] = this->VOI[5] = 1;
196
CHAPTER 19. VTK CODE IN C++: PROCEDURAL CODE AND DERIVED CLASSESDraft Sep. 2009
this->SampleRate[0] = this->SampleRate[1] = this->SampleRate[2] = 1;
}
Next up is the PrintSelf macro, which prints the contents of the member variables. This is very useful for
debugging purposes (we skipped this for the previous two classes which is not ideal!)
void vtkImageExtractVOI::PrintSelf(ostream& os, vtkIndent indent)
{
vtkSimpleImageToImageFilter::PrintSelf(os,indent);
os << indent << " VOI " << this->VOI[0] << ":" << this->VOI[1] << " ";
os << indent << " VOI " << this->VOI[2] << ":" << this->VOI[3] << " ";
os << indent << " VOI " << this->VOI[4] << ":" << this->VOI[5] << "\n";
os << indent << " SampleRate "<<this->SampleRate[0]<<
os << ","<<this->SampleRate[1]<<","<<this->SampleRate[2]<<"\n";
}
Implementation Request Information: Next up is the RequestInformation Method. This is used to alter
the default output settings. We will rst look at this concrete example and then list the most common options.
Do the defaults. Then we set the Spacing, Origin, Dimensions and the WholeExtent (an array of six numbers 0,
dim[0]-1,0,dim[1]-1,0,dim[2]-1) of the output image to reect the user settings for the VOI and the SampleRate.
vtkInformation *inInfo = inputVector[0]->GetInformationObject(0);
vtkInformation *outInfo = outputVector->GetInformationObject(0);
int wholeExtent[6]; inInfo->Get(vtkStreamingDemandDrivenPipeline::WHOLE_EXTENT(), wholeExtent);
double sp[3]; inInfo->Get(vtkDataObject::SPACING(), sp);
double origin[3]; inInfo->Get(vtkDataObject::ORIGIN(), origin);
vtkImageData* input=this->GetImageDataInput(0);
input->GetExtent(wholeExtent);
input->GetSpacing(sp);
input->GetOrigin(origin);
int nc=input->GetNumberOfScalarComponents();
int voi[6],rate[3],outDims[3];
for (int i=0; i < 6; i++ )
voi[i] = this->VOI[i];
for (int i=0; i < 3; i++ )
{
if ( voi[2*i+1] > wholeExtent[2*i+1] )
voi[2*i+1] = wholeExtent[2*i+1];
else if ( voi[2*i+1] < wholeExtent[2*i] )
voi[2*i+1] = wholeExtent[2*i];
if ( voi[2*i] < wholeExtent[2*i] )
voi[2*i] = wholeExtent[2*i];
else if ( voi[2*i] > wholeExtent[2*i+1] )
197
CHAPTER 19. VTK CODE IN C++: PROCEDURAL CODE AND DERIVED CLASSESDraft Sep. 2009
voi[2*i] = wholeExtent[2*i+1];
if ( voi[2*i] > voi[2*i+1] )
voi[2*i] = voi[2*i+1];
if ( (rate[i] = this->SampleRate[i]) < 1 )
rate[i] = 1;
outDims[i] = (voi[2*i+1] - voi[2*i]) / rate[i] + 1;
if ( outDims[i] < 1 )
outDims[i] = 1;
}
// This makes sense for sample rates of 1, 1, 1.
wholeExtent[0] = 0; wholeExtent[1] = outDims[0] - 1;
wholeExtent[2] = 0; wholeExtent[3] = outDims[1] - 1;
wholeExtent[4] = 0; wholeExtent[5] = outDims[2] - 1;
outInfo->Set(vtkStreamingDemandDrivenPipeline::WHOLE_EXTENT(),wholeExtent,6);
double osp[3],outori[3];
for (int ia=0;ia<=2;ia++)
{
osp[ia]=sp[ia]*double(rate[ia]);
outori[ia]=double(voi[ia])*sp[0]+origin[0];
}
outInfo->Set(vtkDataObject::SPACING(), osp, 3);
outInfo->Set(vtkDataObject::ORIGIN(), outori, 3);
Output Information Options
The most common output options that are changed from input to output in image-to-image type lters are:
1. Output Image Dimensions :
outInfo->Set(vtkStreamingDemandDrivenPipeline::WHOLE_EXTENT(),wholeExtent,6);
2. Output Image Spacing :
outInfo->Set(vtkDataObject::SPACING(), osp, 3);
3. Output Image Origin :
outInfo->Set(vtkDataObject::ORIGIN(), outori, 3);
4. Output Image Number Of Components (Frames):
vtkDataObject::SetPointDataActiveScalarInfo(outputVector->GetInformationObject(0), -1, numframes);
5. Output Image Data Type
vtkDataObject::SetPointDataActiveScalarInfo(outInfo, VTK_DOUBLE, -1);
Implementation SimpleExecute: Finally the SimpleExecute method itself. Here we cheat and simply
call vtkImageReslice, since it can do the job just ne.
void vtkImageExtractVOI::SimpleExecute(vtkImageData* input, vtkImageData* output)
{
if (input==NULL) { vtkErrorMacro(<< "Bad Input"); return; }
int dim[3]; output->GetDimensions(dim);
vtkImageReslice* resl=vtkImageReslice::New();
resl->SetInput(input);
198
CHAPTER 19. VTK CODE IN C++: PROCEDURAL CODE AND DERIVED CLASSESDraft Sep. 2009
resl->SetOutputOrigin(output->GetOrigin());
resl->SetOutputSpacing(output->GetSpacing());
resl->SetInterpolationMode(0);
resl->SetOutputExtent(0,dim[0]-1,0,dim[1]-1,0,dim[2]-1);
resl->OptimizationOff();
resl->Update();
output->ShallowCopy(resl->GetOutput());
resl->Delete();
}
19.5 Compiling and Using
The CMakeLists.txt for this project is below. The rst part is fairly standard:
# create the vtk executable
PROJECT(MYLIB)
SET(KITBASE MyLib)
SET(KIT vtk${KITBASE})
INCLUDE (${CMAKE_ROOT}/Modules/FindVTK.cmake)
FIND_PACKAGE(VTK REQUIRED)
IF (USE_VTK_FILE)
INCLUDE(${USE_VTK_FILE})
ENDIF (USE_VTK_FILE)
INCLUDE_DIRECTORIES(${MYLIB_SOURCE_DIR})
# Set the default location for outputting the library
SET (LIBRARY_OUTPUT_PATH ${MYLIB_SOURCE_DIR})
SET(LIBRARY_SRCS
vtkMyUtility.cpp
vtkImageSimpleThreshold.cpp
vtkImageExtractVOI.cpp)
LINK_LIBRARIES(vtkCommon vtkCommonTCL vtkImaging vtkImagingTCL)
ADD_LIBRARY(${KIT} STATIC ${LIBRARY_SRCS})
The exciting part comes at the end. First we use the VTK_WRAP_TCL2 command to wrap all the header les in
LIBRARY_SRCS to create Tcl interface code. Next we add one more library that is the dynamic library that can
be loaded into the vtk interpreter. This is similar to the cpphelloworld example of chapter 17.
INCLUDE(${VTK_CMAKE_DIR}/vtkWrapTcl.cmake)
VTK_WRAP_TCL2 (${KIT}TCL LIBRARY_TCL_SRCS ${LIBRARY_SRCS})
ADD_LIBRARY (${KIT}TCL SHARED ${LIBRARY_TCL_SRCS} ${LIBRARY_SRCS})
199
CHAPTER 19. VTK CODE IN C++: PROCEDURAL CODE AND DERIVED CLASSESDraft Sep. 2009
Example 1: Once the library exists it can be loaded and exercised. See scripts19-1.tcl, script19-2.tcl script19-
3.tcl and script19-4.tcl for examples. The simplest on of these (script19-1.tcl) is listed below:
lappend auto_path [ file dirname [ info script ]]
package require mylibload 1.0
mylibload vtkMyLibTCL
set util [ vtkMyUtility [ newname::vnewobj ] ]
set out [ $util Average 4 11 ]
puts stderr "The average of 4 and 11 is $out"
exit
The library is loaded using the load command, through the script mylibload.tcl this is messy and not worth
discussing in detail here.. Then a new vtkMyUtility object is created. This is then used to invoke the average
command. (While in C++ we do not need an instance of the object to invoke a static member function, e.g. the
Average function could have been invoked as vtkMyUtility::Average(4,11), in Tcl this is necessary most
likely a limitation of the VTK Tcl Wrappers.
Example 2: The following script (script19-3.tcl) illustrates the use of the vtkImageSimpleThereshold class to
threshold an image and generate an output binary mask:
lappend auto_path [ file dirname [ info script ]]
package require loadbioimagesuite 1.0
package require mylibload 1.0
mylibload vtkMyLibTCL
set ana [ vtkpxAnalyzeImageSource [ pxvtable::vnewobj ]]
$ana Load axial.hdr
set img [ $ana GetOutput ]
set orient [ $ana GetOrientation ]
puts stderr "Image Dimensions = [ $img GetDimensions ], Orientation = $orient"
set thr [ vtkImageSimpleThreshold [ pxvtable::vnewobj ]]
$thr SetInput $img
$thr SetLowerThreshold 150
$thr SetUpperThreshold 250
$thr BinaryOutputOn
$thr Update
set anaw [ vtkpxAnalyzeImageWriter [ pxvtable::vnewobj ]]
$anaw SetInput [ $thr GetOutput ]
$anaw SetOrientation $orient
$anaw Save thr.hdr
200
CHAPTER 19. VTK CODE IN C++: PROCEDURAL CODE AND DERIVED CLASSESDraft Sep. 2009
Both these examples fully illustrate the dual-language approach - write in C++, use in Tcl.
19.6 Overall Comments
In general the style of programming that I am advocating (and this class is designed to teach) is one in which
core components/classes are written in C++ for reasons of eciency and reuse. Such components are then to
be packaged into shared libraries and wrapped such that they can be accessed from Tcl. The main program,
including the user interface should most likely be written in Tcl optimally using the [ Incr Tcl ] extensions
which ensures both platform portability as well as relative ease of development. It is fairly common to observe
that the majority of tweaking of a piece of software has to do with the user interface, so the use of an interpreted
language such as Tcl can have a great potential benet in this respect.
The other major advantage of this approach is the separation between algorithms and applications. In this way
the algorithms (neatly packaged in shared libraries) can be reused by other applications by simply changing the
user interface. Essentially a programmer develops: (i) a toolkit - a collection of useful tools and (ii) an application
or set of applications based on the toolkit with additional user interface enhancements. This is the philosophy
used in the design and development of BioImage Suite, for example.
Assignment 19
First ensure that you can compile and run the examples in this handout. Next:
Add two new methods to vtkMyUtility.h. One should compute the average intensity in an image. The
second should return the maximum and minimum intensity in an image. (You should most likely return a
vtkFloatArray in the second case.)
Create a new class (optionally based on vtkImageSimpleThreshold) that takes an image as an input and
two thresholds t
1
and t
2
. The output should be 0 if the intensity is less than t
1
, 1 if the intensity is between
t
1
and t
2
and 2 if the intensity is greater than t
2
. The class should ensure that t
1
< t
2
.
Create a new class (optionally based on vtkImageExtractVOI.h) that samples an image and cuts its size
in half by talking every other voxel in x,y and z-directions respectively. For example a 16 24 32 input
image should result in an 8 12 16 image output.
Add your classes to the CMakeLists.txt le and write Tcl scripts that exercise these new classes.
201
Draft Sep. 2009
Chapter 20
Implementing Point-Based Registration
In this chapter, we will discuss the implementation of one of the most popular surface matching algorithms: the
Iterative Closest Point algorithm (ICP)[?]. Part of the design exercise will involve the division of the implementation
into two parts: (i) an abstract parent class containing functionality common to point-based registration methods
and (ii) a derived class which implements the ICP algorithm. In addition, we will also look at the design of a
graphical user interface for interacting with this algorithm.
20.1 The Iterative Closest Point Algorithm
We review here, briey, the Iterative Closest Point algorithm [?]. This is a registration method whose goal is to
estimate a transformation T that best aligns two point sets X = x
0
, x
1
, . . . , x
n
and Y = y
0
, y
1
, . . . , y
m
, where
in general n ,= m. The exciting part of this algorithm is that it does not assume that the correspondences
between X and Y are known, if they were known, the problem reduces to a simple least squares t of some form
to estimate the transformation. Alternatively, if the transformation T were known, it could be used to estimate
the corresponding points. The way out of this chicken-and-egg problem is to perform an alternating estimation
of (i) the set of correspondences and (ii) the transformation. More formally the algorithm iteratively minimizes
the following two equations (k=number of iteration), until convergence:
Correspondence: c
k
i
=
arg min
yjY
[T
k
(x
i
) y
j
[
2
(20.1)
Transformation: T
k+1
=
arg min
T

n
i=1
[T(x
i
) c
k
i
[
2
(20.2)
In the correspondence estimation step, we take each point x
i
X, and transform it using the current estimate
of the transformation T
k
. Then we look for the point in Y that is closest to T
k
(x
i
). We label this point as the
corresponding point, at iteration k, c
k
i
. This results in a sets of pairs of correspoding points (x
i
, c
k
i
).
In the transformation estimation step, we look for the transformation T that best explains (or describes) this set
of correspondences. T can be either a linear or a non-linear transformation, we will focus on linear transformations
in this chapter.
Once a better version of T is estimated we can use it to estimate a better set of correspondences and so on
until convergence. Note that ICP needs to be initialized fairly close to the true transformation for convergence.
Alternative methods, e.g. the Robust Point Matching method [?] have superior capture range and accuracy.
However, ICP is a simple algorithm which works OK and is a great classroom example.
202
CHAPTER 20. IMPLEMENTING POINT-BASED REGISTRATION Draft Sep. 2009
20.2 Implementation Design
While there is already an existing implementation of ICP in VTK 4.4 vtkIterativeClosestPointTransform
from which some of the code used in our design is based, we will describe a dierent approach here. The goal
of this approach, is to develop both a generic framework for point-based registration methods as well as an
implementation for ICP itself. This is a useful exercise as it demonstrates how object-oriented methodology can
be used to implement a family of algorithms in a clean way.
Generic vs Specic Functionality: In designing our implementation, we can list the following methods and
variables that need to be specied/implemented:
1. Source and Target Data sets
2. Parameters such as the maximum number of iterations, convergence threshold, maximum number of points.
3. Methods for invoking the algorithm and returning the output transformation in an algorithm-independent
way (so the rest of the program can use dierent algorithms with no major changes.)
4. Utility methods for sampling a data-set (to reduce the number of points), estimating the centroid of a data
set and translating points.
5. Parameters specic to the Linear ICP algorithm, such as the exact type of the transformation (rigid,
similarity or ane), the initial step (matching centroids).
6. The method for the main loop of the ICP algorithm.
If we look at this list, one way to break it up is that the rst four items refer to functionality that is generally
useful for point-based registrations, whereas the last two describe linear-ICP specic functionality. A good way to
break up the design, is then to implement rst an abstract class that incorporates items 14 and then derive from
it an ICP implementation that denes 56 and overrides the methods in 3. We discuss both of these class next:
The abstract parent class vtkAbstractPointBasedRegistration
Formally speaking, an abstract class in C++, is one which can never be instantiated. It is simply there as a place
to derive functional classes from. Abstract classes are extremely useful for dening a common interface to more
concrete children classes, such that other parts of the program may interact with any number of these derived
classes using a similar interface.
The Interface: The interface le (vtkAbstractPointBasedRegistration.h) has the following form. We derive
this class from vtkProcessObject which is a fairly lightweight VTK object
1
that does not impose too many
constraints.
class vtkAbstractPointBasedRegistration : public vtkProcessObject
{
public:
vtkTypeRevisionMacro(vtkAbstractPointBasedRegistration,vtkProcessObject);
void PrintSelf(ostream& os, vtkIndent indent);
Next comes the specication of the two data-sets, the source dataset (X) and the target data-set (Y):
1
Perhaps such a class could be also be derived from the new vtkAlgorithm class.
203
CHAPTER 20. IMPLEMENTING POINT-BASED REGISTRATION Draft Sep. 2009
vtkGetObjectMacro(Source, vtkDataSet);
vtkGetObjectMacro(Target, vtkDataSet);
vtkSetObjectMacro(Source, vtkDataSet);
vtkSetObjectMacro(Target, vtkDataSet);
The GetObjectMacro simply creates functions of the form:
vtkDataSet* GetSource() { return this->Source; }
The SetObjectMacro is more interesting. In full this generates a function (we abbreviate somewhat here) that
takes the form:
virtual void SetSource(vtkDataSet* ds) {
if (this->Source!=NULL)
this->Source->UnRegister(this);
this->Source = ds;
if (this->Source != NULL)
this->Source->Register(this);
this->Modified();
}
The Register/UnRegister pair increment and decrement, respectively, the reference count on the objects. In
setting the new source object, we rst decrement the reference count of the previous object used as source. Next
we set the new source object and and then increment its reference counter!
Next come a set of methods for setting and getting the value of various parameters:
vtkSetMacro(MaximumNumberOfIterations, int);
vtkGetMacro(MaximumNumberOfIterations, int);
vtkSetMacro(MaximumNumberOfPoints, int);
vtkGetMacro(MaximumNumberOfPoints, int);
vtkSetMacro(Epsilon, double);
vtkGetMacro(Epsilon, double);
Finally, two extremely important functions, the GetTransformation() and the Run() methods. These are
dened as pure virtual methods (signied by the =0;) at the end, which means two things: (i) this class can
not be instatiated and (ii) any derived class from this class must dene these methods in order to be able to be
instantiated. This is a formal way, in C++, of dening a pure interface for derived classes to implement:
// Get the output transformation. This is purposefully an abstract
// transformation so That derived classes can return different types of transformations
virtual vtkAbstractTransform* GetTransformation() = 0;
// The Run Method Computes the Registration
virtual int Run() = 0;
204
CHAPTER 20. IMPLEMENTING POINT-BASED REGISTRATION Draft Sep. 2009
The rest of the header is fairly trivial:
protected:
vtkAbstractPointBasedRegistration();
virtual ~vtkAbstractPointBasedRegistration();
// Data Members
vtkDataSet* Source,*Target;
int MaximumNumberOfIterations, MaximumNumberOfPoints;
double Epsilon;
// Utility Method
virtual vtkPoints* SampleDataSet(vtkDataSet* input,int NumberOfPoints);
virtual void GetCentroid(vtkDataSet* input,double centroid[3]);
virtual void ShiftPoints(vtkPoints* input,double shift[3],double scale=1.0);
private:
vtkAbstractPointBasedRegistration(const vtkAbstractPointBasedRegistration&);
void operator=(const vtkAbstractPointBasedRegistration&);
};
The Implementation: We next highlight some key points in the implementation (vtkAbstratPointBasedReg-
istration.cpp). First the constructor, which sets the default values for all the parameters (including the critical
NULL settings for all pointer input variables.)
vtkAbstractPointBasedRegistration::vtkAbstractPointBasedRegistration() : vtkProcessObject()
{
this->Source = NULL; this->Target = NULL;
this->MaximumNumberOfIterations = 50;
this->MaximumNumberOfPoints = 200; this->Epsilon=0.001;
}
The destructor simply sets the input pointers to NULL, this is a way of decrementing the reference count of any
objects used as a Source and Target up to this point:
vtkAbstractPointBasedRegistration::~vtkAbstractPointBasedRegistration() {
this->SetSource(NULL); this->SetTarget(NULL);
}
The rest of the code is fairly straightforward and should be easy to follow.
The concrete derived class vtkLinearICPRegistration:
This class implements a form of the ICP algorithm that estimates a linear transformation.
The Interface: The header le (vtkLinearICPRegistration.h) has the form. The rst part is straight-forward:
205
CHAPTER 20. IMPLEMENTING POINT-BASED REGISTRATION Draft Sep. 2009
class vtkLinearICPRegistration : public vtkAbstractPointBasedRegistration
{
public:
static vtkLinearICPRegistration *New();
vtkTypeRevisionMacro(vtkLinearICPRegistration,vtkAbstractPointBasedRegistration);
void PrintSelf(ostream& os, vtkIndent indent);
// Starts the process by translating source centroid to target centroid.
vtkSetMacro(StartByMatchingCentroids, int);
vtkGetMacro(StartByMatchingCentroids, int);
vtkBooleanMacro(StartByMatchingCentroids, int);
The specication of the transformation type is marginally more interesting, in that we dene three more conve-
nience methods. These are useful for making the code more readable to a human!
