Instant Ebook Access, One Click Away – Begin at ebookgate.

com

Insight into Images Principles and Practice for

Segmentation Registration and Image Analysis 1st
Edition Terry S. Yoo

https://ebookgate.com/product/insight-into-images-
principles-and-practice-for-segmentation-registration-and-
image-analysis-1st-edition-terry-s-yoo/

OR CLICK BUTTON

DOWLOAD EBOOK

Get Instant Ebook Downloads – Browse at https://ebookgate.com

Click here to visit ebookgate.com and download ebook now
Instant digital products (PDF, ePub, MOBI) available
Download now and explore formats that suit you...

Image Registration for Remote Sensing 1st Edition

Jacqueline Le Moigne

https://ebookgate.com/product/image-registration-for-remote-
sensing-1st-edition-jacqueline-le-moigne/

ebookgate.com

Microarray Image and Data Analysis Theory and Practice 1st

Edition Luis Rueda

https://ebookgate.com/product/microarray-image-and-data-analysis-
theory-and-practice-1st-edition-luis-rueda/

ebookgate.com

Coaching Science Theory into Practice Wiley Sport Texts

1st Edition Terry Mcmorris

https://ebookgate.com/product/coaching-science-theory-into-practice-
wiley-sport-texts-1st-edition-terry-mcmorris/

ebookgate.com

APA handbook of behavior analysis Vol 2 Translating

principles into practice Gregory J. Madden (Editor)

https://ebookgate.com/product/apa-handbook-of-behavior-analysis-
vol-2-translating-principles-into-practice-gregory-j-madden-editor/

ebookgate.com
Insight into IELTS Updated Edition Vanessa Jakeman

https://ebookgate.com/product/insight-into-ielts-updated-edition-
vanessa-jakeman/

ebookgate.com

Deep Learning for Medical Image Analysis 1st Edition S.

Kevin Zhou

https://ebookgate.com/product/deep-learning-for-medical-image-
analysis-1st-edition-s-kevin-zhou/

ebookgate.com

Middleton s Allergy Principles and Practice

https://ebookgate.com/product/middleton-s-allergy-principles-and-
practice/

ebookgate.com

An Insight into Chemical Engineering 1st Edition M. Subbu

https://ebookgate.com/product/an-insight-into-chemical-
engineering-1st-edition-m-subbu/

ebookgate.com

Image Processing Principles and Applications 1st Edition

Tinku Acharya

https://ebookgate.com/product/image-processing-principles-and-
applications-1st-edition-tinku-acharya/

ebookgate.com
Insight Into Images
Principles and Practice for
Segmentation, Registration,
and Image Analysis

Terry S. Yoo, Editor

National Library of Medicine, NIH
Office of High Performance Computing
and Communications

A K Peters
Wellesley, Massachusetts
Editorial, Sales, and Customer Service Office

AK Peters, Ltd.

Weliesey, MA 02482
888 Worcester Street, Suite 230

www.akpeters.com

Copyright © 2004 by A K Peters, Ltd.

Ali rights reserved. No part of the material protected by this copyright notice may
be reproduced or utilized in any form, electronic or mechanical, including photo­
copying, recording, or by any information storage and retrieval system, without
written permission from the copyright owner.

Library of Congress Cataloging-in-Publication Data

and image analysis / Terry S. Yoo, editor.

Insight into images: principles and practice for segmentation, registration,

p. cm.
I ncludes bibliographical references and index.
ISBN 1 -56881 -217-5
1. Diagnostic imaging-Digital techniques. 2. I mage processing.
I. Yoo, Terry S., 1963-