// Transformation Type, Rigid, Similarity, Affine
vtkGetMacro(TransformationType,int);
vtkSetClampMacro(TransformationType,int,0,2);
// Three convenience methods
virtual void SetTransformationTypeToRigid() { this->SetTransformationType(0);}
virtual void SetTransformationTypeToSimilarity() { this->SetTransformationType(1);}
virtual void SetTransformationTypeToAffine() { this->SetTransformationType(2);}
Next we explicitly override the two pure virtual methods; this is the heart of the new functionality.
virtual vtkAbstractTransform* GetTransformation();
virtual int Run();
The rest of the denition is fairly straight-forward. Our algorithm will store the transformation in an instance of
vtkTransform
protected:
vtkLinearICPRegistration();
virtual ~vtkLinearICPRegistration();
// Data Members
int StartByMatchingCentroids, TransformationType;
vtkTransform *OutputTransformation;
private:
vtkLinearICPRegistration(const vtkLinearICPRegistration&); // Not implemented.
void operator=(const vtkLinearICPRegistration&); // Not implemented.
};
206
CHAPTER 20. IMPLEMENTING POINT-BASED REGISTRATION Draft Sep. 2009
The Implementation: First the constructor/destructor pair. There is one pointer data member that is
allocated in the constructor and deleted in the destructor. (This symmetry is a useful check, make sure that
anything created in the constructor is deleted in the destructor!)
vtkLinearICPRegistration::vtkLinearICPRegistration() : vtkAbstractPointBasedRegistration(){
this->StartByMatchingCentroids=1; this->TransformationType=2;
this->OutputTransformation=vtkTransform::New();
}
vtkLinearICPRegistration::~vtkLinearICPRegistration(){
this->OutputTransformation->Delete();
}
Next we dene one of the two originally pure-virtual methods, the GetTransformation method. We simply return
the member variable OutputTransformation. Here, we have an example of what is known in C++ as polymorphism.
The returned variable (OutputTransformation) is not an instance of vtkAbstractTransform, but an instance of
vtkTransform which is a descendent of vtkAbstractTransform. You can always safely return an instance of derived
class in the place of an instance of the specied class. (Naturally the reverse is not possible, one cannot return a
vtkAbstractTransform when a vtkTransform is requested!)
vtkAbstractTransform* vtkLinearICPRegistration::GetTransformation() {
return this->OutputTransformation;
}
The heart of the implementation is the Run() method. We will omit some of the code here. We will use
the notation // OMITTED: code to do something where this takes place. First we get our act together by
checking that the input data sets exist. Then we initialize two key objects: (i) the Locator object for quickly
nding nearest points and (ii) the LandmarkTransform for estimating transformations from sets of corresponding
points. Next we get the OutputTransformation in shape (note the PostMultiply call!) and sample the input
data-set to reduce the number of points for computational reasons (The SampledSourcePoints are the xs of
the equations). Finally, we allocate the CorrespondingPoints (the cs of the equations):
int vtkLinearICPRegistration::Run() {
// OMITTED: Code to check that inputs are OK
// Create locator
vtkPointLocator* Locator = vtkPointLocator::New();
Locator->SetDataSet(this->Target); Locator->BuildLocator();
// Get The Landmark Transform All Set
vtkLandmarkTransform* LandmarkTransform=vtkLandmarkTransform::New();
switch(this->TransformationType) {
case 0: LandmarkTransform->SetModeToRigidBody(); break;
case 1: LandmarkTransform->SetModeToSimilarity(); break;
case 2: LandmarkTransform->SetModeToAffine(); break;
}
this->OutputTransformation->Identity(); this->OutputTransformation->PostMultiply();
// Get The Point Sets Ready
207
CHAPTER 20. IMPLEMENTING POINT-BASED REGISTRATION Draft Sep. 2009
vtkPoints* SampledSourcePoints=this->SampleDataSet(this->Source,
this->MaximumNumberOfPoints);
vtkPoints* CorrespondingPoints=vtkPoints::New();
CorrespondingPoints->SetNumberOfPoints(SampledSourcePoints->GetNumberOfPoints());
The next step performs some initial alignment by matching the centroids, this can oftentimes be helpful as a
crude estimate of the overall translation:
double offset[3] = { 0.0,0.0,0.0};
if (this->StartByMatchingCentroids) {
double source_centroid[3]; this->GetCentroid(this->Source,source_centroid);
double target_centroid[3]; this->GetCentroid(this->Target,target_centroid);
for (int ia=0;ia<=2;ia++)
offset[ia]=target_centroid[ia]-source_centroid[ia];
this->ShiftPoints(SampledSourcePoints,offset,1.0);
}
int NumberOfLandmarks=SampledSourcePoints->GetNumberOfPoints();
int NumberOfIterations = 1;
With the preliminaries out of the way we now move on to the alternating estimation itself. The UpdateProgress
method is used to provide feedback to the GUI we will see how to respond to this from the Tcl-based graphical
user interface later.
// Provide Feedback to GUI
this->UpdateProgress(0.0);
// Begin Alternating Estimation
double previousmaxdist=0.0;
while (NumberOfIterations <= this->MaximumNumberOfIterations) {
The rst part is the correspondence estimation. Note the use of the Locator object which makes a complex
search problem trivial!
// 1. Find Correspondences
for(int i = 0; i < NumberOfLandmarks;i++) {
// Get a Point
double x[3]; SampledSourcePoints->GetPoint(i,x);
// Find where this point was at the last estimate!
double tx[3]; LandmarkTransform->TransformPoint(x,tx);
// Find Corresponding point using locator
int id=Locator->FindClosestPoint(tx);
CorrespondingPoints->SetPoint(i,this->Target->GetPoint(id));
}
208
CHAPTER 20. IMPLEMENTING POINT-BASED REGISTRATION Draft Sep. 2009
Next we estimate the transformation using an instance of vtkLandmarkTransform.
LandmarkTransform->SetSourceLandmarks(SampledSourcePoints);
LandmarkTransform->SetTargetLandmarks(CorrespondingPoints);
LandmarkTransform->Update();
We concatenate this with any initial oset (from pre-aligning the centroids) to get the estimate of the complete
transformation, and call UpdateProgress to notify the user interface that we have news:
this->OutputTransformation->Identity();
this->OutputTransformation->Translate(offset);
this->OutputTransformation->Concatenate(LandmarkTransform);
double progress=double(NumberOfIterations)/double(this->MaximumNumberOfIterations);
this->UpdateProgress(progress);
Next comes some code for checking for convergence. When we are done, we clean up any temporary objects
allocated in the Run method.
// OMITTED: Code to check for convergence
vtkDebugMacro(<<"End of Iteration " << NumberOfIterations <<"\n");
++NumberOfIterations;
}
Locator->Delete(); SampledSourcePoints->Delete();
CorrespondingPoints->Delete(); LandmarkTransform->Delete();
this->UpdateProgress(1.0);
return 1;
}
Compiling using CMake and Loading into Tcl
We use a fairly straight-forward CMakeLists.txt. The only new point here, is that we need to explicitly tell the
VTK Tcl Wrappers that vtkAbstractPointBasedRegistration as an abstract class (i.e. it cannot be instantiated).
This is accomplished using the SET_SOURCE_FILES_PROPERTIES command as shown below:
SET(LIBRARY_SRCS
vtkAbstractPointBasedRegistration.cpp
vtkLinearICPRegistration.cpp
)
SET_SOURCE_FILES_PROPERTIES(
vtkAbstractPointBasedRegistration.cpp
ABSTRACT
209
CHAPTER 20. IMPLEMENTING POINT-BASED REGISTRATION Draft Sep. 2009
Figure 20.1:
The GUI
for the
ICP al-
gorithm.
)
The end result of the compilation is a shared library that can be accessed from our Tcl user-interface code. This
is loaded in the usual way as demonstrated by the package le (loadpointreg.tcl) below:
package provide loadpointreg 1.0
package require mylibload 1.0
mylibload vtkPointRegTCL
On MS-Windows, you should change the debug to release (in mylibload), if you compile the release version
of the DLL. There is often a substantial speed up upon switching from debug to release versions.
20.3 Designing the Graphical User Interface
One of the goals of this class is to provide you with the skills to implement a complete application. This includes
both the algorithms and the graphical user interface which makes using the algorithms easier.
The overall GUI is shown in Figure 20.1. It consists of a viewer on the right and a set of controls for setting
parameters etc. on the left. The GUI is implemented using two simple [ Incr] Tcl Classes, the SimpleSurfaceViewer
class (simplesurfaceviewer.tcl) which captures the functionality of the viewer (on the right) and the master class
PointBasedRegistration (pointbased.tcl) which is the overall control that contains the viewer.
We will not go through this code line-by-line. Instead, we will focus on some key elements.
The Simple Surface Viewer
We will use the vtkTkRenderWidget class to embed a vtk Viewer into our Tcl-based GUI. This has some issues
with the standard vtk Interactors, a workaround is provided by the vtkinteraction Tcl package this is part of the
VTK distribution. The header includes all the usual good stu, including the request for vtkinteraction.
210
CHAPTER 20. IMPLEMENTING POINT-BASED REGISTRATION Draft Sep. 2009
package provide SimpleSurfaceViewer 1.0
lappend auto_path [ file dirname [ info script ]]
package require vtk
package require vtkinteraction
package require Itcl
package require Iwidgets
The class denition follows. Essentially we keep handles to the following objects: (i) referencesurface and tar-
getsurface these are vtkPolyData objects of the two surfaces to be aligned, (ii) the transformlter, a vtkTrans-
formPolyData lter for transforming the surface, (iii) the transformation, (iv) the two actors used to display the
surfaces and (v) the renderer and the renderwidget. These are dened rst, followed by the constructor and the
destructor:
itcl::class SimpleSurfaceViewer {
protected variable referencesurface [ vtkPolyData New ]
protected variable transformfilter [ vtkTransformPolyDataFilter New ]
protected variable targetsurface [ vtkPolyData New ]
protected variable transformation 0
protected variable referenceactor [ vtkActor New ]
protected variable targetactor [ vtkActor New ]
protected variable renderwidget 0
protected variable renderer 0
protected variable basewidget 0
private common thisparam
constructor { base } { set basewidget $base
set thisparam($this,referencemode) "Wireframe"
set thisparam($this,targetmode) "Wireframe"
$this Initialize }
destructor {
$referencesurface Delete; $transformfilter Delete
$targetsurface Delete;
$referenceactor Delete $targetactor Delete
$renderer Delete; $renderwidget Delete }
The rest of the header denes some interface and implementation methods. The rst method, is the method
SetSurfacesAndTransformation, which will be used by the main GUI class to provide updates as to the current
state of the registration.
public method SetSurfacesAndTransformation { reference target transform }
public method SetSurfaceMode { md }
public method ResetRenderer { }
public method UpdateDisplay { }
public method GetThisPointer { } { return $this }
protected method Initialize { }
protected method CreatePipelines { }
protected method AddInteractor { renderwidget renwin }
211
CHAPTER 20. IMPLEMENTING POINT-BASED REGISTRATION Draft Sep. 2009
};
The Implementation: We describe the implementation below, highlighting key points:
itcl::body SimpleSurfaceViewer::CreatePipelines { } {
# Omitted code to generate the two actors and their pipelines
# Good idea to use a simple object such as a sphere source
# to initialize the surfaces with default shapes until
# the user properly specifies them!
}
The Initialize method demonstrates how to create and add a vtkTkRenderWidget.
itcl::body SimpleSurfaceViewer::Initialize { } {
iwidgets::Labeledframe $basewidget -labelpos nw -labeltext "Surface Viewer"
pack $basewidget -side top -expand true -fill both
set labelframe [ $basewidget childsite ]
set buttonbar [ frame $labelframe.a1 ]
pack $buttonbar -side bottom -expand false -fill x -pady 2
# Create the Viewer First and add the actors to it
$this CreatePipelines
set renderwidget [ vtkTkRenderWidget $labelframe.a2 ]
pack $renderwidget -side top -expand true -fill both
set renwin [ $renderwidget GetRenderWindow ]
set renderer [ vtkRenderer New ]
$renwin AddRenderer $renderer
$renderer AddActor $referenceactor; $renderer AddActor $targetactor
# Create the Buttons Next
# Omited code -- see the file for details
# Key function to properly add interaction -- see below
$this AddInteractor $renderwidget $renwin
}
The interaction is added to the vtkTkRenderWidget using a call to the Tcl function ::vtk::bind_tk_widget
as shown below:
itcl::body SimpleSurfaceViewer::AddInteractor { renderwidget renwin } {
::vtk::bind_tk_widget $renderwidget $renwin
update idletasks
[ $renwin GetInteractor ] ExposeEvent
}
212
CHAPTER 20. IMPLEMENTING POINT-BASED REGISTRATION Draft Sep. 2009
The rest of the world communicates with this viewer primarily using the method below. It can be used to provide
the triad of (i) the reference surface, (ii) the target surface and (iii) the current state of the transformation (or 0
to use an identity). Note the use of the ShallowCopy method to properly and eciently create a link to the input
surfaces.
itcl::body SimpleSurfaceViewer::SetSurfacesAndTransformation { reference
target transformation } {
$referencesurface ShallowCopy $reference; $targetsurface ShallowCopy $target
if { $transformation != 0 } {
$transformfilter SetTransform $transformation; $transformfilter Update
} else {
set temp [ vtkIdentityTransform New ];
$transformfilter SetTransform $temp; $temp Delete
}
}
The rest of the code provides additional display functionality:
itcl::body SimpleSurfaceViewer::SetSurfaceMode { md } {
# Omitted code to switch surface display from surface to wireframe or points
}
itcl::body SimpleSurfaceViewer::ResetRenderer { } {
# Reset the camera to show all objects
$renderer ResetCamera; $renderer ResetCameraClippingRange
$this UpdateDisplay
}
itcl::body SimpleSurfaceViewer::UpdateDisplay { } {
# Force a render update
[ $renderer GetRenderWindow ] Render
}
The Master Class PointBasedRegistration
This class leverages the viewer class dened above to provide a complete application around our ICP implemen-
tation. We use a notebook widget to break up the user interface into three tabs: (i) The Input tab where
we dene the input surfaces, (ii) the Compute tab where we specify algorithm parameters and execute the
algorithm and (iii) the Output tab where we can see and save the output matrix.
The script header contains the usual package require statements. The last two statements require the Surface
Viewer and our newly compiled shared library which includes the ICP code. This is where the C++ code is loaded
into the Tcl script.
package provide PointBasedRegistration 1.0
lappend auto_path [ file dirname [ info script ]]
# Omitted standard stuff
package require SimpleSurfaceViewer
213
CHAPTER 20. IMPLEMENTING POINT-BASED REGISTRATION Draft Sep. 2009
package require loadpointreg
The Class Interface: The class interface has the usual mix of methods and variables. We highlight the
ExitCommand which is invoked by the exit button. This contains code to do proper cleanup on MS-Windows.
The Initialize command creates the user interface one tab at a time:
itcl::class PointBasedRegistration {
protected variable referencesurface [ vtkPolyData New ]
protected variable targetsurface [ vtkPolyData New ]
protected variable currentregistration 0
protected variable simplesurfaceviewer 0
protected variable notebook 0
private common thisparam
protected variable basewidget 0
protected variable outputbox 0
protected variable outputmatrix "1 0 0 0\n 0 1 0 0\n0 0 1 0\n0 0 0 1"
constructor { base } { set basewidget $base; $this Initialize }
destructor { $referencesurface Delete; $targetsurface Delete
catch [ $currentregistration Delete ]
catch [ itcl::delete obj $simplesurfaceviewer ] }
# Omitted method definitions
public method ExitCommand { }
protected method Initialize { }
};
The Initialize method (see below) creates the Graphical User Interface (GUI). This consists of three widgets along
the top, with the viewer on the right, a notebook widget on the left and a divider along the middle. Each tab on
the notebook is created using a separate helper method. Finally we create the viewer frame.
Creating the GUI:
itcl::body PointBasedRegistration::Initialize { } {
$this ResetParameters
toplevel $basewidget; wm geometry $basewidget 600x400
wm title $basewidget "Point Based Registration"; update
set notebook [ iwidgets::tabnotebook $basewidget.left -tabpos n -width 250 ]
set middle [ frame $basewidget.middle -width 5 -bg black ]
set viewframe [ frame $basewidget.right -width 500 ]
pack $notebook $middle -side left -expand false -fill y
pack $viewframe -side right -expand true -fill both
# Create Notebook Tabs
$this CreateInputGUI [ $notebook add -label "Input" ]
214
CHAPTER 20. IMPLEMENTING POINT-BASED REGISTRATION Draft Sep. 2009
$this CreateComputeGUI [ $notebook add -label "Compute" ]
$this CreateOutputGUI [ $notebook add -label "Output" ]
$notebook view "Input"
# The Viewer Frame
frame $viewframe.2; pack $viewframe.2 -side bottom -expand false -fill x
eval "button $viewframe.2.2 -text Exit -command { $this ExitCommand }"
pack $viewframe.2.2 -side right -padx 1 -expand false
iwidgets::entryfield $viewframe.2.1 -textvariable \
[ itcl::scope thisparam($this,icp_status) ] -width 50 -labeltext "Status:"
pack $viewframe.2.1 -side left -padx 1 -expand true -fill x
set simplesurfaceviewer [ CreateSurfaceViewer $viewframe.1 ]
eval "wm protocol $basewidget WM_DELETE_WINDOW { $this ExitCommand }"
}
The following methods create the individual tab-GUIs. These are shown here in a highly abbreviated form:
itcl::body PointBasedRegistration::CreateSurfaceViewer { par } {
return [ [ SimpleSurfaceViewer \#auto $par ] GetThisPointer ]
}
itcl::body PointBasedRegistration::ResetParameters { } {
# Omitted Code to set default values for various parameters
}
itcl::body PointBasedRegistration::CreateInputGUI { base } {
# Omitted Code to create the input GUI for loading the
# two surfaces
}
itcl::body PointBasedRegistration::CreateComputeGUI { base } {
eval "button $base.bot -text \"Compute Registration\" \
-command { $this ComputeRegistration }"
pack $base.bot -side bottom -expand false -fill x
iwidgets::Labeledframe $base.top -labelpos nw -labeltext "Parameters" -relief ridge
pack $base.top -side top -expand true -fill both -pady 2
set w [ $base.top childsite ]
set k 0
iwidgets::entryfield $w.$k -labeltext "Max Num Points:" -width 5
-validate integer \
-textvariable [ itcl::scope thisparam($this,icp_numberofpoints) ] -relief sunken
pack $w.$k; incr k
# Omitted code to add more entryfields for the rest of the ICP Parameters
}
itcl::body PointBasedRegistration::CreateOutputGUI { base } {
# Omitted code to create the GUI for displaying and saving the matrix
}
215
CHAPTER 20. IMPLEMENTING POINT-BASED REGISTRATION Draft Sep. 2009
Finally, the methods for loading and saving objects:
itcl::body PointBasedRegistration::LoadSurface { mode inputfilename } {
# Omitted Code to load the surface from a .vtk file
}
itcl::body PointBasedRegistration::SaveMatrix { } {
# Omitted Trivial code for saving the matrix
}
Invoking and Interacting with vtkLinearICPRegistration: The heart of the PointBasedRegistration
class is the
ComputeRegistration and RegistrationProgressUpdate methods that follow:
itcl::body PointBasedRegistration::ComputeRegistration { } {
# Omitted code that checks that both surfaces are OK
# Initialize the Display
$simplesurfaceviewer SetSurfacesAndTransformation $referencesurface $targetsurface 0
$simplesurfaceviewer UpdateDisplay
# Create the ICP Registration Class
set icp [ vtkLinearICPRegistration New ]
set currentregistration $icp
$icp SetSource $referencesurface; $icp SetTarget $targetsurface
$icp SetMaximumNumberOfIterations $thisparam($this,icp_numberofiterations)
$icp SetMaximumNumberOfPoints $thisparam($this,icp_numberofpoints)
$icp SetStartByMatchingCentroids $thisparam($this,icp_alignoriginsfirst)
if { $thisparam($this,icp_transformationmode) == "Rigid" } {
$icp SetTransformationTypeToRigid
} elseif { $thisparam($this,icp_transformationmode) == "Similarity" } {
$icp SetTransformationTypeToSimilarity
} else {
$icp SetTransformationTypeToAffine
}
Next we attach an observer to the class. This results in the function RegistrationProgressUpdate being called
each time the C++ code invokes the update progress method (this generates the ProgressEvent event), or when
it is done. Finally the registration is invoked using the Run method:
# Observer Stuff
eval "$icp AddObserver ProgressEvent { $this RegistrationProgressUpdate }"
eval "$icp AddObserver EndEvent { $this RegistrationProgressUpdate }"
# Execute
$icp Run
216
CHAPTER 20. IMPLEMENTING POINT-BASED REGISTRATION Draft Sep. 2009
# Omitted code to capture the matrix etc
# Finally switch to the output pane
set thisparam($this,icp_status) "Done with Registration"
$notebook view "Output"
}
The Progress Update method rst updates the viewer with the current transformation estimate (currentregistration
= icp, see above). Next we update the display and the status label on the bottom:
itcl::body PointBasedRegistration::RegistrationProgressUpdate { } {
$simplesurfaceviewer SetSurfacesAndTransformation \
$referencesurface $targetsurface [ $currentregistration GetTransformation ]
$simplesurfaceviewer UpdateDisplay
set thisparam($this,icp_status) [ format "Registration Progress %.2f "\
[ expr 100.0 * [ $currentregistration GetProgress ] ] ]
update idletasks
}
The Exit Command: Finally a clean exit command to avoid those annoying crash-on-exit issues in MS-
Windows:
itcl::body PointBasedRegistration::ExitCommand { } {
vtkCommand DeleteAllObjects
exit
}
Invoking: An example is shown in script20-2.tcl. This includes loading two sphere surfaces by default:
lappend auto_path [ file dirname [ info script ]]
package require PointBasedRegistration
wm withdraw . ; update
set pb [ PointBasedRegistration \#auto .a ]
$pb LoadSurface Reference sphere1.vtk
$pb LoadSurface Target sphere2.vtk
217
CHAPTER 20. IMPLEMENTING POINT-BASED REGISTRATION Draft Sep. 2009
20.4 An Alternative Implementation Using BioImage Suite Interactors
Instead of using the standard vtk interactors via the ::vtk::bind_tk_widget command, we can use our custom
BioImage Suite interactors via the use of the vtkpxGUIRenderer class. This provides a rich GUI for manipulating
the viewer, as shown in Figure 20.2.
The use of the BioImage Suite viewers is described in two classes (i) BioImageSuiteSimpleSurfaceViewer this
derives from SimpleSurfaceViewer and replaces some of the functionality there and (ii) BioImageSuitePointBase-
dRegistration this similarly derives from PointBasedRegistration and replaces some of the functionality there.
The heart of the new functionality is in three methods in BioImageSuiteSimpleSurfaceViewer. First the new
AddInteractor method creates a vtkpxGUIRenderer class (its GUI is accessed by pressing the shift key while
clicking with the right mouse button in the viewer.)
itcl::body BioImageSuiteSimpleSurfaceViewer::AddInteractor { renderwidget renwin } {
update idletasks; set bioimagesuite_viewer [ vtkpxGUIRenderer New ]
$bioimagesuite_viewer BindMouseEvents $renderwidget \
"$this HandleMouseEvent" "$this HandleUpdateEvent"
$bioimagesuite_viewer Initialize $renderwidget $renderer 0
set renderwindow $renwin
}
This requires two additional helper methods for processing expose and mouse events:
itcl::body BioImageSuiteSimpleSurfaceViewer::HandleUpdateEvent { args } {
if {$renderwindow != 0 } { $renderwindow Render }
}
itcl::body BioImageSuiteSimpleSurfaceViewer::HandleMouseEvent { mouse stat \
x1 x2 widgetname args} {
if { $bioimagesuite_viewer == 0 } { return }
# Need to flip y-axis vtk counts from bottom tk counts from top !!!
set wd [ $renderwindow GetSize ]
set x2 [ expr [lindex $wd 1 ] - $x2 ]
$bioimagesuite_viewer HandleMouseButtonEvent $mouse $stat $x1 $x2
}
The BioImageSuitePointBasedRegistration overrides a total of two short methods from its parent class: (i) The
CreateSurfaceViewer method which creates the instance of the BioImageSuiteSimpleSurfaceViewer. (ii) A slightly
Figure 20.2: The helper GUI of the vtkpxGUIRenderer class.
218
CHAPTER 20. IMPLEMENTING POINT-BASED REGISTRATION Draft Sep. 2009
dierent ExitCommand to handle properly the GUI constructs in BioImage Suite.
itcl::body BioImageSuitePointBasedRegistration::CreateSurfaceViewer { par } {
return [ [ BioImageSuiteSimpleSurfaceViewer \#auto $par ] GetThisPointer ]
}
itcl::body BioImageSuitePointBasedRegistration::ExitCommand { } {
destroy $basewidget ; update idletasks; exit
}
The use of these classes is illustrated in script20-3.tcl.
219
Draft Sep. 2009
Chapter 21
Implementing Intensity Based
Segmentation
In this chapter, we will discuss the implementation of three common image segmentation, or labeling, algorithms.
For the purposes of this chapter, segmentation is the process of assigning a label to each voxel that denes the
class (often equivalent to tissue type) it belongs in. For example, in MRI brain images the goal is to classify each
voxel as belonging to white matter, gray matter, cerebrospinal uid (CSF) or background (often the last two are
combined into a single class). The three algorithms are (i) a manual multiple threshold segmentation method, (ii)
an automated k-means clustering based approach [?] and (iii) an automated method that uses a Markov Random
Field model for image smoothness [?, ?]. A BioImage Suite component for accessing these algorithms is also
described.
21.1 Introduction The Three Algorithms
The Multiple Threshold Algorithm
This is a fairly trivial method that may be useful for initialization purposes. Given (a) an input image I which is
a collection of voxels i(x), where x is the voxel index, and (b) a set of M thresholds t
i
, where i = 0 : M 1
and t
i
> t
i1
we output a label image L (similarly a collection of voxels l(x) which takes values 0 : M. The
segmentation uses the following simple rule:
l(x) =
_
_
_
0 : i(x) t
0
r : t
r1
i(x) < t
r
, r (1, M 1)
M : i(x) > t
M1
This type of algorithm is often useful for CT images where the intensities are calibrated.
The K-means Clustering Algorithm
The K-means clustering algorithm (see [?]) can be used to essentially estimate the thresholds above. In our
implementation we assume that the intensity in each tissue class can be modeled as having a normal distribution
with mean and standard deviation . We will use the notation p(i[l = c) = ^(
c
,
c
) to dene this, where l
is the label and c is the class number. The goal of the version of the k-means algorithm that we will describe is
to estimate the class parameters (, ) for all classes and then to assign each voxel to the class that maximizes
the function:
220
CHAPTER 21. IMPLEMENTING INTENSITY BASED SEGMENTATION Draft Sep. 2009

l(x) =
arg max
l
p(i(x)[l(x) = c)
=
arg max
l
1
_
2
2
c
e
(i(x)c)
2
2
2
c
(21.1)
A simpler form of the algorithm assumes that all
c
s have the same value. This reduces the problem to estimating
the means only. The procedure can be described in recipe form as:
1. Decide on the number of classes M
2. Assign class centroids
c
and optionally standard deviations
c
for each class. The most common way to
do this is to equally space the
i
s and set all
i
s to some constant value.
3. Estimate the labels l(x) using equation 21.1. This is an exhaustive optimization compute p(i[l = c) for
all ls and pick the l that maximizes this probability.
4. Estimate a new set of
c
s and
c
s by computing the mean and standard deviation of the intensities of
the voxels labeled as having class c.
5. Repeat steps 3-4 until the parameters
c
and
c
converge.
Note that, since the spatial position of the voxels x does not enter into the calculation, a quick way to implement
this method is by rst forming the image histogram and performing all operations on the histogram itself. This
can speed up the iterations by a factor of 15000 or so in the case of an 128 128 128 image whose histogram
can be described by 128 bins.
Imposing Local Smoothness using Markov Random Fields
The key weakness of the previous method is that, as noted, the spatial position of each voxel is not used during the
segmentation. Spatial homogeneity is often a powerful constraint that can be used to improve the segmentation
in the presence of image noise. This can be loosely thought of as nding the label at each voxel that is an optimal
constraint between (i) the intensity at that voxel and (ii) the labels assigned to its neighboring voxels.
Markov Random Fields: The probabilistic structure used, most frequently, to capture such homogeneity is
to model the label (classication) image as a Markov Random Field. This (and we are skipping a lot of math here)
basically reduces to describing the probability of each voxel belonging to class l, as having a Gibbs distribution of
the form:
p(l(x)) = k
1
exp(W(L(
x
), l)) (21.2)
where k
1
is a normalization constant, L(
x
) is the set of labels of the neighboring voxels and W is a positive
semi-denite function. This is a way of converting an energy-function like smoothness term into a probability
distribution for integration into a probabilistic framework. The function W can take many forms, the one we will
use here is[?]:
W(l(X
n
) =

Rx
(l(x

) l(x)) (21.3)
This essentially counts the number of voxels in the neighborhood of the voxel at location x that have labels
dierent from it.
Overall Formulation using the Expectation-Minimization Framework: We rst dene the vector
as the collection of the means and standard deviations of all C classes, i.e. = [
0
,
0
, . . . ,
c1
,
c1
]. The
goal of the segmentation is to estimate both the set of labels L and the class parameters , given the image I.
We can express this mathematically as:

L,

=
arg max
L,
p(L, [I) (21.4)
221
CHAPTER 21. IMPLEMENTING INTENSITY BASED SEGMENTATION Draft Sep. 2009
As is commonly done, this can be solved iteratively in the same spirit as the EM-framework as:
E-Step:
k
=
arg max

p([I, L
k1
), M-Step: L
k
=
arg max
L
p(L[I,
k
) (21.5)
where k is the iteration number. In the E-Step we estimate a new set of parameters
k
given the current
classication L
k1
. In the M-Step, using the newly estimated
k
we estimate a new classication L
k
.
E-Step: This is straightforward. For each class i we estimate the mean and standard deviation of the intensities
I of all the voxels where M = i. This is identical to the procedure used in the k-means algorithm above.
M-Step: This takes the form of a Bayesian a-posterior maximization. First we express

l(x) =
arg max
l
log p(l(x)[i(x),
k
, L(
x
)
k
1
+ log p(i(x),
k
[l(x))
. .
Data Adherence Term
+ logp(l[L(
x
))
. .
MRF Smoothness Term
(21.6)
where k
1
is a constant. This equation is easily maximized by a greedy search strategy as M can only take values
of 1 . . . C. The prior term on the classication, p(L), can be dened by modeling L as a Markov random eld
(see discussion above and equation 21.3). We express the likelihood (or data-adherence) term for each possible
value of l(x) = c as:
p(i(x),
k
[l(x) = c) = p(i(x)[
k
, l(x) = c) (21.7)
which is similar to the model previously used in equation 21.1.
Hopefully the math will become easier to follow with a quick look at the code implementing it.
21.2 Algorithm Implementation
All three algorithms are implemented as classes deriving from the vtkSimpleImageToImageFilter class. The scope
is similar to material described in chapter 17, and you are strongly urged to re-read the notes from that chapter
again.
The Multi Threshold Method
The header le of this method (vtkMultiThresholdSegmentation.h) is fairly standard. In addition to an input
image the user needs to specify an array of M thresholds that will results in M + 1 output classes.
class vtkMultiThresholdSegmentation : public vtkSimpleImageToImageFilter
{
public:
static vtkMultiThresholdSegmentation *New();
vtkTypeMacro(vtkMultiThresholdSegmentation,vtkSimpleImageToImageFilter);
// The Input here is N Thresholds resulting in N+1 output classes
vtkSetObjectMacro(Thresholds,vtkDoubleArray);
vtkGetObjectMacro(Thresholds,vtkDoubleArray);
protected:
vtkMultiThresholdSegmentation();
virtual ~vtkMultiThresholdSegmentation();
virtual int RequestInformation(vtkInformation *vtkNotUsed(request), vtkInformationVector **inputVector,
222
CHAPTER 21. IMPLEMENTING INTENSITY BASED SEGMENTATION Draft Sep. 2009
vtkInformationVector *outputVector);
virtual void SimpleExecute(vtkImageData* in,vtkImageData* out);
vtkDoubleArray* Thresholds;
};
The implementation (vtkMultiThresholdSegmentation.cpp) is fairly straightforward. First we dene the RequestIn-
formation method to explicitly specify that the output image will always have type short and a single frame
regardless of what the input image is like:
223
CHAPTER 21. IMPLEMENTING INTENSITY BASED SEGMENTATION Draft Sep. 2009
int vtkMultiThresholdSegmentation::RequestInformation(
vtkInformation *request,
vtkInformationVector **inputVector,
vtkInformationVector *outputVector)
{
vtkDataObject::SetPointDataActiveScalarInfo(outputVector->GetInformationObject(0), -1, 1);
vtkDataObject::SetPointDataActiveScalarInfo(outputVector->GetInformationObject(0), VTK_SHORT, -1);
}
The thresholding takes place in the SimpleExecute method. This is fairly trivial:
void vtkMultiThresholdSegmentation::SimpleExecute(vtkImageData* input,vtkImageData* output )
{
// Omitted Code to check that inputs are OK
int num=this->Thresholds->GetNumberOfTuples();
int NumberOfClasses=num+1;
double* thr=new double[num];
double* outthr=new double[num];
for (int ia=0;ia<num;ia++)
thr[ia]=this->Thresholds->GetComponent(ia,0);
// Omitted code that sorts thresholds and places them into array outthr[]
vtkDataArray* inpdata=input->GetPointData()->GetScalars();
vtkDataArray* outdata=output->GetPointData()->GetScalars();
outdata->FillComponent(0,0.0); int numvoxels=inpdata->GetNumberOfTuples();
for (int voxel=0;voxel<numvoxels;voxel++) {
int c=0,done=0; double v=inpdata->GetComponent(voxel,0);
while (c<num && done==0) {
if (v<outthr[c]) {
outdata->SetComponent(voxel,0,c);
done=1;
}
else
c++;
}
if (done==0)
outdata->SetComponent(voxel,0,NumberOfClasses-1);
}
delete [] thr; delete [] outthr;
}
The K-Means Method
This is a more interesting class. It takes a number of arguments, namely (i) Number of Class this species the
number of unique class labels, (ii) Iterations the maximum number of iterations, (iii) Number Of Bins this
denes the resolution of the histogram, (iv) Convergence this sets the thereshold that if the maximum change
in the the estimation of the class falls below which, the algorithm is set to have converged and (v) MaxSigmaRatio
which constrains the range of the values of the standard deviations. Setting this to 1 ensures that all standard
deviations are equal. Setting this to 0 results in an unconstrained estimation of the standard deviations.
224
CHAPTER 21. IMPLEMENTING INTENSITY BASED SEGMENTATION Draft Sep. 2009
class vtkKMeansSegmentation : public vtkSimpleImageToImageFilter
{
public:
static vtkKMeansSegmentation *New();
vtkTypeMacro(vtkKMeansSegmentation,vtkSimpleImageToImageFilter);
// Omitted Code
// Set and Get Macros for the five input parameters
// NumberOfClasses, Iterations, NumberOfBins, Convergence, MaxSigmaRatio
// Get The Histogram and the class Parameters if desired
vtkGetObjectMacro(Histogram,vtkImageData);
vtkGetObjectMacro(Parameters,vtkDoubleArray);
protected:
vtkKMeansSegmentation();
virtual ~vtkKMeansSegmentation();
virtual int RequestInformation(vtkInformation *vtkNotUsed(request), vtkInformationVector **inputVector,
vtkInformationVector *outputVector);
virtual void SimpleExecute(vtkImageData* in,vtkImageData* out);
// Helper Methods
virtual double Metric(double x,double m,double sigma2);
virtual vtkImageData* CreateHistogram(vtkImageData* input,int NumBins);
virtual int InitializeParameters(vtkImageData* histogram,int numclasses,
vtkDoubleArray* params);
// Omitted class parameter definition
};
Implementation: This is found in vtkKMeansSegmentation.cpp. We will only highlight key pieces of code
here.
The rst interesting point is that we dene the Gaussian distribution in a virtual method called Metric. If we
wanted to replace this with a dierent distribution, we could derive a new class from vtkKMeansSegmentation
and simply override the Metric method.
double vtkKMeansSegmentation::Metric(double x,double m,double sigma2) {
return (x-m)*(x-m)/(-2.0*sigma2)* 1.0/sqrt(2.0*vtkMath::Pi()*sigma2);
}
The next interesting point is the creation of the histogram. This uses the vtkImageAccumulate lter. Note that,
the vtkImageAccumulate Filter is designed to generate a histogram for color images (RGB). It assumes that the
input image will have upto three components and produces a histogram for each. We specify the position, spacing
and number of bins using the SetComponentOrigin, SetComponentSpacing and SetComponentExtent methods
of the vtkImageAccumulate class respectively. Since we are dealing with only grayscale images we set the extent
for components 2 and 3 (which do not exist) to have a single bin.
vtkImageData* vtkKMeansSegmentation::CreateHistogram(vtkImageData* input,int NumBins) {
if (input==NULL) return NULL; if (NumBins<4) NumBins=4; double range[2];
input->GetPointData()->GetScalars()->GetRange(range); double minv=range[0],
225
CHAPTER 21. IMPLEMENTING INTENSITY BASED SEGMENTATION Draft Sep. 2009
maxv=range[1]; this->HistogramOrigin=minv; this->HistogramSpacing=1.0;
this->NumberOfBins=NumBins; int drange=int(maxv-minv+1); if (drange<
this->NumberOfBins) this->NumberOfBins=drange; while(drange>
this->NumberOfBins*this->HistogramSpacing) this->HistogramSpacing+=1.0;
this->NumberOfBins=int(drange/this->HistogramSpacing+0.5);
vtkImageAccumulate* accumulate=vtkImageAccumulate::New();
accumulate->SetInput(input);
accumulate->SetComponentOrigin(this->HistogramOrigin,0.0,0.0);
accumulate->SetComponentSpacing(this->HistogramSpacing,1.0,1.0);
accumulate->SetComponentExtent(0,this->NumberOfBins-1,0,0,0,0);
accumulate->Update();
vtkImageData* out=vtkImageData::New(); out->ShallowCopy(accumulate->GetOutput());
accumulate->Delete();
return out;
}
The class parameters are initialized in the method InitializeParameters. The means are evenly spaced in the
intensity range and the standard deviation is set to be constant. The parameters are stored in a 3-component
double array which has one tuple for each class. The tuples form the vector [ NumVoxels, Mean, Standard
Deviation ] where NumVoxels is the number of voxels having this class label.
Finally the segmentation itself is implemented in the SimpleExecute method as usual. The rst part deals with
initialization issues:
void vtkKMeansSegmentation::SimpleExecute(vtkImageData* input,vtkImageData* output )
{
// Omitted code: check for valid input image
this->Histogram=this->CreateHistogram(input,this->NumberOfBins);
this->Parameters=vtkDoubleArray::New();
this->InitializeParameters(this->Histogram,this->NumberOfClasses,this->Parameters);
this->UpdateProgress(0.05);
int dim[3]; this->Histogram->GetDimensions(dim);
vtkDataArray* data=this->Histogram->GetPointData()->GetScalars();
double* mean =new double[this->NumberOfClasses], *sigma2=new double[this->NumberOfClasses];
double* sum =new double[this->NumberOfClasses], *sum2 =new double[this->NumberOfClasses];
double* num =new double[this->NumberOfClasses],
for (int j=0;j<this->NumberOfClasses;j++) {
sigma2[j]=this->Parameters->GetComponent(j,2); sigma2[j]*=sigma2[j];
mean[j] =this->Parameters->GetComponent(j,1);
}
Next we launch into the iterative estimation of labels and class parameters. Note that, rather than iterating on
a voxel-by-voxel basis, we operate instead on the image histogram for eciency.
int iter=1; double error=this->Convergence+1.0;
226
CHAPTER 21. IMPLEMENTING INTENSITY BASED SEGMENTATION Draft Sep. 2009
while (iter <= this->Iterations && error > this->Convergence)
{
error=0.0; double totalnum=0.0;
for (int i=0;i<this->NumberOfClasses;i++) { sum[i]=0.0;sum2[i]=0.0;num[i]=0.0;}
For each bin, we compute the actual intensity v and then nd which class has the best Metric i.e. the highest
likelihood. We assign the bin to this class and use it to form the sums needed to compute the class mean and
standard deviation:
for (int bin=0;bin<dim[0];bin++) {
double v=this->HistogramOrigin+double(bin)*this->HistogramSpacing;
double numv=data->GetComponent(bin,0);
double bestp=0.0; int bestc=0;
for (int c=0;c<this->NumberOfClasses;c++) {
double p=this->Metric(v,mean[c],sigma2[c]);
if (p>bestp) { bestp=p; bestc=c; }
}
num[bestc]+=numv; sum[bestc]+=v*numv; sum2[bestc]+=v*v*numv;
totalnum+=numv;
}
Once the sums are complete we compute the class means and standard deviations. We also check for convergence.