RC78.7. D531567 2004

616.07'54-dc22
2003068974

Printed in Canada
08 07 06 05 04 10 9 8 7 6 5 4 3 2 1
To my parents, T.S.Y.
To my wife Geanne, Baby Ada and my parents, D.N.M.
To Konrad, Marcin and Tomasz C.1.
To Stephen P izer, G.D.S.
To the ITK team at the lab, J.C.G.
To my family and parents, JK
To Geoff, L.N.
For Dan and Marlene, R.T.W
To my beloved wife and my parents, Yinpeng Jin
To Daniel and Camilo, L.1.
To the friendships formed while writing ITK, S.R.A.
To my grandmother, PKS.
To Rebecca and her big heart, B.B.A.
To lab friends old and new, T.A.S.
To my daughter, Y.Z.
To my parents and friends in the ITK group, T.C.
The purpose of computing is Insight, not numbers.
-Richard Hamming
 Contents

Foreword xi

Introduction and Basics 1

1 I ntroduction 3
1.1 Medical Image Processing . . . . . . . . . . . 4
1 .2 A Brief Retrospective on 3D Medical Imaging 5
1 .3 Medical Imaging Technology . . . . . . 6
1 .4 Acquisition, Analysis, and Visualization . 16
1 .5 Summary . . . . . . . . . . . . . . . . . 17

2 Basic I mage Processi ng and Linear Operators 19

2.1 Introduction . . 19
2.2 Images . . . . . 20
2.3 Point Operators . 21
2.4 Linear Filtering . 25
2.5 The Fourier Transform . 37
2.6 Summary . . . . . . . . 45

3 Statistics of Pattern Recog n ition 47

3.1 Introduction . . . . . . . . . . 47
3.2 Background . . . . . . . . . . . 49
3.3 Quantitative Comparison of Classifiers 56
3.4 Classification Systems . . . . . . . . 59
3.5 Summary of Classifiers' Performance 84
3.6 Goodness-of-Fit . . . . . . . . . . . 87

vii
viii Contents

3.7 Conclusion . . . . . . . . . . . . . . . . . . 92
3.8 Appendix: Extruded Gaussian Distributions . 93

4 Nonli near I mage Fi ltering with Partial Differential Equations 1 03

4. 1 Introduction . . . . . . . . . . . . . .
. . 1 03
4.2 Gaussian Blurring and the Heat Equation · 1 03
4.3 Numerical Implementations . . . . . . . · 1 10

II Segmentation 119

5 Segmentation Basics 1 21
5. 1 Introduction . . . . · 121
5.2 Statistical Pattern Recognition . · 1 23
5.3 Region Growing . . . . . . . . · 1 24
5.4 Active SurfaceslFront Evolution . 1 26
5.5 Combining Segmentation Techniques 1 27
5.6 Looking Ahead . . . . . . . . . . . . 1 28

6 Fuzzy Connected ness 1 31

6. 1 Background . . . . . . . . . . . · 131
6.2 Outline of the Chapter . . . . . · 133
6.3 Basic Notations and Definitions 136
6.4 Theory . . . . . . . . . . 138
6.5 Methods and Algorithms . 151
6.6 Applications . . . . . 161
6.7 Concluding Remarks . . . 171

7 Markov Random Field M odels 1 81

7.1 Markov Random Field Models: Introduction and Previous Work . 181
7.2 Gibbs Prior Model Theories . . . . . . . . . . . . . . 1 82
7.3 Bayesian Framework and Posterior Energy Function . 185
7.4 Energy Minimization . . . 186
7.5 Experiments and Results . . . . . . . . . . . . . . . . 1 86

8 Isosu rfaces and Level Sets 1 93

8.1 Introduction . . . . . · 193
8.2 Deformable Surfaces . · 195
8.3 Numerical Methods · 1 99
8.4 Applications . 208
8.5 Summary . . . . . . · 214
Contents ix

9 Deformable Models 219

9.1 Introduction . . . 219
9.2 Previous Work . . . . . . . . 220
9.3 Deformable Model Theories . 222
90 4 Experiments, Results, and Applications in ITK . 230

III Registration 237

1 0 Medical I mage Registratio n : Co ncepts and I m plementation 239

1 0. 1 Introduction . . . . . . . . . . . .
. . . . . . . . . . . .
. . . . 239
1 0.2 Image Registration Concepts . . . . . . . . . . . . . . . . . . . . 240
1 0.3 A Generic Software Framework for Image to Image Registration . 242
l OA Examples . . . . . . . . . . . . . . . .
. . . . . . . . . . .
. . . 295