If the means remain close to their previous values, this implies that the algorithm has converged and we need no
more iterations. An interesting twist here (omiited from this code) is that we compute the maximum standard
deviation of all the classes and can enforce all other classes to have standard deviations at least as big as a certain
fraction (MaxSigmaRatio) of this. This type of constraint adds stability to the estimation:
for (int c=0;c<this->NumberOfClasses;c++) {
double m=sum[c]/num[c];
if (fabs(m-mean[c])>error) error=fabs(m-mean[c]);
mean[c]=m; sigma2[c]=(sum2[c]/num[c]-mean[c]*mean[c]);
double s=sqrt(sigma2[c]);
}
++iter;
}
Finally, once the iterative estimation of class parameters has converged, we re-estimate class labels, this time on
a voxel-by-voxel basis so that we can create the nal segmentation map:
// Omitted code sort classes to have ascending means
vtkDataArray* inpdata=input->GetPointData()->GetScalars();
vtkDataArray* outdata=output->GetPointData()->GetScalars();
outdata->FillComponent(0,-1.0);
int numvoxels=inpdata->GetNumberOfTuples();
for (int voxel=0;voxel<numvoxels;voxel++) {
227
CHAPTER 21. IMPLEMENTING INTENSITY BASED SEGMENTATION Draft Sep. 2009
double bestp=0.0; int bestc=0;
double v=inpdata->GetComponent(voxel,0);
for (int c=0;c<this->NumberOfClasses;c++) {
double p=this->Metric(v,mean[c],sigma2[c]);
if (p>bestp || c==0) { bestp=p; bestc=c; }
}
outdata->SetComponent(voxel,0,bestc);
}
delete [] mean; delete [] sigma2; delete [] sum; delete [] sum2; delete [] num;
}
The MRF Segmentation Method
Things will begin to appear to get more complex here. There is, however, nothing particularly complex about any
of the code following, but it might be a little longer than can be easily described in a handout.
The class denition (vtkMRFSegmentation.h) is fairly straight-forward. The class takes a number of key pa-
rameters namely (i) NumberOfIterations, (ii) NumberOfClasses (iii) Smoothness the weight of the MRF term.
Secondary parameters include ConvergencePercentage and MRFIterations (the number of iterations in the M-
Step). Finally, we also use an ImageNoise term that adds some stability to the estimation more on this later.
class vtkMRFSegmentation : public vtkSimpleImageToImageFilter {
public:
static vtkMRFSegmentation *New();
vtkTypeMacro(vtkMRFSegmentation,vtkSimpleImageToImageFilter);
// Omitted Code, Set/Get Macros for Class Parameters such as
// NumberOfIterations, MRFIterations, NumberOfClasses, Smoothness,ImageNoisePercentage
// ConvergencePercentage
// Initial Segmentation Map as Image
vtkSetObjectMacro(InitialSegmentation,vtkImageData);
vtkGetObjectMacro(InitialSegmentation,vtkImageData);
// Initial Segmentation Map as Image
vtkGetObjectMacro(Parameters,vtkDoubleArray);
protected:
vtkMRFSegmentation();
virtual ~vtkMRFSegmentation();
virtual int RequestInformation(vtkInformation *vtkNotUsed(request), vtkInformationVector **inputVector,
vtkInformationVector *outputVector);
virtual void SimpleExecute(vtkImageData* in,vtkImageData* out);
// E-Step Is Easy for the most_part
virtual void DoExpectationStep(vtkImageData* intensities,vtkImageData* classification,
vtkDoubleArray* params,int numclasses);
// M-Step (...) indicates lots of parameters!
virtual double Metric(double x,double m,double sigma2);
virtual double ComputeLogLikelihoodProbability(...)
virtual double ComputeLogMRFPrior(...)
virtual double ComputeTotalMinusLogProbability(...)
virtual int UpdateVoxel(int eveni,int evenj,int evenk,int pass);
228
CHAPTER 21. IMPLEMENTING INTENSITY BASED SEGMENTATION Draft Sep. 2009
virtual int ComputeMRFIncrementsAndWeights(vtkImageData* img,int incr[6],double wgt[6]);
virtual int ClassifyVoxel(....)
virtual double DoMaximizationStep(vtkImageData* intensity_image, vtkImageData* label_image,
vtkDoubleArray* params, int numclasses,
double smoothness, int maxiter);
//Omitted Code -- definition of class variables
};
The Implemetation: This consists of about 400 lines of code in total which we will not describe in huge
detail. Essentially the ow goes as follows: SimpleExecute initializes the process and calls the Expectation and
Maximization procedures in an alternating fashion until convergence:
void vtkMRFSegmentation::SimpleExecute(vtkImageData* input,vtkImageData* output)
{
# Omitted Code to check that inputs are valid
# Compute Image Noise Variance
double range[2]; input->GetPointData()->GetScalars()->GetRange(range);
this->ImageNoiseVariance=pow(range[0]+
this->ImageNoisePercentage*0.01*(range[1]-range[0]),2.0);
# Create Ouput Image
vtkImageData* classification=vtkImageData::New();
classification->CopyStructure(output);
classification->AllocateScalars();
classification->GetPointData()->GetScalars()->CopyComponent(0,
this->InitialSegmentation->GetPointData()->GetScalars(),0);
int iterations=1;
while(iterations<=this->NumberOfIterations)
{
this->DoExpectationStep(input,classification,this->Parameters,this->NumberOfClasses);
double changed=this->DoMaximizationStep(input,classification,this->Parameters,
this->NumberOfClasses,this->Smoothness,this->MRFIterations);
if (changed<this->ConvergencePercentage)
iterations=this->NumberOfIterations+1;
++iterations;
}
output->ShallowCopy(classification); classification->Delete();
this->UpdateProgress(1.0);
}
The Expectation Step: The expectation step is also fairly straight-forward and similar in scope to the code in
vtkKMeansSegmentation for estimating parameters.
void vtkMRFSegmentation::DoExpectationStep(vtkImageData* intensities,
vtkImageData* classification,
vtkDoubleArray* params,int numclasses)
229
CHAPTER 21. IMPLEMENTING INTENSITY BASED SEGMENTATION Draft Sep. 2009
{
vtkDataArray* intens=intensities->GetPointData()->GetScalars();
vtkDataArray* labels=classification->GetPointData()->GetScalars();
double* sum =new double[numclasses],*sum2 =new double[numclasses];
int* count=new int[numclasses];
for (int ia=0;ia<numclasses;ia++) {
sum[ia]=0.0; sum2[ia]=0.0; count[ia]=0; }
int numvoxels=intens->GetNumberOfTuples();
for (int i=0;i<numvoxels;i++) {
double v=intens->GetComponent(i,0);
int l=(int)labels->GetComponent(i,0);
if (l>=0 && l < numclasses) {
sum[l]+=v; sum2[l]+=v*v; count[l]+=1; }
}
double totalnp=double(numvoxels);
for (int i=0;i<numclasses;i++) {
double numv=count[i]; params->SetComponent(i,0,numv);
double mean=sum[i]/numv; params->SetComponent(i,1,mean);
double mean2=sum2[i]/numv; params->SetComponent(i,2,sqrt(mean2-mean*mean));
}
}
The Maximization Step: The M-Step is a little bit more involved. We rst note that we are using rst-order
neighborhoods for the MRF, i.e. the voxels immediately adjacent in the x,y and z-directions. Hence we can
update the labels in two passes by thinking of the image as a chess-board. We rst update all the black squares
(keeping the labels of the white squares xed) and then update all the white squares (keeping the black squares
xed). In addition, we randomize the order in which we start.
This method (DoMaximizationStep) can be simply thought of as a smart way to call the ClassifyVoxel method
(described next) in the appropriate order.
double vtkMRFSegmentation::DoMaximizationStep(vtkImageData* intensity_image,
vtkImageData* label_image,
vtkDoubleArray* params, int numclasses,
double smoothness,int maxiter)
{
vtkDataArray* intensities=intensity_image->GetPointData()->GetScalars();
vtkDataArray* labels=label_image->GetPointData()->GetScalars();
int nt=intensities->GetNumberOfTuples();
int dim[3]; intensity_image->GetDimensions(dim);
int incr[6]; double weights[6];
this->ComputeMRFIncrementsAndWeights(intensity_image,incr,weights);
int done=0,iter=0; int tenth=nt/11;
double sumchanged=0.0;
while (done==0 && iter<maxiter) {
double total=0.0, changed=0.0;
int order=(vtkMath::Random()>0.5);
for (int pass=0;pass<=1;pass++) {
int realpass=pass;
230
CHAPTER 21. IMPLEMENTING INTENSITY BASED SEGMENTATION Draft Sep. 2009
if (order==1) realpass=1-pass;
// Omitted Code, loop over k (z-axis), j (y-axis)
// Compute voxel index
int vox=k*incr[5]+j*incr[3]+1;
for (int i=1;i<dim[0]-1;i++) {
// Check for voxel color (i.e. black square or white square)
// If yes update
if ( this->UpdateVoxel(eveni,evenj,evenk,pass)==1) {
changed+=this->ClassifyVoxel(vox,intensities,labels,params,
numclasses,incr,weights,smoothness);
++total;
}
++vox;
....
}
changed=100.0*changed/total; sumchanged+=changed;
if (changed<this->ConvergencePercentage) done=1;
++iter;
}
return sumchanged;
}
Computing the Posterior: The hard work of the M-Step is done by the ClassifyVoxel method. This takes
one voxel, specied by voxel_index and computes its posterior probability (equation 21.6). for each possible
class assignment. Then, we set the class label for this voxel to be the one that corresponds to the class that
maximimizes the posterior probability:
int vtkMRFSegmentation::ClassifyVoxel(int voxel_index,vtkDataArray* intensities,
vtkDataArray* labels,vtkDoubleArray* params,
int numclasses,int incr[6],double wgth[6],
double smoothness)
{
int current_label=(int)labels->GetComponent(voxel_index,0);
int bestclass=0;
double bestprob=0.0;
double v=intensities->GetComponent(voxel_index,0);
for (int cl=0;cl< numclasses;cl++) {
double prob=this->ComputeTotalMinusLogProbability(v,voxel_index,cl,labels,
params,numclasses,incr,
wgth,smoothness);
if (cl==0 || prob<bestprob) {
bestprob=prob; bestclass=cl;
}
}
if (bestclass!=current_label) {
labels->SetComponent(voxel_index,0,bestclass); return 1;
}
return 0;
}
231
CHAPTER 21. IMPLEMENTING INTENSITY BASED SEGMENTATION Draft Sep. 2009
The probability is computed by the ComputeTotalMinusLogProbability method. This is also straightforward
as it delegates all the work to two other functions (i) ComputeLogLikelihoodProability the data term and (ii)
ComputeLogMRFPrior the smoothness term. We examine these, in order, next:
double vtkMRFSegmentation::ComputeTotalMinusLogProbability(double intensity,
int current_voxel, int current_label,
vtkDataArray* labels,vtkDoubleArray* parameters,
int numclasses, int incr[6],double wgth[6],double smoothness)
{
double mlterm=this->ComputeLogLikelihoodProbability(intensity,current_label,parameters);
double pmrf =smoothness*this->ComputeLogMRFPrior(labels,current_voxel,
current_label,numclasses,incr,wgth);
return -mlterm+pmrf;
}
Computing the Log Likelihood: The log likelihood probability has an interesting twist. We model each voxel as
having intensity y = x + n, where x is the true intensity and n is the image noise (zero mean, gaussian). The
distribution of x depends on its current label l and has mean
l
and standard deviation
l
. The distribution of
y has mean
l
and variance
2
l
+
2
n
where
n
is the standard deviation of the noise. This added term is useful
because it adds stability to the estimation, i.e. all classes will have standard deviation at least equal to sigma
n
.
Finally, as before, the distribution is implemented in a separate function, Metric, so that it can easily be changed.
double vtkMRFSegmentation::ComputeLogLikelihoodProbability(double intensity,
int current_label,vtkDoubleArray* params)
{
double totalprob=0.0;
double mean=params->GetComponent(current_label,1);
double sigma=params->GetComponent(current_label,2);
double variance=sigma*sigma+this->ImageNoiseVariance;
totalprob=log(this->Metric(intensity,mean,variance));
return totalprob;
}
Computing the MRF Term: This implements the MRF model desribed in equation 21.3. The only added twist is
that we weight the eect of each voxel based on its distance from the current voxel. (If the images are isotropic
all weights will take value=1).
double vtkMRFSegmentation::ComputeLogMRFPrior(vtkDataArray* labels,int vox,
int current_label,int numclasses,
int incr[6],double wgth[6])
{
double sum=0.0;
for (int i=0;i<=5;i++) {
int l=(int)labels->GetComponent(vox+incr[i],0);
if (l!=current_label) sum+=wgth[i];
}
232
CHAPTER 21. IMPLEMENTING INTENSITY BASED SEGMENTATION Draft Sep. 2009
}
return sum;
}
Other Functions: The Metric function implements the Gaussian distribution. The UpdateVoxel method determines
whether a voxel is on the right color of the checkerboard and hence whether it should be updated (this is
called from within the DoMaximizationStep method). Finally the ComputeMRFIncrementsAndWeights method
computes both the osets (in raster-voxel order) and the weights of neighboring voxels for use in the MRF model
computation.
Compiling using CMAKE
These classes are placed in a library called MRFSegm. The CMakeLists.txt le is straight-forward and will not be
described here. The libraries can be loaded into Tcl using the script loadmrfsegm.tcl shown below:
package provide loadmrfsegm 1.0
if { $tcl_platform(platform) == "windows" } {
set name debug/vtkMRFSegmTCL.dll
} else {
set name libvtkMRFSegmTCL.so
}
load $name;unset name
21.3 A Complete Intensity-Based Segmentation Application
We implement this as a BioImage Suite application, shown in Figure 21.1, following the guidelines in chapter
14. The implementation consists of two les, the main script (segmtool.tcl) and the segmentation control class
(mrfutility.tcl). We examine these in turn next.
The Main Application
This is dened in the le segmtool.tcl and is similar to the mytool.tcl script described in chapter 14. There
is nothing exciting here, other than the fact that we will use the Objectmap viewer so that we can overlay the
results of the segmentation on the original image, if we choose to.
lappend auto_path [ file dirname [ info script ]]
package require loadbioimagesuite 1.0
package require pxitclbaseimageviewer 1.0
package require mrfutility 1.0
# Eliminate the default tk window
wm withdraw .
# Initialize a base application with some default settings
233
CHAPTER 21. IMPLEMENTING INTENSITY BASED SEGMENTATION Draft Sep. 2009
Figure 21.1: The Main application with the added MRFSegmentation menu.
# See bioimagesuite/main/pxitclbaseimageviewer.tcl for all the options
set baseviewer [ pxitclbaseimageviewer \#auto 0 ]
$baseviewer configure -appname "BioImage Suite::Intensity Segmentation Tool"
# Omitted code .. define default choices of controls
$baseviewer configure -show_standard_images 1
....
# Create an Objectmap Viewer
$baseviewer InitializeObjectmapViewer .[ pxvtable::vnewobj ] 1
# Add a submenu for our Segmentation Control
set menubase [ $baseviewer cget -menubase ]
set mb [ menu $menubase.[pxvtable::vnewobj] -tearoff 0 ]
$menubase add cascade -label "MRFSegmentation" -menu $mb -underline 0
# Create the segmentation control, add it to the menu and register it with
# the viewer (addcontrol)
set myutil [ mrfutility \#auto $baseviewer ]
$myutil Initialize [ $baseviewer GetBaseWidget ].[ pxvtable::vnewobj ]
$myutil AddToMenuButton $mb
$baseviewer AddControl $myutil
# Omitted code -- the rest is similar to mytool.tcl (chapter16)
234
CHAPTER 21. IMPLEMENTING INTENSITY BASED SEGMENTATION Draft Sep. 2009
Figure 21.2: The three tabs of the Segmentation Control.
The Segmentation Control
The Segmentation Control is implemented as an [ Incr ] Tcl class deriving from pxitclbaseimagecontrol. This
consists of three tabs each having controls for setting the parameters for one of the three algorithms described
earlier. The three tabs are shown in Figure 21.2
In the class header there is a statement:
package require loadmrfsegm
for loading our newly implemented methods into the Tcl interpreter.
The class denition is as follows. It consists of essentially ve types of methods, (i) the constructor/destructor
pair, (ii) initialization methods (iii) methods for generating the GUI for the individual methods called by the
initialization methods, (iii) a method to add this control the menu and (iv) methods to perform the segmentation
and deal with the results.
itcl::class mrfutility {
inherit pxitclbaseimagecontrol
protected common thisparam
constructor { par } { pxitclbaseimagecontrol::constructor $par } { InitializeControl }
destructor { }
# initialization methods
public method InitializeControl { }
public method Initialize { inpwidg }
# interface creation methods
protected method CreateOutputImageGUIControl { guiname name widget }
protected method CreateMultiThresholdControl { name }
protected method CreateKMeansControl { name }
protected method CreateMRFControl { name }
# Add this control to a Menu Button
public method AddToMenuButton { mb args }
# Computational Utility Stuff
public method CheckImage { image name operation verbosity }
public method ComputeMultiThresholdSegmentation { }
public method ComputeKMeansSegmentation { }
public method ComputeMRFSegmentation { }
235
CHAPTER 21. IMPLEMENTING INTENSITY BASED SEGMENTATION Draft Sep. 2009
# Deal with Results
public method ProcessResult { image guiname opname }
}
Initialization Code: Nothing special here. The Initialize function creates a toplevel widget, packs a notebook
in it and then calls three helper methods to add the controls for each tab.
itcl::body mrfutility::InitializeControl { } {
# Omitted Code which sets some default parameters
}
itcl::body mrfutility::Initialize { widget } {
if { $initialized == 1 } { return $basewidget }
set basewidget [toplevel $widget ]; wm geometry $basewidget 610x450
set notebook $basewidget.notebook; iwidgets::tabnotebook $notebook -tabpos w
CreateMultiThresholdControl [ $notebook add -label "MultiThreshold" ]
CreateKMeansControl [ $notebook add -label "KMeans" ]
CreateMRFControl [ $notebook add -label "MRF" ]
pack $notebook -side top -fill both -expand t -padx 5
set initialized 1; SetTitle "Intensity Segmentation Tool"
eval "wm protocol $basewidget WM_DELETE_WINDOW { wm withdraw $basewidget }"
return $basewidget
}
Creating the GUI for each Method: Three very similar methods (CreateMultiThresholdControl, CreateK-
MeansControl and CreateMRFControl) create the interface for each algorithm. We will look at one of them in
detail CreateKMeansControl, the others are very similar.
This control consists of three parts. At the top there is a series of iwidgets::entryfield widgets for setting
the values of the input parameters. In the middle there is a button (base.but) for executing the segmentation,
and at the bottom there is a pxitclimageGUI control for storing the output segmentation.
itcl::body mrfutility::CreateKMeansControl { base } {
iwidgets::labeledframe $base.frame0 -labelpos nw \
-labeltext "KMeans Segmentation"
pack $base.frame0 -fill both -expand f -pady 5
set w [ $base.frame0 childsite ]; set k 0
iwidgets::entryfield $w.$k -labeltext "Classes:" \
-textvariable [ itcl::scope thisparam($this,kmeans_numclasses) ] \
-width 2 -validate integer
pack $w.$k -side top -expand true -fill x; incr k
# Omitted code for entryfileds for Iterations, Bins etc.
236
CHAPTER 21. IMPLEMENTING INTENSITY BASED SEGMENTATION Draft Sep. 2009
# Create a compute button
eval "button $base.but -text \"Compute K-Means Segmentation\" \
-command { $this ComputeKMeansSegmentation }"
pack $base.but -side top -expand t -fill x
# Create a pxitclimageGUI to store output -- more later
set widg [ $this CreateOutputImageGUIControl "kmeansoutput" "K-Means Output" $base.bot ]
pack $widg -side bottom -expand f -fill x
}
The Image GUIs are created by the CreateOutputImageGUIControl Function below. We add two functions, one
to display the segmentation as an image (Display) and one to display it as an overlay (Display Mask).
itcl::body mrfutility::CreateOutputImageGUIControl { guiname name widget } {
set thisparam($this,$guiname) [ [ pxitclimageGUI \#auto ] GetThisPointer ]
$thisparam($this,$guiname) configure -description $name
$thisparam($this,$guiname) Initialize $widget
set bbut [ $thisparam($this,$guiname) cget -browsebutton ]; pack forget $bbut
$thisparam($this,$guiname) AddFunction "$parent SetResultsFromObject" "Display" "$this"
$thisparam($this,$guiname) AddFunction "$parent SetMaskFromObject" "Display Mask" "$this"
return $widget
}
Computational Code: The three algorithms are executed using the methods ComputeMultiThresholdSeg-
mentation, ComputeKMeansSegmentation and method ComputeMRFSegmentation respectively. As before, we
will look at one of these in detail, ComputeKMeansSegmentation. We rst check that the current image of the
viewer (stored in the currentimage variable) exists. If it does, we create the vtkKMeansSegmentation object
and set its parameters. The segmentation is invoked using the Update method. The result is then handled using
the ProcessResult method. Nothing complicated here:
itcl::body mrfutility::ComputeKMeansSegmentation { } {
# Omitted code to check that currentimage exists
WatchOn
set segm [ vtkKMeansSegmentation New ]
$segm SetInput [ $currentimage GetImage ]
$segm SetNumberOfBins $thisparam($this,kmeans_numbins)
$segm SetNumberOfClasses $thisparam($this,kmeans_numclasses)
# ...
$segm Update
WatchOff
$this ProcessResult [ $segm GetOutput ] "kmeansoutput" "km"
$segm Delete
}
The fancy coding is in the ProcessResult method. This does three things: (i) it copies the segmentation result
237
CHAPTER 21. IMPLEMENTING INTENSITY BASED SEGMENTATION Draft Sep. 2009
into the image currentresults, (ii) it sends the segmentation result to the viewers mask image (the overlay) and
(iii) it stores it into the appropriated pxitclimageGUI for later use.
itcl::body mrfutility::ProcessResult { image guiname opname } {
# Step 1
$currentresults ShallowCopyImage $image
$currentresults configure -filename [ AddPrefix [ $currentimage cget
-filename ] $opname ]
# Step 2
$parent SetMaskFromObject $currentresults $this
# Step 3
set gui $thisparam($this,$guiname)
[ $gui GetObject ] ShallowCopy $currentresults; $gui Update
}
238
Draft Sep. 2009
Chapter 22
More Advanced Image Algorithm/Filters
and Templated Execution
In the rst part of this chapter we discuss the overall algorithm/lter pipeline for image processing in VTK
working our way from vtkAlgorithm to vtkImageAlgorithm and certain specic examples. Next we describe
the implementation of multi-threaded templated image-to-image lter implementation. Such implementations,
while somewhat more complex, can oer substantial speed improvements over the more simple lters previously
described.
22.1 Introduction
In previous chapters we have considered simple image-to-image lters deriving from the class vtkSimpleImage-
ToImageFilter. A common aspect of all the lters we discussed was the reliance on convenience methods such
as GetComponent, SetComponent, SetScalarComponentFromDouble and GetScalarComponentAsDouble for
manipulating images. These convenience methods hide the actual underlying data type of the image. They, at
least to the programmer, treat images the same regardless of whether the image has type short or oat.
1
This convenience often comes at the cost of computational overhead. Ultimately the fastest way to access data
is by direct pointer manipulation. This makes the programming arcane and practically returns the programmer
to the age of assembly language. However, such operations can dramatically improve the computational speed of
many operations. The usual path is to implement a class using the convenience methods and if it becomes too
critical in a large system, re-implement it carefully, once and for all, with lower level operations.
Modern C++ is beginning to provide an additional mechanism to encapsulate pointer operations with less over-
head, using a type of class called iterators. Many iterators are dened in the Standard Template Library (STL),
which unfortunately we will not have time to get into this semester. However, VTK denes two simple iterators
that we will explore in the context of the example to follow.
Iterators leverage templating for execution speed. An iterator is programmed in a generic (not type specic)
setup, unlike the Get/Set methods above which take double inputs. Templated implementations are used to
generate specic versions of the iterators for specic types such as oats, ints etc.
In this chapter we will explore the VTK pipeline in more depth using concrete VTK classes as examples.
1
Naturally in the Set methods if the value specied is not valid for the type, we have unexpected results. For example,
consider setting the value of a voxel in a short image to 10.2. This will most likely be truncated to 10 as short images do
not handle decimal points. More interesting artifacts occur when the value is outside the range of the type. For example,
consider the setting the value of a voxel in an 1-byte image to 300.
239
CHAPTER 22. MORE ADVANCED IMAGE ALGORITHM/FILTERS AND TEMPLATED
EXECUTION Draft Sep. 2009
22.2 The VTK Object Hierarchy
22.2.1 The vtkAlgorithmClass and its Helpers
As stated in the VTK manual, vtkAlgorithm is the parent class for all sources, lters, and sinks in VTK. It
denes a generalized interface for executing data processing algorithms. Pipeline connections are associated with
input and output ports that are independent of the type of data passing through the connections. Instances may
be used independently or within pipelines with a variety of architectures and update mechanisms. Pipelines are
controlled by instances of vtkExecutive. Every vtkAlgorithm instance has an associated vtkExecutive when it is
used in a pipeline. The executive is responsible for data ow.
A general vtkAlgorithm can have an arbitrary number of inputs and outputs of varying types (e.g. vtkImageData,
vtkPolyData etc). In general on both the input and output sides of the algorithm, the algorithm has ports (i.e.
InputPorts and OutputPorts) which are used to connect this to the rest of the pipeline. These enable the setting
of inputs and the getting of the outputs of the algorithm. The algorithm has as many input ports as it has inputs
and as many output ports as it has outputs. For example a lter that takes two images as input, adds them and
outputs the sum as an output will have 2 input ports and 1 output port.
vtkInformation: The information about the data types is store in a vtkInformation object (which we saw 3
chapters back briey). This contains things, in the case of images, such as the potential output dimensions,
output spacing etc. of the algorithm.
vtkInformation objects are stored together in vtkInformationVector objects. Each vtkAlgorithm has two arrays of
these namely:
vtkInformationVector* InputPortInformation;
vtkInformationVector* OutputPortInformation;
These arrays contain one vtkInformationVector for each port (either input, or output respectively). Each vtkIn-
formationVector in turn contains a number of vtkInformation objects (one per connection to the port).
Lets refer back now to chapter 19, as per the code below:
int vtkImageExtractVOI::RequestInformation(vtkInformation *request,
vtkInformationVector **inputVector,
vtkInformationVector *outputVector)
{
vtkInformation *inInfo = inputVector[0]->GetInformationObject(0);
vtkInformation *outInfo = outputVector->GetInformationObject(0);
Here inInfo is the vtkInformation object for input port 0, connection 0. Similarly outInfo is the equivalent for
output port 0, connection 0. In many cases, (e.g. classes derived from vtkImageAlgorithm such as vtkSimpleIm-
ageToImageFilter), we have by default a single input port and a single output port. This makes life simpler.
One we have a pointer to the appropriate vtkInformation object, then the lter can make appropriate decisions,
provide necessary information about what it will do. The best documentation for this, unfortunately, is in the
source code.
240
CHAPTER 22. MORE ADVANCED IMAGE ALGORITHM/FILTERS AND TEMPLATED
EXECUTION Draft Sep. 2009
Accessing and Manipulating the Ports: Key aspects of the interface of vtkAlgorithm for communicating
through the ports involves the methods (abbreviated from vtkAlgorithm.h) below:
1. vtkInformation* GetInputPortInformation(int port); this returns the information object asso-
ciated with an input port. There is one input port per kind of input to the algorithm. Each input port tells
executives what kind of data and downstream requests this algorithm can handle for that input.
2. vtkInformation* GetOutputPortInformation(int port); Get the information object associated
with an output port. There is one output port per output from the algorithm. Each output port tells
executives what kind of upstream requests this algorithm can handle for that output.
3. int GetNumberOfInputPorts(); and int GetNumberOfOutputPorts(); These output the number
of input and output ports:
4. vtkDataObject* GetOutputDataObject(int port); Get the data object that will contain the algo-
rithm output for the given port.
5. vtkDataObject *GetInputDataObject(int port, int connection); Get the data object that
will contain the algorithm input for the given port and given connection.
6. virtual void SetInputConnection(int port, vtkAlgorithmOutput* input); and
void SetInputConnection(vtkAlgorithmOutput* input); These methods represent the new way
to set the inputs of a lter. The old style SetInput methods are still supported in many derived classes.
2
7. vtkAlgorithmOutput* GetOutputPort(int index); and vtkAlgorithmOutput* GetOutputPort().
These return the output of the lter. Note that these are not vtkDataObjects but richer structures con-
taining more pipeline info.
8. vtkDataObject* GetOutputDataObject(int port) this method can be used to get the actual output
vtkDataObject. Many derived class also denes the more legacy style GetOutput function.
9. vtkDataObject *GetInputDataObject(int port, int connection); this is the input equivalent
of the previous method.
22.2.2 Imaging Filters and the VTK Pipeline
Almost all Image-related algorithms derive directly or indirectly from the vtkImageAlgorithm class. The exception
to this rule is old-style legacy code (from VTK 4.4) which derives from vtkImageToImageFilter (whose ancestry
is vtkAlgoritm vtkProcessObject vtkSource vtkImageSource vtkImageToImageFilter!)
Implementing an Image-to-image lter or perhaps an image reader or writer class (e.g. reading in Analyze les)
is most easily done by deriving from one of the following classes.
vtkImageAlgorithm this is the generic parent class for everything else.
vtkImageReader2 this is a good parent class for classes that read in images.
vtkImageWriter this is a good parent class for classes that write images to le.
vtkSimpleImageToImageFilter this is the simplest way to do this. It operates on the whole image (i.e. no
multithreading) we have seen examples of this before.
vtkThreadedImageAlgorithm This can handle threaded execution where the image is split into pieces and
the processing done in parallel on separated threads this is ideal for processing where each pixel is more
less treated independently.
Some Special Functionality in vtkImageAlgorithm and Handling the Pipeline: vtkImageAlgorithm
is more specialized subclass of vtkAlgorithm, where the inputs and outputs are assumed to be of type vtkImageData
(i.e. images!). Some key new methods here are (as dened in vtkImageAlgorithm.h):
vtkImageData* GetOutput();
vtkImageData* GetOutput(int);
virtual void SetOutput(vtkDataObject* d); Get the output data object for a port on this algorithm.
2
As best as I can tell, if you are creating your vtkDataObjects manually and then using them as inputs to classes, you
still need to use SetInput type methods.
241
CHAPTER 22. MORE ADVANCED IMAGE ALGORITHM/FILTERS AND TEMPLATED
EXECUTION Draft Sep. 2009
void SetInput(vtkDataObject *);
void SetInput(int, vtkDataObject*);
Set an input of this algorithm.
The pipeline interface is dened by the three methods:
int RequestInformation(vtkInformation* request,
vtkInformationVector** inputVector, vtkInformationVector* outputVector);
int RequestUpdateExtent(vtkInformation*,vtkInformationVector**,vtkInformationVector*);
int RequestData(vtkInformation *request,
vtkInformationVector** inputVector,vtkInformationVector* outputVector);
Essentially these three methods handle the pipeline update process.
Consider the case of a pipeline consisting of three algorithms A B C.
1. The rst step in the pipeline the so called Request Information pass i.e. the downstream lter inquires
of the upstream object as to what it can/will produce. In this case C calls the RequestInformation method
of B and similarly B calls the method of A.
2. The second step in the pipeline is the so called RequestUpdateExtent pass where the upstream object
tells the downstream object now requests a specic piece of the output (if threaded execution is used).
Again C asks for a piece of data from B which then asks for the necessary piece of data from A etc.
3. Finally, in the third step the RequestData step, the data is produced (i.e. the actual ltering operation!)
and fed forward from A to B to C.
For a lter to connect to the pipeline properly it must be able to handle all three of these steps. vtkImageAlgorithm
denes a basic implementation for these which is then overridden in derived classes.
22.2.3 A concrete example from vtkSimpleImageToImageFilter
1. RequestInformation This is actually dened in vtkImageAlgorithm. It assumes that the output image
will have exactly the same properties as the input image, hence things just get copied over!
int vtkImageAlgorithm::RequestInformation(
vtkInformation* request,
vtkInformationVector** inputVector,
vtkInformationVector* outputVector)
{
// do nothing except copy scalar type info
this->CopyInputArrayAttributesToOutput(request,inputVector,outputVector);
return 1;
}
2. RequestUpdateExtent this lter can only produce 1 piece (i.e. the whole image, no threading is
available), so we tell the down stream lter that this is all it can get!
int vtkSimpleImageToImageFilter::RequestUpdateExtent (
vtkInformation * vtkNotUsed(request),
vtkInformationVector **inputVector,
vtkInformationVector *vtkNotUsed( outputVector ))
242
CHAPTER 22. MORE ADVANCED IMAGE ALGORITHM/FILTERS AND TEMPLATED
EXECUTION Draft Sep. 2009
{
vtkInformation *inInfo = inputVector[0]->GetInformationObject(0);
// always request the whole extent
inInfo->Set(vtkStreamingDemandDrivenPipeline::UPDATE_EXTENT(),
inInfo->Get(vtkStreamingDemandDrivenPipeline::WHOLE_EXTENT()),6);
return 1;
}
3. RequestData in vtkSimpleImageToImageFilter, this method serves to take the complex VTK pipeline
interface and reduce it to calling a simpler method SimpleExecute which does the work. This is dened in the
derived classes, see examples from previous chapters.
virtual void SimpleExecute(vtkImageData* input, vtkImageData* output) = 0;
The top part of the code extracts the information from the pipeline and checks the validity of the input. Then
the output data is allocated nally SimpleExecute is called.
int vtkSimpleImageToImageFilter::RequestData(
vtkInformation* vtkNotUsed( request ),
vtkInformationVector** inputVector,
vtkInformationVector* outputVector)
{
// get the data object
vtkInformation *outInfo = outputVector->GetInformationObject(0);
vtkImageData *output = vtkImageData::SafeDownCast(
outInfo->Get(vtkDataObject::DATA_OBJECT()));
vtkInformation *inInfo = inputVector[0]->GetInformationObject(0);
vtkImageData *input = vtkImageData::SafeDownCast(
inInfo->Get(vtkDataObject::DATA_OBJECT()));
int inExt[6];
input->GetExtent(inExt);
// if the input extent is empty then exit
if (inExt[1] < inExt[0] || inExt[3] < inExt[2] || inExt[5] < inExt[4])
{ return 1; }
// Set the extent of the output and allocate memory.
output->SetExtent(
outInfo->Get(vtkStreamingDemandDrivenPipeline::WHOLE_EXTENT()));
output->AllocateScalars();
thisSimpleExecute(input, output);
243
CHAPTER 22. MORE ADVANCED IMAGE ALGORITHM/FILTERS AND TEMPLATED
EXECUTION Draft Sep. 2009
return 1;
}
22.3 A Threaded Algorithm vtkImageShiftScale
The vtkThreadedImageAlgorithm class is the parent class from which multithreaded (i.e. parallel, able to run
using multiple threads concurrently) VTK image processing algorithms are commonly derived from. The code
in this is not for the faint-hearted. We simply note that essentially this class reroutes the pipeline such that its
derived classes need to dene a method:
virtual void ThreadedRequestData(vtkInformation *request,
vtkInformationVector **inputVector,
vtkInformationVector *outputVector,
vtkImageData ***inData,
vtkImageData **outData,
int extent[6], int threadId);
which will handle the requests for updates on pieces of the output image. We will look at a concrete example,
vtkImageShiftScale.
22.3.1 The Header vtkImageShiftScale.h
vtkImageShift takes an image I as input and outputs an image J of the form J = (I + a) b, where a is
the shift and b is called the scale. This makes it embarrassingly parallel-izable, since each voxel can more or
less be processed individually. This lter derives from vtkThreadedImageAlgorithm. The header le is fairly
straight-forward:
#include "vtkThreadedImageAlgorithm.h"
class vtkImageShiftScale : public vtkThreadedImageAlgorithm
{
public:
static vtkImageShiftScale *New();
vtkTypeRevisionMacro(vtkImageShiftScale,vtkThreadedImageAlgorithm);
void PrintSelf(ostream& os, vtkIndent indent);
// Description:
// Set/Get the shift value.
vtkSetMacro(Shift,double);
vtkGetMacro(Shift,double);
// Description:
// Set/Get the scale value.
vtkSetMacro(Scale,double);
vtkGetMacro(Scale,double);
244
CHAPTER 22. MORE ADVANCED IMAGE ALGORITHM/FILTERS AND TEMPLATED
EXECUTION Draft Sep. 2009
// Description:
// Set the desired output scalar type. The result of the shift
// and scale operations is cast to the type specified.
vtkSetMacro(OutputScalarType, int);
vtkGetMacro(OutputScalarType, int);
void SetOutputScalarTypeToDouble()
{this->SetOutputScalarType(VTK_DOUBLE);}
// Omitted Code ...
// Description:
// When the ClampOverflow flag is on, the data is thresholded so that
// the output value does not exceed the max or min of the data type.
// By default, ClampOverflow is off.
vtkSetMacro(ClampOverflow, int);
vtkGetMacro(ClampOverflow, int);
vtkBooleanMacro(ClampOverflow, int);
protected:
vtkImageShiftScale();
~vtkImageShiftScale();
double Shift;
double Scale;
int OutputScalarType;
int ClampOverflow;
virtual int RequestInformation(vtkInformation*,vtkInformationVector**,
vtkInformationVector*);
virtual void ThreadedRequestData(vtkInformation*,vtkInformationVector**,
vtkInformationVector*,vtkImageData*** inData,
vtkImageData** outData,
int outExt[6],
int threadId);
private:
vtkImageShiftScale(const vtkImageShiftScale&); // Not implemented.
void operator=(const vtkImageShiftScale&); // Not implemented.
};
The key things to note are the two methods near the bottom namely (i) RequestInformation and (ii) Thread-
edRequestData. The rst is trivial in this case and is used to set the output scalar type (if dierent from the
input).
int vtkImageShiftScale::RequestInformation(vtkInformation*,
vtkInformationVector**,
vtkInformationVector* outputVector)
{
// Set the image scalar type for the output. -1 = same as input, so do nothing
if(this->OutputScalarType != -1)
{
245
CHAPTER 22. MORE ADVANCED IMAGE ALGORITHM/FILTERS AND TEMPLATED
EXECUTION Draft Sep. 2009
vtkInformation* outInfo = outputVector->GetInformationObject(0);
vtkDataObject::SetPointDataActiveScalarInfo(
outInfo, this->OutputScalarType, -1);
}
return 1;
}
22.3.2 Templated Execution
Unlike ITK, which we will discuss in the next chapter, VTK does not use templates in the class interface. This
results in the VTK lter classes not being templated in themselves. However, VTK does use templates in the
implementation of some of the code for eciency. Since, however, the classes themselves are not templated, such
lters cannot have templated member functions. The solution to this quandary is the use of ordinary, specially
named templated procedures, inserted in the VTK source which are called from within the member functions.
These procedures are often specially named to reduce the likelihood of naming conicts at link time. (Naming
conicts occur when more than one function/procedure has the same exact name and argument list). In general
their names begin with the class name of the class whose methods will call them. This is best illustrated by means
of the following example.
In vtkImageShiftScale, the input point to the lter is the ThreadedRequestData method (which is the rough
equivalent of SimpleExecute in the classes deriving from vtkSimpleImageToImageFilter).
This then calls an ordinary function called vtkImageShiftScaleExecute1 which is templated by the type of the
input image. This in turn calls a second function vtkImageShiftScaleExecute which is templated over both
the type of the input image and the type of the output image. The reason for this double hop, is that we use
(implicitly) switch statements to check for data type. Doing this in a single step, would require N
2
cases (where
N is the number of possible data types) whereas doing this in two steps reduces the coding complexity to N
followed by N more (= 2N total cases).
The key to understanding the code below is to understand the operation of the vtkTemplateMacros.
Step 1: ThreadedRequestData: This is the main method of the lter. The method takes seven inputs: (i)
Inputs 1-3 are the necessary information vectors etc., inputs 4-5 are the input and output arrays.
More interestingly, outExt[6] is an array species the part of the image to update and threadId this is the
thread id.
In multi-threaded lters, the execution the lter can be split over a number of processors, if these are available.
To accomplish this, the output is split into pieces, and each thread is responsible for computing the result for its
piece. For example, in the case of this lter, vtkImageShiftScale, thread 1 may be computing the output for the
top half of the image and thread 2 for the bottom half. (Naturally not all operations can be multi-threaded.).
outExt[6] denes the region over which this thread is responsible. It is a six-component array of the form
(xmin, xmax, ymin, ymax, zmin, zmax).
void vtkImageShiftScale::ThreadedRequestData(vtkInformation*,
vtkInformationVector**,
vtkInformationVector*,
vtkImageData*** inData,
vtkImageData** outData,
int outExt[6],
int threadId)
{
246
CHAPTER 22. MORE ADVANCED IMAGE ALGORITHM/FILTERS AND TEMPLATED
EXECUTION Draft Sep. 2009
vtkImageData* input = inData[0][0];
vtkImageData* output = outData[0];
switch(input->GetScalarType())
{
vtkTemplateMacro(
vtkImageShiftScaleExecute1(this, input, output, outExt, threadId,
static_cast<VTK_TT*>(0)));
default:
vtkErrorMacro("ThreadedRequestData: Unknown input ScalarType");
return;
}
}
The messy part of this function is the use of the vtkTemplateMacro. Expanding this macro would result (approx-
imately) in code of the form:
switch (inData->GetScalarType()) {
// BEGIN MACRO
case double:
vtkImageShiftScaleExecute1(this,input,output,outExt,id,static_cast<double *>0);
break;
case float:
vtkImageShiftScaleExecute1(this,input,output,outExt,id,static_cast<float *>0);
break;
// OMMITTED CODE
// similarly for long, unsigned long, int, unsigned int, short,
// unsigned short, char and unsigned char
// END MACRO
default:
vtkErrorMacro(<< "Execute: Unknown ScalarType");
return;
}
Also note, that since vtkImageShiftScaleExecute1 is not a member function of vtkImageShiftScale, we pass
the this pointer as an explicit argument, so that this function can access class data members for more information
otherwise they would all have to be passed as parameters.
Step 2: vtkImageShiftScaleExecute1
Please note that this function is not a member of vtkImageShiftScale. It is an ordinary c-like function which simply
happens to be in the same le as the implementation of vtkImageShiftScale. This enable its use without necessary
needing to declare it in any header le this follows the solid principles of PIMPL or private implementation!
template <class T>
void vtkImageShiftScaleExecute1(vtkImageShiftScale* self,
vtkImageData* inData,
vtkImageData* outData,
int outExt[6], int id, T*)
247
CHAPTER 22. MORE ADVANCED IMAGE ALGORITHM/FILTERS AND TEMPLATED
EXECUTION Draft Sep. 2009
{
switch (outData->GetScalarType())
{
vtkTemplateMacro(
vtkImageShiftScaleExecute(self, inData,
outData, outExt, id,
static_cast<T*>(0),
static_cast<VTK_TT*>(0)));
default:
vtkErrorWithObjectMacro(
self, "ThreadedRequestData: Unknown output ScalarType");
return;
}
}
This function is templated using type T, which is the type of the input image (e.g. short, oat etc.). It then uses
the vtkTemplateMacro macro which creates another set of case statements where this time the conditional is on
the output type.
This switch method ends up calling the nal function vtkImageShiftScaleExecute, where the actual work will take
place.
Step 3: vtkImageShiftScaleExecute
vtkImageShiftScaleExecute is were the seemingly trivial work of addition and multiplication takes place. This
is a double-templated function over input type IT and output type OT. Its rst argument is a pointer to the
instance of the class from which it was called, whereas the rest are information about the operation. The last
two arguments are zeros and are unused, the important aspect of these last two arguments is that they explicitly
dene the input and output types for the templating.
The code follows I have rearranged it marginally for greater clarity. First we get the key class parameters from
the calling class, using the self pointer:
template <class IT, class OT>
template <class IT, class OT>
void vtkImageShiftScaleExecute(vtkImageShiftScale* self,
vtkImageData* inData,
vtkImageData* outData,
int outExt[6], int id,
IT*, OT*)
{
// Get the shift and scale parameters values.
double shift = self->GetShift();
double scale = self->GetScale();
// Clamp pixel values within the range of the output type.
int clamp = self->GetClampOverflow();
Next we get the range of types for the current type. These would be 0 and 255 in the case of unsigned char, 0
to 65535 for short etc.
248
CHAPTER 22. MORE ADVANCED IMAGE ALGORITHM/FILTERS AND TEMPLATED
EXECUTION Draft Sep. 2009
// for preventing overflow
double typeMin = outData->GetScalarTypeMin();
double typeMax = outData->GetScalarTypeMax();
Next we create two iterators. Iterators are special classes (originally dened in the C++ Standard Template
Library) that allow the easy looping through data structures of arbitrary complexity. Iterators are the modern
way of direct pointer manipulation.
In this case, the iterators are not pointers but are simply allocated on the stack (i.e. like ordinary variables). The
rst iterator (inIT) is simply used to traverse through the input image, whereas the second iterator (outIt) does
the same job for the second image. The second iterator is of type vtkImageProgressIterator and in addition to
looping over the image, periodically calls the self->UpdateProgress() method to keep the user informed of
the process of the lter.
The iterator breaks the image region (dened by outExt) over which it will iterate into a number of continuous
(in memory) parts called spans. For example if we have as input a 161616 image and outExt=[0, 7, 0, 7, 0, 7],
the rst span would be from voxel (0, 0, 0) to (7, 0, 0). At this point, we will need to jump to (0, 1, 0) and begin
a second span.
// Create iterators for the input and output extents assigned to
// this thread.
vtkImageIterator<IT> inIt(inData, outExt);
vtkImageProgressIterator<OT> outIt(outData, outExt, self, id);
// Loop through output pixels.
while (!outIt.IsAtEnd())
{
IT* inSI = inIt.BeginSpan();
OT* outSI = outIt.BeginSpan();
OT* outSIEnd = outIt.EndSpan();
if (clamp)
{
while (outSI != outSIEnd)
{
// Pixel operation
double val = (static_cast<double>(*inSI) + shift) * scale;
if (val > typeMax)
{
val = typeMax;
}
if (val < typeMin)
{
val = typeMin;
}
*outSI = static_cast<OT>(val);
++outSI;
++inSI;
}
}
else
{
249
CHAPTER 22. MORE ADVANCED IMAGE ALGORITHM/FILTERS AND TEMPLATED
EXECUTION Draft Sep. 2009
while (outSI != outSIEnd)
{
// Pixel operation
*outSI = static_cast<OT>((static_cast<double>(*inSI) + shift) * scale);
++outSI;
++inSI;
}
}
inIt.NextSpan();
outIt.NextSpan();
}
}
The work of the lter is done by the middle portion. The input value can be accessed by pointer dereferencing
(.e.g. *inSI). This could be written in longhand as:
// Get input
IT inval = (IT) (*inSI);
// Perform the operation in double
double val = ((double)(*inSI) + shift) * scale;
// If we are checking for clamping verify legal range
if (clamp){
if (val > typeMax)
val = typeMax;
if (val < typeMin)
val = typeMin;
}
// Set the Output
*outSI = (OT)(val);
Note: The two functions vtkImageShiftScaleExecute1 and vtkImageShiftScaleExecute do not appear in any header
(.h) le. They are an example of private implementation, and are only known to other functions in the same source
(.cpp) le. However, in this case, the order in which the functions appear in the .cpp le is important. A function
can not call a function that has not already been dened or at the very least its interface has been dened.
Hence, in the source le, vtkImageShiftScaleExecute appears rst followed by vtkImageShiftScaleExecute1 and
vtkImageShiftScale::ThreadedExecute appears last.
Unfortunately, for many lters, the VTK iterators are not suciently powerful. Many lters (e.g. vtkImageGaus-
sianSmooth) rely on direct pointer manipulations instead.
22.4 STL Notes:
If you want understand the operations of these iterators better, I suggest reading through some of the tutorials
on the STL e.g. http://www.cprogramming.com/tutorial/stl/iterators.html. The STL denes many
interesting classes, that can be of great use. It is, unfortunately beyond the scope of this introductory material.
250
Draft Sep. 2009
Chapter 23
Copying Data Objects and Breaking the
Pipeline
23.1 Introduction
Most of the VTK examples you will see are essentially single pipelines. This gives the appearance that the only
way to process data in VTK is to take raw data as an input, pass it through various lters for manipulation and
display the output in a single connected pipeline. In larger projects, however, we often need to do a small amount
of processing and return the result. Later this result, perhaps as a response to some user input via a GUI, will
then be processed some more to yield a dierent output etc.
In this chapter we discuss ways of breaking the single pipeline and returning intermediate results. Key to this
process is the ability to copy data-objects. We rst discuss three dierent methods for doing this at varying
levels of completenesss. This is followed by two concrete examples. The rst shows how to properly extract an
isosurface from an image and the second how to compute a gradient of smoothed image. In both examples, the
focus is on proper implementation and returning of the data-object as opposed to the details of the algorithms
themselves.
23.2 CopyStructure, ShallowCopy and DeepCopy
It is often desirable to make a copy of a dataset. Datasets such as vtkPolyData and vtkImageData support, at
least, three dierent means of copying. These are:
1. CopyStructure: This copies the geometric structure of an input data-set. In the case of vtkImageData
this simply copies the image dimensions, origin, spacing, number of scalar components and data type (and
a few other miscellaneous members). It essentially creates an image of the same size as the input, but
does not allocate memory. This is very useful as an initialization of a lter method, where the output image
is of the same size and type as the input image:
vtkImageData* out=vtkImageData::New();
out->CopyStructure(input);
// Now modify as desired
out->SetScalarTypeToFloat();
// At this point only allocate memory
out->AllocateScalars();
251
CHAPTER 23. COPYING DATA OBJECTS AND BREAKING THE PIPELINE Draft Sep. 2009
2. ShallowCopy: ShallowCopy is essentially the same as CopyStructure with the key addition that all array
data (e.g. intensities) are linked! Shallow Copy for a vtkImageData (in a very simplied form)
vtkImageData* out=vtkImageData::New();
out->ShallowCopy(input);
// or equivalently
out->CopyStructure(input);
out->GetPointPointData()->SetScalars(input->GetPointData()->GetScalars());
All pointer-based objects, which include all vtkDataArray structures are simply passed as pointers (with
appropriate reference count increases). No new memory is allocated to store these. Any modication to
the intensities in out, also modies the intensities in input as their intensities are stored in the exact
same array!
ShallowCopy is extremely useful for preserving the results of a lter or a combination of lters (a pipeline)
while destroying the actual pipeline. This is illustrated in the two examples later in this document.
3. DeepCopy: DeepCopy creates a complete duplicate version of a data-object. This allocates all necessary
memory and creates a separate but identical object to the input.
vtkImageData* out=vtkImageData::New();
out->DeepCopy(input);
Use DeepCopy sparingly, unless you absolutely need a complete duplicate copy of an image prior to modi-
cation.
23.3 Example 1: Extracting a Surface from a LevelSet Function
Consider, for example, an implementation of a Levelset segmentation algorithm, which results in a distance
map (or levelset function) stored in an image LevelsetImage. Either during the evolution of the levelset, or at
the end, we may need to return a surface extracted from the zero-levelset. (The zero-levelset conventionally
represents the output surface of the segmentation.) This may be conveniently implemented in a member function
GetZeroSurface.
A rst attempt at implementing this function can take the form:
vtkPolyData* vtkMyLevelsetFilter::GetZeroSurface() {
// this->LevelsetImage is of type vtkImageData
vtkContourFilter *ContourFilter = vtkContourFilter::New();
ContourFilter->SetInput(this->LevelsetImage);
ContourFilter->SetValue(1, 0.0);
ContourFilter->Update();
return ContourFilter->GetOutput();
}
This will work, but there is a key problem associated with it. The problem is that we return a pointer to the
output of a lter, without returning the actual lter itself. This can mess up the reference counting scheme in
VTK.
252
CHAPTER 23. COPYING DATA OBJECTS AND BREAKING THE PIPELINE Draft Sep. 2009
Consider the case where we call this lter once. The lter ContourFilter is rst created. Then we set the image
this->LevelsetImage as its input, which results in the reference counter of the image being incremented by
one.
Then we call the lters Update function. This, incidentally, is critical. VTK pipelines operate on a lazy executing
scheme, so the lter will not do anything unless it has to, using Update forces the lter to go to work. (In normal
pipelines, i.e. source to display, updating the display propagates an update event backwards through the pipeline
and forces all intermediate lters to update!).
The lter results in a surface (vtkPolyData) which is then returned to the user.
Next time we call the lter, we create a new ContourFilter object and set this->LevelsetImage as its input,
which results in the reference counter of the image, again, being incremented by one.
If we call this function 100 times then LevelsetImage will have a reference count of 100 which means that it
will never be deleted even when the LevelsetFilter class is deleted, resulting in a potential memory leak in a large
piece of software.
The correct solution to this problem, is to copy the result of the lter output, delete the lter, and return this
copy.
vtkPolyData* vtkMyLevelsetFilter::GetZeroSurface()
{
vtkContourFilter *ContourFilter = vtkContourFilter::New();
ContourFilter->SetInput(this->LevelsetImage);
ContourFilter->SetValue(1, 0.0);
ContourFilter->Update();
vtkPolyData* zerosurface=vtkPolyData::New();
zerosurface->ShallowCopy(ContourFilter->GetOutput());
ContourFilter->Delete();
return zerosurface;
}
Here, we rst create a temporary surface (zerosurface). Next we perform a shallow copy operation which copies
the contents of the output of the ContourFilter to this temporary surface. Then the ContourFilter is deleted,
cleaning up all reference counting issues. At this point we return the zerosurface object to the calling code.
The calling function is then responsible to delete the zerosurface object when it is done with it, this object has
no attachments to any lingering pipeline code.
The Recipe: In the many cases where one needs to use a VTK pipeline to generate an output data structure
(as opposed to an output display or le) and return it, the following recipe can be very useful:
Create the pipeline
Call the Update function of the last lter this will invoke, in turn, the Update functions of all the
previous lters
Create a new output structure (e.g. image or surface most likely)
ShallowCopy the output of the nal lter to the new output structure.
Delete all lters in the pipeline.
Return the output structure.
253
CHAPTER 23. COPYING DATA OBJECTS AND BREAKING THE PIPELINE Draft Sep. 2009
A slight variation on the above example, in which the level is specied as opposed to assumed to be zero, is given
in the class vtkMyUtility.cpp.
23.4 Example 2: An Image Processing Example
Consider the case where one needs to compute the gradient of an image at a specic scale. This requires (i) rst
smoothing the image and (ii) computing the gradient. This operation can be accomplished by a pipeline shown
below:
vtkImageData* vtkMyUtility::SmoothImageAndComputeGradient(vtkImageData* input,
double sigma,int dimensionality){
vtkImageGaussianSmooth* sm=vtkImageGaussianSmooth::New();
sm->SetInput(input);
sm->SetStandardDeviations(sigma,sigma,sigma);
sm->SetDimensionality(dimensionality);
vtkImageGradient* gradient=vtkImageGradient::New();
gradient->SetInput(sm->GetOutput());
sm->Delete();
gradient->SetDimensionality(dimensionality);
gradient->Update();
vtkImageData* grad=vtkImageData::New();
grad->ShallowCopy(gradient->GetOutput());
gradient->Delete();
return grad;
}
Note that we follow the same recipe. The pipeline is rst created. Then, the Update function of the last lter
gradient->Update() is called to force execution. Next, we create a new output data structure (grad) and
perform the shallow copy operation. The pipeline is deleted as usual and the output image is then returned.
23.5 Implementation
Both of these examples are implemented in a class vtkMyUtility. The header le of this has the form:
class vtkMyUtility : public vtkObject {
public:
static vtkMyUtility *New();
vtkTypeMacro(vtkMyUtility,vtkObject);
// Example 1 -- Extract Iso-Contour
vtkPolyData* ExtractContour(vtkImageData *img,double level);
// Example 2 -- Smooth Image and Compute Gradient
static vtkImageData* SmoothImageAndComputeGradient(vtkImageData* img,double sigma,
int dimensionality);
protected:
254
CHAPTER 23. COPYING DATA OBJECTS AND BREAKING THE PIPELINE Draft Sep. 2009
};
A script (script23-1.tcl) exercises these functions and outputs a surface and a gradient image respectively.
23.6 Note:
There is a bug in VTK 5.2 in the ShallowCopy method of vtkImageData. We have this patched in the version of
VTK you get with BioImage Suite 3 and are in the process of submitting a x to VTK itself.
255
Draft Sep. 2009
Chapter 24
The Insight Toolkit
24.1 Introduction
The Insight Toolkit (ITK) is an open source software toolkit for registration and segmentation. It is implemented
in C++ and uses the CMake build environment in fact CMake was developed for the ITK project. ITK was
started in 1999 under a contract by the US National Library of Medicine of the National Institutes of Health to
a consortium of academic and industrial partners.
The ITK software guide [?], which is available at the ITK webpage www.itk.org, has a good description of the
algorithmic content of ITK. A set of presentations submitted to the Insight Journal is another good resource
(http://hdl.handle.net/1926/161).
While ITK can be thought o as a rst-cousin of VTK, there is one critical dierence. ITK is implemented
using generic programming principles. It uses templates both for the algorithm implementation and, unlike VTK,
the class interfaces themselves. This type of heavily templated C++ code challenges many compilers and it can
take much longer to compile. The other dierence, is that the memory model depends on smart pointers that
maintain a reference count to objects. Smart pointers can be allocated on the stack, and when scope is exited,
the smart pointers disappear and decrement their reference count to the object that they refer to. There is no
need to call itkFilter->Delete(), unlike VTK lters.
The use of ITK, especially by beginners, is more challenging than VTK. The use of generic programming techniques
assumes a rm grounding in templated programming in general, and the Standard Template Library in particular.
The use of templated lters also makes the use of the toolkit from languages other than C++ less elegant than
VTK.
The use of templates in the interface has one negative consequence. While the use of templated classes can
simplify the lter design there is no need for multiple switch statements as is the case in VTK templated
implementations it results in the need to explicitly specify image types at compile time. In contrast, in VTK
one can allocate a vtkImageData object rst, then dynamically set its type, and potentially even change its type
later. In ITK images need to be allocated explicitly with a xed type at compile time, which is a limitation in
using it to develop larger systems.
To elaborate this further, consider the case of allocating an image. In VTK, this is accomplished as:
vtkImageData* img=vtkImageData::New();
img->SetScalarTypeToFloat();
256
CHAPTER 24. THE INSIGHT TOOLKIT Draft Sep. 2009
By contrast, in ITK, the same task requires either:
itk::Image< float , 3 >::Pointer img=itk::Image< float , 3 >::New();
where 3 is the image dimension and oat is the image type. Alternatively, we can use the typedef construct to
create a shorthand for the complex type. This results in the following code segment:
typedef itk::Image< float , 3 > ImageType
typename ImageType::Pointer img=ImageType::New();
As you can see, things can get ugly pretty quickly.
1
Learning to program using ITK, i.e. implementing algorithms by leveraging ITK code, is beyond the scope of this
class. However, there are a lot of algorithms in ITK that can be usefully exploited for many tasks. In addition ITK
has a nice image I/O framework which supports a large number of image formats. In the rest of this handout,
we will focus on using ITK in a VTK-centric environment. This will be accomplished by hiding ITK code inside
functions which take VTK images as inputs and return VTK images as outputs. Within such functions, we can
convert the images to ITK data structures, do the operation and convert. In this way ITK code is safely packaged
in a VTK wrapper.
The techniques we will use to accomplish this task are similar to those described in our discussion on templated
VTK lters. We will again use ordinary specially named templated functions, inserted in VTK source which are
called from within the member functions. All ITK code will reside primarily within these templated functions.
24.2 The vtkITKMyUtility Class
This is a simple class which implements three static member functions: (i) CurvatureAnisotropicDiusion this
calls an ITK smoothing lter by the same name, (ii) LoadImage this can be used to load an image using the
ITK IO Factory and (iii) SaveImage this can be used to save an image. In the case of the last two functions, the
le type is automatically determined by the image name! Note that the header below has no traces of anything
to do with ITK:
#include <vtkObject.h>
class vtkImageData;
class vtkITKMyUtility : public vtkObject {
public:
static vtkITKMyUtility *New();
vtkTypeMacro(vtkITKMyUtility,vtkObject);
static vtkImageData* CurvatureAnisotropicDiffusionFilter(vtkImageData* input,
double Conductance=1.0,
double TimeStep=0.15,int NumberOfIterations=8);
1
It must be acknowledged that the ability to have code that can handle images of arbitrary types and dimensions has
its own aesthetic appeal. However, inside the code all the types are of some user dened type which can be dizzying until
one gets used to it.
257
CHAPTER 24. THE INSIGHT TOOLKIT Draft Sep. 2009
static vtkImageData* LoadImage(char* fname);
static int SaveImage(vtkImageData* input,char* filename);
protected:
vtkITKMyUtility() {};
virtual ~vtkITKMyUtility() {};
};
Each of the three static member functions calls special templated utility ordinary functions placed in vtkITKMyU-
tility.cpp to do the templated operations. Within these ordinary functions we will make use of two convenience
classes that come as part of a supplementary distribution called InsightApplications that is also available from the
ITK web-page. These classes are VTKImageToImageFilter and ImageToVTKImageFilter. The notation used here
is that the word Image signies an ITK image, whereas the word VTKImage is used to signify a vtkImageData
structure.
These lters make use of pairs of classes known as importers and exporters. These can export and import an
image from/to a naked C-like pointer. VTK has a pair of these classes called vtkImageExport and vtkImageImport.
They convert from VTK Images to ITK Images and back.
24.3 Curvature Anisotropic Diusion Filtering
This is a type of nonlinear image smoothing that tries to smooth uniform areas while preserving sharp disconti-
nuities. It is available in ITK as part of the itk::CurvatureAnisotropicDiffusionImageFilter class.
2
The Filtering Function: The real work is handled by the following doubly-templated function note that
this is not a member of vtkITKMyUtility but just an ordinary function:
template <class IT,int dimension>
void vtkITKMyUtilityCurvatureAnisotropicDiffusionSmoothImage(vtkImageData* input,
vtkImageData* output,
double Conductance,
double TimeStep,
int NumberOfIterations)
{
// Define the parts (types) of the ITK pipeline.
typedef itk::Image<IT, dimension> ImageType;
typedef itk::VTKImageToImageFilter<ImageType> VTKImageToImageFilterType;
typedef itk::ImageToVTKImageFilter<ImageType> ImageToVTKImageFilterType;
typedef itk::CurvatureAnisotropicDiffusionImageFilter<ImageType, ImageType > FilterType;
// From VTK to ITK
typename VTKImageToImageFilterType::Pointer importer=VTKImageToImageFilterType::New();
importer->SetInput(input);
importer->Update();
// Create the itk::CurvatureAnisotropicDiffusionImageFilter and connect it
typename FilterType::Pointer filter = FilterType::New();
2
ITK also makes use of C++ namespaces these are similar to the Tcl namespaces that we looked at earlier.
258
CHAPTER 24. THE INSIGHT TOOLKIT Draft Sep. 2009
filter->SetInput(importer->GetOutput());
filter->SetTimeStep(TimeStep);
filter->SetNumberOfIterations(NumberOfIterations);
filter->SetConductanceParameter( Conductance );
filter->Update();
// Back to VTK
typename ImageToVTKImageFilterType::Pointer exporter=ImageToVTKImageFilterType::New();
exporter->SetInput(filter->GetOutput());
exporter->Update();
output->DeepCopy(exporter->GetOutput());
}
The code is fairly straightforward, once one gets over the messy type denitions at the top. To avoid constantly
writing things like Image<IT,dimension>, we create shorthands for all the types at the top using the typedef
operator.
Next we convert the input image, which comes as a vtkImageData to an itkImage using a properly templated
instance of VTKImageToImageFilter. Following this we are into straight ITK code as lifted from one of the
examples that came with ITK. This naturally results in an itkImage which we then convert back to VTK using
an instance of ImageToVTKImageFilter.
Finally the result is copied to an output image using a DeepCopy operation which ensures that the data will still
be around once the pipeline is deleted. Crossing over toolkits has overheads this is one of them.
Note that we do not delete the ITK objects. ITK uses smart pointers which essentially delete themselves. ITK
came into being a few years after VTK and in some respects has some signicant improvements. This is one of
them.
An Intermediate Function: The function above is called by an intermediate function that takes care of the
second template argument, the image dimension. This is fairly straightforward:
template <class IT>
void
vtkITKMyUtilityCurvatureAnisotropicDiffusionSmoothImage1(vtkImageData* indata,
vtkImageData* outdata,double Conductance,double TimeStep,
int NumberOfIterations,int dimension,IT *)
{
if (dimension==2)
vtkITKMyUtilityCurvatureAnisotropicDiffusionSmoothImage<IT,2>(indata,outdata,Conductance,
TimeStep,NumberOfIterations);
else
vtkITKMyUtilityCurvatureAnisotropicDiffusionSmoothImage<IT,3>(indata,outdata,Conductance,
TimeStep,NumberOfIterations);
}
The Class Member Function: The two functions above are buried inside vtkITKMyUtility.cpp and are
inaccessible to the outside world. Outside code access the ltering operation through the following class member
259
CHAPTER 24. THE INSIGHT TOOLKIT Draft Sep. 2009
function:
vtkImageData*
vtkITKMyUtility::CurvatureAnisotropicDiffusionFilter(vtkImageData* input,
double Conductance,double TimeStep,int NumberOfIterations)
{
if (input==NULL)
return NULL;
vtkImageData* output=vtkImageData::New();
output->CopyStructure(input);
int dimension=3;
int dim[3]; input->GetDimensions(dim);
if (dim[2]==1)
dimension=2;
switch (input->GetScalarType())
{
vtkTemplateMacro7(vtkITKMyUtilityCurvatureAnisotropicDiffusionSmoothImage1, input,output,
Conductance,TimeStep,NumberOfIterations,dimension,
static_cast<VTK_TT *>(0));
}
return output;
}
This rst creates the output image. Then it checks whether the input image is really a 2D Image. If it is, the
dimension variable is set to 2 otherwise it stays at 3. Finally, we use the vtkTemplateMacro7 to call the previous
ordinary function. The vtkTemplateMacros were discussed in more detail in chapter 20.
24.4 The LoadImage Function
ITK has a very nice Image Factory IO setup. The user only needs to specify an image lename. The factory
instatiates the appropriate Image Reader class based on the lename (e.g. Analyze, TIF, PNG) and loads the
image.
The workhorse function: As before, we will use a combination of an ordinary templated function and a
non-templated member function to perform the operation. First, the actual method that loads the image:
template <class IT>
vtkImageData* vtkITKMyUtilityLoadImage(char* fname)
{
typedef itk::Image< IT, 3 > ImageType;
typedef itk::ImageToVTKImageFilter<ImageType> ImageToVTKImageFilterType;
typedef itk::ImageFileReader< ImageType > ReaderType;
260
CHAPTER 24. THE INSIGHT TOOLKIT Draft Sep. 2009
typename ReaderType::Pointer reader = ReaderType::New();
reader->SetFileName(fname);
reader->Update();
typename ImageToVTKImageFilterType::Pointer exporter=ImageToVTKImageFilterType::New();
exporter->SetInput(reader->GetOutput());
exporter->Update();
vtkImageData* output=vtkImageData::New();
output->DeepCopy(exporter->GetOutput());
return output;
}
This is similar to the smoothing lter above, other than for the fact that here we only have an output.
The Interface Function: The outside class accesses the Load Image functionality through the following
member function. This has two parts: (i) First, we read the image information to identify the image type using
a trick posted on the ITK-users mailing list. (ii) Then an explicit switch statement is used to call the workhorse
function above with the proper template argument:
vtkImageData* vtkITKMyUtility::LoadImage(char* filename)
{
// Some of this code derives from code posted by Hideaki Hiraki
// on the Insight-users mailing list
itk::ImageIOBase::Pointer imageIO;
imageIO=itk::ImageIOFactory::CreateImageIO(filename, itk::ImageIOFactory::ReadMode);
imageIO->SetFileName(filename);
imageIO->ReadImageInformation();
switch( imageIO->GetComponentType() ){
case itk::ImageIOBase::UCHAR:
return vtkITKMyUtilityLoadImage<unsigned char>(filename);
break;
case itk::ImageIOBase::CHAR:
return vtkITKMyUtilityLoadImage<char>(filename);
break;
case itk::ImageIOBase::SHORT:
return vtkITKMyUtilityLoadImage<short>(filename);
break;
case itk::ImageIOBase::FLOAT:
return vtkITKMyUtilityLoadImage<float>(filename);
break;
// .... Lots more case statements omitted to save space.
}
return NULL;
}
261
CHAPTER 24. THE INSIGHT TOOLKIT Draft Sep. 2009
The implementation of the Save Image function is similar to the smoothing lter and will not be discussed in any
detail here.
24.5 The CMakeLists.txt File
This is an example of a combined VTK/ITK project. Both VTK and ITK must be found.
PROJECT(VTKITK)
SET(KITBASE VTKITK)
SET(KIT vtk${KITBASE})
INCLUDE (${CMAKE_ROOT}/Modules/FindVTK.cmake)
FIND_PACKAGE(VTK REQUIRED)
IF (USE_VTK_FILE)
INCLUDE(${USE_VTK_FILE})
ENDIF(USE_VTK_FILE)
FIND_PACKAGE(ITK REQUIRED)
IF (USE_ITK_FILE)
INCLUDE(${USE_ITK_FILE})
ENDIF(USE_ITK_FILE)
INCLUDE_DIRECTORIES(${VTKITK_SOURCE_DIR})
SET (LIBRARY_OUTPUT_PATH ${VTKITK_SOURCE_DIR})
SET(LIBRARY_SRCS
vtkITKMyUtility.cpp)
LINK_LIBRARIES(
vtkCommon
vtkCommonTCL
${ITK_LIBRARIES})
INCLUDE(${VTK_CMAKE_DIR}/vtkWrapTcl.cmake)
VTK_WRAP_TCL2L (${KIT}TCL LIBRARY_TCL_SRCS ${LIBRARY_SRCS})
ADD_LIBRARY (${KIT}TCL SHARED ${LIBRARY_TCL_SRCS} ${LIBRARY_SRCS})
The script script24-2.tcl is used to exercise the code. This loads an image in the new NIfTI format, smooths
it and saves it out in analyze format!
lappend auto_path [ file dirname [ info script ]]
package require mylibload 1.0
mylibload vtkVTKITKTCL
set util [ vtkITKMyUtility New ]
262
CHAPTER 24. THE INSIGHT TOOLKIT Draft Sep. 2009
set inimg [ $util LoadImage axial_small.hdr ]
puts stderr "Image Loaded [ $inimg GetDimensions ]"
puts stderr "Calling Curvature Anistropic Diffusion Filter"
set img [ $util CurvatureAnisotropicDiffusionFilter $inimg 1.0 0.05 8 ]
puts stderr "Done [ $img GetDimensions ] on to saving"
$util SaveImage $img "itksmooth.nii.gz"
exit
24.6 An integrated lter vtkitkSignedDanielssonDistanceMapImage-
Filter
In this Section we describe an example of wrapping an ITK image-to-image lter inside a VTK image-to-image
lter. This demonstrates a means of leveraging existing ITK within a mostly VTK-based piece of software.
In general our approach consists of:
1. Writing a standard VTK-based lter deriving from vtkSimpleImageToImageFilter.
2. Invoking the ITK code from the SimpleExecute command of the lter. In our case this calls a separate
method RunFilter.
3. The Run Filter method simply checks whether the image is 2D or 3D and calls a templated function (not
a member function of the class) to actually do the processing.
The SignedDanielssonDistanceMapImageFilter takes in a binary image and computes an image whose value is
signed distance map from the boundary of the binary image. This is often useful as an initialization step for
Levelset algorithms. The value is negative inside the object and positive outside. The implementation of the
actual algorithm consists of 3 layers.
1. The SimpleExecute Method: This simply checks for inputs and invokes RunFilter.
void vtkitkSignedDanielssonDistanceMapImageFilter::SimpleExecute(vtkImageData* input,vtkImageData* output)
{
if (input==NULL)
{
vtkErrorMacro(<<"Bad Inputs to vtkitkSignedDanielssonDistanceMapImageFilter");
return;
}
this->RunFilter(input,output);
}
2. The RunFilter Method: This checks whether the image is 2D or 3D and invokes a standard function
vtkitkSignedDanielssonDistanceMapImageFilter RunFilter which is templated over the image dimension (either 2
or 3).
int vtkitkSignedDanielssonDistanceMapImageFilter::RunFilter(vtkImageData* input,vtkImageData* output)
263
CHAPTER 24. THE INSIGHT TOOLKIT Draft Sep. 2009
{
int dim[3]; input->GetDimensions(dim);
if (dim[2]==1)
return vtkitkSignedDanielssonDistanceMapImageFilter_RunFilter<2>(input,output,this);
return vtkitkSignedDanielssonDistanceMapImageFilter_RunFilter<3>(input,output,this);
}
3. The Actual Code: This has six parts:
1. First dene all the types this is part of the joy of running ITK.
2. Take VTK image make it oat and threshold this both accomplished using vtkImageThreshold,
3. Convert VTK image output to ITK image,
4. Run the ITK Code,
5. Convert the ITK output to VTK (note the DeepCopy needed at the end!).
6. Cleanup all VTK lters, ITK lters use smart pointers so they take care of themselves!
1. Create the denitions: The typedef command used to rename messy things to simpler things down the road.
}
template <int ImageDimension>
int vtkitkSignedDanielssonDistanceMapImageFilter_RunFilter(vtkImageData* input,vtkImageData* output,vtkitkSignedDanielssonDistanceMapImageFilter* self)
{
typedef float PixelType;
typedef itk::Image<PixelType,ImageDimension> ImageType;
typedef itk::VTKImageImport<ImageType> ImportType;
typedef itk::VTKImageExport<ImageType> ExportType;
typedef itk::SignedDanielssonDistanceMapImageFilter<ImageType,ImageType> DMapFilterType;
2. Threshold and make oat: The input image is thresholded to ensure that it is binary. We pass the value of
the this pointer to this non-class member function as self so that we can access the necessary parameters for
thresholding. The output of the threshold lter is a oat binary image.
}
vtkImageThreshold* thresholder=vtkImageThreshold::New();
thresholder->SetInput(input);
thresholder->SetOutputScalarTypeToShort();
thresholder->SetOutValue(self->GetOutValue());
thresholder->SetInValue(self->GetInValue());
thresholder->ThresholdBetween(self->GetLowerThreshold(),self->GetUpperThreshold());
thresholder->SetOutputScalarTypeToFloat();
thresholder->ReplaceInOn();
thresholder->ReplaceOutOn();
3. Convert VTK to ITK: This is boilerplate as before.
264
CHAPTER 24. THE INSIGHT TOOLKIT Draft Sep. 2009
}
vtkImageExport *vtkInputExporter = vtkImageExport::New();
vtkInputExporter->SetInput(thresholder->GetOutput());
typename ImportType::Pointer itkInputImporter = ImportType::New();
ConnectPipelines(vtkInputExporter, itkInputImporter);
itkInputImporter->Update();
4. Run the ITK Code: This is the heart of the function. Instantiate the lter, set the input and update.
}
typename DMapFilterType::Pointer dmap = DMapFilterType::New();
dmap->SetInput(itkInputImporter->GetOutput());
dmap->Update();
5. Convert ITK to VTK: This is also boilerplate as before. Note the DeepCopy at the end. This is required as a
ShallowCopy will point to ITK data-structures which will be automatically deleted at the end of the function.
}
// Step 5: Push it out to VTK
typename ExportType::Pointer itkExporter = ExportType::New();
itkExporter->SetInput(dmap->GetOutput());
itkExporter->Update();
vtkImageImport* vtkImporter=vtkImageImport::New();
ConnectPipelines(itkExporter, vtkImporter);
vtkImporter->Update();
output->DeepCopy(vtkImporter->GetOutput());
6. Cleanup: We delete all VTK lters. The ITK lters are self-deleting.
}
// Step 6: Cleanup VTK filters only
vtkImporter->Delete();
thresholder->Delete();
vtkInputExporter->Delete();
}
The use of this class is demonstrated in script script24-3.tcl.
265
CHAPTER 24. THE INSIGHT TOOLKIT Draft Sep. 2009
24.7 Concluding Remarks
The Insight Toolkit (ITK) is another large object-oriented library that oers a great deal of functionality for medical
image analysis. It has among others, some very nice implementations of the Levelset method, Finite Element
Code, Registration code etc. Unfortunately, in my opinion, the use of a fully generic programming style requires
a level of C++ expertise that makes it dicult to recommend whole-heartedly to a beginner. The methodology
presented in this chapter aims to demonstrate how one can take advantage of ITK code by neatly packaging it
inside VTK.
266
Draft Sep. 2009
Part VII
Appendices
267
Draft Sep. 2009
Appendix A
Introduction to Computer Graphics -
Part I
Computer Graphics has become an integral part for visualizing any type of data (images, surfaces etc.). In the rst
chapter, we will discuss some basic computer graphics concepts such as a camera, dierent types of projections,
camera options in VTK and go over a quick introduction to texture mapping.
A.1 Introduction
Computer graphics has now become an integral part of entertainment (games, movies) as well as research for
visualizing data (medical, seismic, oil exploration, atmospheric, computational uid dynamics and so on).
In this chapter, we will look at a brief introduction to computer graphics and focus on topics that you will most
probably run into during this course or in your research.
A.2 Cameras
In computer graphics, to create a visual representation of a scene or a dataset, we need to actually place a camera
in the scene or in front of the dataset. A real world camera analogy will help us understand some of the similarities
and dissimilarities with the synthetic camera in the computer graphics world.
In computer graphics, as in the real world, a scene is dened by objects in the scene. But, just as you need to
position your camera and point it at a subset of objects in the scene to take a photograph of the scene, you need
to specify the position of the virtual camera and point it at objects/data before you can see a visual representation
of it. Very often one ends up with a blank screen due to the fact that the camera position is correct but it is
pointing in the opposite direction to the where the objects are relative to the camera. Just as in the real world,
you can move around an object and take photographs of it from dierent angles, you can move around data and
examine dierent facets of it by looking at its visual representation on the screen. [Note: In computer graphics, if
we want to take a photograph of an object from a dierent viewpoint, we can either move the camera around the
object or rotate/transform the object in front of the static camera to see dierent facets of the object.] Therefore,
we can move, rotate, orient and snap the shutter of the virtual camera to capture a 2D image.
In actuality, a camera in computer graphics is nothing but a piece of code that produces an image for you. As you
can imagine, the camera has its own coordinate system (u, v, n) in order to facilitate transformations around
objects in the scene which are dened in their own coordinate system (x, y, z).
268
APPENDIX A. INTRODUCTION TO COMPUTER GRAPHICS - PART I Draft Sep. 2009
Figure A.1: The left column describes concepts in the real world camera whereas the right column describes concepts in
the virtual/computer graphics world.
269
APPENDIX A. INTRODUCTION TO COMPUTER GRAPHICS - PART I Draft Sep. 2009
Figure A.2: The gure shows a schematic of the projector rays that emanate from the center of projection (center of
the camera) and pass through each point of an object to intersect the projection plane
A.3 Viewing Pipeline
Naturally generating an image using a virtual camera in computer graphics is a bit more than merely pushing
a button on your digital camera. But, the considerations that one needs to keep in mind during the process
of generating an image are very similar to what one would consider during the process of actually composing a
photograph in the real world.
For image acquisition/generation, we need to specify the following:
Specication of projection type
Specication of viewing parameters (camera position, direction in which the camera is pointing etc.)
Clipping in 3D (objects outside the visual frame/view volume are clipped out)
Projection and Display (Orthographic/Perspective projection and the nal transformations with respect to
display the image on the computer screen).
Let us now closely look at what each of the above items entails.
A.3.1 Projection
In general, projection consists of transforming an n-dimensional system into a system with a dimension less than
n. Since in computer graphics, the scenes and objects are mostly 3-dimensional, we shall focus on the 3D to 2D
projection where the three dimensional world is projected onto a viewing plane to create a 2D image representation
of the scene.
Projection Rays The projection of 3D objects onto a image/viewing plane can be dened by projection rays
called Projectors. These projectors start from a center of projection (the eye of the viewer or the center of the
camera). They pass through each point of an object and intersect a projection plane to form the projection.
Figure A.2.
Center of Projection The center of projection for a camera is generally a nite distance away from the
projection plane. For some projections (orthographic), it is convenient to think of the center of projection as
innitely away. Figure A.3 shows dierent options for center of projection where the projectors could either be
convergent or parallel.
270
APPENDIX A. INTRODUCTION TO COMPUTER GRAPHICS - PART I Draft Sep. 2009
Figure A.3: This gure shows two dierent types of projectors and their implication on the center of projection. The
left image shows the case for perspective projection, whereas the right image shows orthographic projection.
Figure A.4: The left and middle gure show dierent views of the 1-Point perspective projection where lines parallel to
the X-, Y-axes do not converge. The right image shows a schematic of perspective projection with the center of projection,
projection plane and the normal to the projection plane.
A.3.1.1 Perspective Projection
Perspective projection can be loosely dened as the projection when any set of parallel lines (not parallel to
the projection plane) converge to a vanishing point. Depending on the number of vanishing points, there are
1-point perspective, 2-point perspective and 3-point perspective projections. In 1-point perspective projection,
lines parallel to x, y do not converge. Figure A.4 shows a side view (left image) and a front view (middle image)
where the vanishing point can be seen. A schematic of perspective projection is seen in the right image which
shows the center of projection, projection plane and the normal to the projection plane.
For perspective projection in computer graphics, the view volume is in the shape of a frustum after all the
parameters for the camera are dened. Everything inside the view volume gets displayed on the screen whereas
everything outside the view volume gets clipped/culled out. The view volume is dened by the extents of the
camera (left, right, top, bottom, near plane and far plane). Figure A.5 shows a diagrammatic representation of
the view volume in the shape of a frustum for perspective projection.
Figure A.5: The view volume for perspective projection is a frustum whose shape is dened by specifying the left, right,
top, bottom, near plane and far plane. Everything in the scene that falls in the view volume gets displayed on the screen
whereas everything outside the volume gets clipped/culled out.
271
APPENDIX A. INTRODUCTION TO COMPUTER GRAPHICS - PART I Draft Sep. 2009
Figure A.6: The left image shows a schematic of orthographic projection where the projectors are parallel and the center
of projection is at innity. The right image shows that the view volume in this projection is a rectangular parallelepiped
that is dened by the camera parameters.
Figure A.7: These images show the eect of varying the up vector, viewing direction and camera position on the nal
output image.
A.3.1.2 Orthographic Projection
Orthographic projection can be dened as the projection when the projectors are parallel to each other and the
center of projection is at innity. The left image in Figure A.6 shows a simple example of orthographic projection
of an object onto the image plane. The right image depicts the view volume for orthographic projection. The
view volume in this case is specied using left, right, top, bottom, near and far planes just as in perspective
projection. But, since the center of projection is at innity, the view volume is a rectangular parallelepiped. As
in perspective projection, everything inside the view volume is displayed on the screen whereas everything outside
the view volume gets clipped/culled out.
A.3.1.3 Some other camera parameters
In addition to the projection and camera position, the view direction as well as the up vector are also critical to
the correct generation of an image. The up vector can be thought of as the direction that one considers up in
the scene. If you are taking landscape photograph, your up vector is (0, 1, 0) whereas if you are taking a portrait
photo, the up vector will be (1, 0, 0) since now up is in the direction of the X-axis. Figure A.7 shows the eect
of varying the camera position, the up vector and the view direction.
272
APPENDIX A. INTRODUCTION TO COMPUTER GRAPHICS - PART I Draft Sep. 2009
Figure A.8: This shows the process of mapping a texture onto a polygon. The texture coordinates are always between
0 to 1 and are independent of the geometry. Note: For black and white printouts, the texture coordinates are below the
vertex coordinates and are between 0.0 and 1.0.
A.4 Computer graphics in VTK
VTK (Visualization toolkit) contains extensive functionality to display graphics and visualize data. We will focus
on a large part of the rest of the course on using VTK, so I shall not delve into too much detail other than the
computer graphics capabilities.
In VTK, they have used the Scene/Theatre paradigm in the virtual setting with Actors, Props, Lights and Camera.
Actors represent Images (2D/3D), Surfaces, Filter output, pretty much any data element. Props are the objects
added to the renderer to create a scene. Props generally specify information about the orientation, size, origin,
position of the Actor and are used extensively in VTK.
A.4.1 Camera in VTK
Now that we know all that we need to know about cameras, let us look at the available functionality in VTK.
VTK has a class called vtkCamera, that is basically a virtual camera for 3D rendering. It allows the users to
move, rotate, orient the camera around a focal point. It allows the user ne tuned control over the camera and
has some other functions to modify the current camera parameters to change camera settings on the y.
There are a lot more graphics/visualization features in VTK that we shall examine more closely later in the course.
A.5 Texture mapping
Texture mapping is the process of mapping an image/texture onto a geometric shape (triangle, polygon, sphere,
tetrahedron etc.) to increase the visual appeal of the model for realism. A texture is typically a 2-D image (though
it could be a 1D/3D image). Image elements are called texels (texture elements). The mapping of the texture
to a surface determines the correspondence, i.e., how the texture lies on the surface. As is probably evident, it is
much easier to map a texture to a triangle since texture coordinates can be easily specied for a triangle. On the
other hand, for a complex arbitrary 3-D shape, it is much harder to specify a one-to-one mapping between the
texture and points on the 3-D shape. Figure A.8 shows a schematic that describes the process of texture mapping.
A simple brick texture is mapped onto a polygon which has texture coordinates. The texture coordinates specify
the way in which the texture is glued onto the polygon.
There are many possible mappings that dene how a texture is mapped onto a surface. Some of the more popular
ones are
Orthogonal mapping - Straightforward one-to-one mapping without any transformations
273
APPENDIX A. INTRODUCTION TO COMPUTER GRAPHICS - PART I Draft Sep. 2009
Spherical mapping - Use to map a two dimensional image/texture onto a sphere.
Cylindrical mapping - To map a texture onto a cylinder.
Texture Chart - An explicitly dened mapping between points in a texture and the triangle that they map
to on an object.
274
Draft Sep. 2009
Appendix B
Introduction to Computer Graphics -
Part II
Computer Graphics has become an integral part for visualizing any type of data (images, surfaces etc.). In the
second chapter, we will discuss some computer graphics concepts such as dierent types of data, visualization
methods to visualize the data including volume rendering, colormaps and a quick introduction to lighting and
shading.
B.1 Introduction
Data for image processing can be either 2D/3D/4D image data or surfaces that are created or extracted from
the image data. Surface data can be visualized using simple computer graphics primitives such as points, lines
and polygons. 2D image data can be visualized using texture mapping techniques. 3D image data can be
visualized using isosurface rendering or direct volume rendering techniques. In order to visualize data, colors play
an important role in conveying dierences in the data and convey shape and structure cues. Colormaps are a way
to assign color in visualizations. Lighting and shading are extremely useful in conveying shape and structure of
data (especially 3D data).
B.2 Types of Data
Data is generally in the form of images or surfaces (also known as meshes). Images can be 2D, 3D or 4D (time
series data). Images are generally a 2D/3D/4D array of values representing data such as intensity, orientation of
tensors (DTI), activation in MR Spectroscopy or fMRI. Each individual element in a 2D image can be represented
by a pixel (picture element). Similarly, each element in a 3D image can be represented by a voxel (volume
element).
A Mesh is a collection of vertices in space connected by edges. The faces that are formed from connecting the
edges form a primitive (triangle, polygon etc.). A collection of such geometric primitives form a surface that is
then lit and shaded to show shape and structure.
Figure B.1 shows three dierent representations of the same data. The leftmost image shows a volume rendering
of the MNI head volume. The middle image shows a mesh extracted from the data while the rightmost image
shows a surface representation of the same data.
275
APPENDIX B. INTRODUCTION TO COMPUTER GRAPHICS - PART II Draft Sep. 2009
Figure B.1: Three dierent types of representations of the same data. The leftmost image is a volume rendering of
the MNI head volume. The middle image is a mesh extracted from the same 3D data. The rightmost image is a surface
representation of the mesh generated by drawing faces and shading them.
B.3 Visualizing data
To understand the data better we need to create a visual representation of the data. Visualizing meshes and
surfaces consists of drawing points, lines and geometric primitives such as triangles, polygons etc. The shape cues
are provided by lighting and shading, which we will discuss in section B.5 of the handouts in this chapter.
B.3.1 Visualizing images
2D images are visualized using texture mapping techniques which were introduced in the previous class. A polygon
is drawn on the screen and a texture is created from the data. The texture is then mapped onto the polygon for
display purposes. All the operations (such as image smoothing, edge detection etc.) are applied to the 2D image
and not to the polygon/texture. The texture map is then updated with the ltered image and the results shown
as a texture mapped polygon on the screen.
3D images are visualized by isosurface rendering or volume visualization techniques. An isosurface is a 3D analog
of an isocontour. It is a surface that represents points of a constant value (e.g. pressure, temperature, velocity,
density) within a volume of space; in other words it is a level set of a continuous function whose domain is
3D-space. In Figure B.1, the rightmost image is an isosurface rendering of the 3D data visualized in the leftmost
image.
B.3.1.1 Volume Rendering
Volume rendering is dened as the process of generating a 2D image directly from 3D volumetric data. Due to
the fact that a visual representation is obtained directly from 3D data, it is also sometimes referred to as Direct
Volume Rendering.
The idea behind volume rendering is to simulate the transport of light through matter. Some of the light gets
absorbed and some passes through. Based on the property of the material, which in our case is the intensity of
the voxel at each 3D location, some voxels let more light pass through than some others.
The volume rendering equation below is evaluated at each point along the ray
I =
_
D
0
c
_
s(x())
_
exp
_

_

0
(s(x(

)))d

_
d
where x() is the viewing ray parameterized by the distance to the viewpoint, c(s(x) is the color assigned by
the colormap and (s(x)) is the extinction coecient assigned to each possible material in the volume, D is the
276
APPENDIX B. INTRODUCTION TO COMPUTER GRAPHICS - PART II Draft Sep. 2009
Figure B.2: This image shows the four ray traversal methods (L to R): First, Average, MIP and Composite.
maximum distance that the ray traverses before it goes outside the volume.
This can be approximated to the following
C
i+1
= A
i
C
i
+ (1 A
i
) a
i
c
i
A
i+1
= A
i
+ (1 A
i
) a
i
where C
i
stands for the accumulated color along the ray and A
i
represents the accumulated opacity along the
ray. The current voxels color and opacity, which is obtained from the colormap, is represented as c
i
and a
i
respectively.
Ray traversal methods Based on dierent ray traversals, dierent visualizations are generated. The simplest
ray traversal method is rst-hit ray traversal method in which the rst non-zero voxel encountered is used for
color and opacity lookup. The second method averages the intensity along the voxels to produce an X-ray eect.
The third method is called Maximum Intensity Projection (MIP) where, as the name suggests, the maximum
intensity encountered along the ray is used to lookup the color and opacity for that pixel. MIP is widely used for
visualizing vessels in Magnetic Resonance Angiography. The last and most commonly used ray traversal method
is called Accumulate, where the color and opacity are accumulated along the way which allows for transparent
layers to be visible and takes into account multiple material properties. Figure B.2 shows the four dierent ray
traversal methods.
Volume rendering has been implemented in multiple ways to evaluate the volume rendering integral and produce
a visual representation of 3D data. We are going to look at raycasting, 2D and 3D texture-mapping based volume
rendering. Shear warp and splatting are the other popular volume rendering techniques.
Raycasting is a volume rendering technique in which for every pixel in the image, a ray is cast into the 3D space
containing the volume. Figure B.3 shows a schematic of the raycasting process. For every non-zero voxel, the
color and opacity for that voxel is looked up into a color lookup table (known as a transfer function and commonly
referred to as a colormap). The color and opacity are then combined with the accumulated color and opacity
along the ray. After the ray has gone through the volume, the accumulated color and opacity are assigned to
that pixel. The speed and quality of the raycasting algorithm is dependent on the steps of the ray as it traverses
through the volume. Newer advances in graphics hardware have made it possible to implemented a hardware
accelerated raycasting algorithm.
2D Texture mapping based volume rendering is a technique that uses the texture mapping capabilities of
graphics hardware to perform volume rendering. The volume is viewed as a set of slices parallel to the coordinate
planes. Polygons are drawn in front of each other with each polygon being textured with dierent slices of the
data. A blending operation is then performed to obtain a visual representation of the data. Figure B.4 shows a
schematic where the 2D polygon slices, the texture mapped polygons and the blended volume rendered image is
shown.
There are 3 stacks of 2D textures, each aligned to one of the axes (x, y and z). These slices are called axis aligned
as they are aligned to one of the coordinate axes. The blend operation simulates bilinear interpolation and since it
is completely dependent on graphics hardware, it is considerably faster than raycasting but requires more memory
(3 copies of same data in 3 textures).
277
APPENDIX B. INTRODUCTION TO COMPUTER GRAPHICS - PART II Draft Sep. 2009
Figure B.3: This image shows a schematic of the raycasting process. A ray traverses the volume and accumulates the
color and opacity along the way.
Figure B.4: This image shows a schematic of texture mapping based volume rendering process. The polygon slices, the
texture mapped polygons and the blended volume rendered image are seen.
3D Texture mapping based volume rendering is a technique that uses more advanced graphics hardware
capabilities which allow for 3D textures. Now the data is loaded into a single 3D texture and slices are drawn
parallel to the viewing direction. This simulates trilinear interpolation and is much faster since it happens on
graphics hardware. The speed and quality of the visualization are directly dependent on the number of slices
drawn perpendicular to the viewing direction (sampling the data). Figure B.5 shows a schematic of 3D texture
mapping based volume rendering. The slices are drawn perpendicular to the view vector as can be seen in the left
image. The slices are textured using a single 3D texture and therefore eliminate the need to have three copies of
the same data in graphics memory.
B.4 Colormaps
Colormaps are a way to apply color and opacity to data. Images contain numbers that represent a quantity such
as intensity, pressure and so on. To display this quantity and distinguish between dierent values we need to
assign colors to dierent values of that quantity. Colormaps are used to assign color and opacity to values in
2D/3D data during the process of visualizing the data. Basically, colormaps are lookup tables where the intensity
(or any other quantity in the data) is looked up to obtain color and opacity.
The simplest colormap is a grayscale colormap in which the lowest intensity is mapped to black and the highest
intensity is mapped to white and all the intermediate intensities are shades of gray.
Additionally, the lower threshold and upper threshold for a colormap can be specied. It ensures that all the
values below the lower threshold are set to zero (black) while all the values above the higher threshold are set to
one (white).
Some standard colormaps that are commonly used are rainbow, reverse rainbow and heat/temperature color map.
278
APPENDIX B. INTRODUCTION TO COMPUTER GRAPHICS - PART II Draft Sep. 2009
Figure B.5: This image shows a schematic of the 3D texture mapping based volume rendering process. The view-aligned
polygon slices, the 3D texture mapped polygons and the blended volume rendered image are seen.
Figure B.6: This image shows the standard colormaps (top to bottom): Greyscale, Rainbow, Reverse Rainbow, Temper-
ature colormap.
In the rainbow colormap, the rainbow is used to map the quantity to color. The reverse rainbow colormap only
reverses the lookup table. The heat/temperature colormap is widely used since it is perceptually constant like
the grayscale colormap. Even though the rainbow colormap is widely used and is the default setting in a lot of
software, it has been shown that it is a bad choice and leads to inaccurate readings by viewers using the colormap
to make decisions based on the visualizations. Figure B.6 shows all the four colormaps discussed above.
One can also dened custom colormaps by varying the lookup table based on ones needs. This invariably leads
to more interesting results but also requires a lot of tedious manipulation by hand.
B.5 Illumination
Illumination can be dened as the transport of energy from light sources to surfaces and points in the scene. It
includes direct and indirect illumination.
Lighting is dened as the process of computing the luminous intensity (i.e., outgoing light) at a particular 3-D
point, usually on a surface. Shading is dened as the process of assigning colors to pixels.
B.5.1 Phong lighting model
The Phong lighting model is the most commonly used empirical lighting model in computer graphics. It simulates
realistic lighting by using a simple combination of three components: ambient, diuse and specular lighting.
Ambient lighting is an approximation for indirect illumination. Diuse lighting is directional lighting and depends
on the position of the light source. Specular lighting is what gives the surface a shiny, polished look. A light
279
APPENDIX B. INTRODUCTION TO COMPUTER GRAPHICS - PART II Draft Sep. 2009
Figure B.7: This image shows a volume rendering obtained using a custom colormap. The custom colormap too has
been shown here. The right image is another volume rendering of the with a dierent custom color map.
Figure B.8: This image shows the three components: ambient, diuse, specular lighting and the white specular highlight
too can be seen on the sphere.
shining on a specular surface causes a bright spot known as a specular highlight. Where these highlights appear
is a function of the viewers position, so specular reectance is view-dependent. Figure B.8 shows the three
components of the phong lighting model.
I
total
= k
a
I
ambient
+
#lights

i=1
I
i
_
k
d
(n L
i
) +k
s
(v R
i
)
n
shiny
_
where I
total
is the total color of each pixel, I
ambient
is the ambient light intensity, k
a
, k
d
, k
s
represent the ambient,
diuse, specular coecients that control the contribution of the ambient, diuse, specular components. n is the
normal vector, L
i
is the light vector, v is the view vector, R
i
is the reection vector that is computed using Snells
Law. n
shiny
is the specular exponent that determines the size of the specular highlight.
B.5.2 Shading
Shading is the process of applying the computed color in the lighting model to every point on the surface. Flat
shading uses the normal for each face in the polygonal model. The lighting model is computed once for a point
on the face and same computed color is applied to all the points on that face. Gouraud Shading uses vertex
normals and for points in between two vertices, it interpolates the result of the colors computed at each vertex to
give a smoother appearance. Phong shading uses vertex normals and actually computes the vertex normals for
each vertex on each face and evaluates the lighting model on each point on the surface. This produces the best
quality but is computationally expensive as compared to at and Gouraud shading. Figure B.9 shows the type
of normals used for each type of shading and the bottom row shows the resulting images obtained by using the
three dierent types of shading.
280
APPENDIX B. INTRODUCTION TO COMPUTER GRAPHICS - PART II Draft Sep. 2009
Figure B.9: Flat, Gouraud and Phong Shading. The top row shows the normals used in each type of shading. The
bottom row shows a sphere shaded using the three dierent shading models.
281
Draft Sep. 2009
Appendix C
Final Exam
Programming
1. Implement, using C++ and VTK, a reasonably complex medical image analysis algorithm of your choice.
This must be an algorithm that you have not previously implemented in C++ converting code from
MATLAB is allowed. The implementation must use VTK data structures for images, surfaces etc. If you
need suggestions, I have placed several papers in a subdirectory called final/papers; any one of these
will be ne.
(a) The algorithm must be implemented in C++.
(b) All code must be in classes deriving from VTK classes (e.g. vtkProcessObject)
2. Compile the algorithm into a shared library that is accessible from (i.e. loadable into) Tcl.
3. Implement using Tcl a command-line script for quickly testing the algorithm. This should load a synthetic
image of your choice, execute the algorithm and generate some output result.
4. Implement using Tcl/[Incr] Tcl a complete application for interacting with your algorithm. This must
include:
(a) A graphical user interface for parameter setting and executing the algorithm.
(b) An integrated interactive viewer for displaying the results.
(c) You may leverage BioImage Suite components if you wish to do so.
Report
The report should have three parts:
1. A brief (1-2 page) description of the selected algorithm.
2. A description of the implementation strategy (e.g. how the algorithm was broken up into dierent classes
etc.), and description of the functions in your code. (This is the heart of the report 4-6 pages, longer if
needed).
3. A brief (2 pages) Users Guide for your application. Including snapshots of key GUI elements etc.
The lecture notes for Sessions 20 and Session 21 are typical examples of what is expected for Parts 1 and 2.
282
Draft Sep. 2009
Appendix D
Code License
All example code from this book is made available under the following BSD-style open source license.
Copyright (c) 2006 Xenophon Papademetris, All rights reserved.
Redistribution and use in source and binary forms, with or without
modification, are permitted provided that the following conditions are met:
* Redistributions of source code must retain the above copyright
notice, this list of conditions and the following disclaimer.
* Redistributions in binary form must reproduce the above copyright
notice, this list of conditions and the following disclaimer in the
documentation and/or other materials provided with the distribution.
* The name of the Insight Software Consortium, nor the names of any
consortium members, nor of any contributors, may be used to endorse
or promote products derived from this software without specific prior
written permission.
* Modified source versions must be plainly marked as such, and must not
be misrepresented as being the original software.
THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDER AND CONTRIBUTORS AS
ISAND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO,
THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHORS OR CONTRIBUTORS BE
LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
POSSIBILITY OF SUCH DAMAGE.
283