1 1 Non-Rigid I mage Registration 307

1 1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . .
. . . 307
1 1 .2 Optical Flow: Fast Mono-Modality Non-Rigid Registration . 3 12
1 1 3.
. Variational Framework for Computational Anatomy . . 321
1 10 4 Review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 338

IV Hybrid Methods - Mixed Approaches to Segmentation 349

1 2 Hybrid Segmentatio n Methods 35 1

1 2. 1 Introduction . . . . . . .
. . . 35 1
1 2.2 Review of Segmentation Methods . 353
1 2.3 Hybrid Segmentation Engine . . . 357
120 4 Hybrid Segmentation: Integration of FC, VD, and DM . 358
1 2.5 Evaluation of Segmentation . 365
1 2.6 Results . . . . 367
1 2.7 Conclusions . . . . . . .
. . 375

389
I ndex
 Foreword

This book is the product of a extraordinary collaboration among commercial

groups, researchers from computer science departments and engineering schools,
and researchers and scientists from medical schools and radiology departments.
The ultimate result has been the Insight Toolkit (or ITK). ITK is an open source
library of software components for data segmentation and registration, provided
through federal funding administered by the National Library of Medicine, NIH,
and available in the public domain. ITK is designed to complement visualization
and user interface systems to provide advanced algorithms for filtering, segmenta­
tion, and registration of volumetric data. Created originally to provide preprocess­
ing and analysis tools for medical data, ITK is being used for image processing
in a wide range of applications from handwriting recognition to robotic computer
vision.

History and P h i l os ophy of IlK

In August of 1 99 1 , the National Library of Medicine (NLM) began the acquisition

of the Visible Human Project Male and Female (VHP) datasets. The VHP male
dataset contains 1 ,97 1 digital axial anatomical images ( 1 5 Gbytes), and the VHP
female dataset contains 5, 1 89 digital images (39 Gbytes). However, researchers
have complained that they are drowning in data, due in part to the sheer size of the
image information and the available anatomical detail. Also, the imaging com­
munity continues to request data with finer resolution, a pressure that will only
compound the existing problems posed by large data on analysis and visualiza­
tion. The NLM and its partner institutes and agencies seek a broadly accepted,
lasting response to the issues raised by the segmentation and registration of large
3D medical data.
The Insight Software Research Consortium was formed in 1 999 with the mis­
sion to create public software tools in 3D segmentation and deformable and rigid

xi
xii Foreword

registration, capable of analyzing the head-and-neck anatomy of the Visible Hu­

man Project data. The result has been the Insight ToolKit (ITK), a publicly avail­
able software package for high-dimensional (3D, 4D, and higher) data processing
supported by the NLM and its partner institutes and federal agencies. At the time
of this writing, the software is soon to be released as ITK lA, its third release
incorporating software contributions from new contributors including Harvard
Medical School, the Georgetown ISIS Center, and the Mayo Clinic. The most
aggressive hope for this project has been the creation of a self-sustaining open­
source software development community in medical image processing based on
this foundation. By creating an open repository built from common principles, we
hope to help focus and concentrate research in this field, reduce redundancy of de­
velopment, and promote cross-institution communication. The long-term impact
of this initiative may well reach beyond the initial mission of archiving computer
algorithms. The team of contractors and subcontractors includes:

• GE Global Research
• Harvard BWH Surgical Planning Lab
• Kitware, Inc.
• Insightful, Inc.

• The University of Pennsylvania

• The University of North Carolina at Chapel Hill
• The University of Tennessee at Knoxville
• Columbia University
• The University of Pittsburgh
• The University of Utah
• Georgetown University
• The University of Iowa

• The Imperial College/Kings College London

• Cognitica, Inc.
• The Mayo Clinic

• Carnegie Mellon University Robotics Institute

• The University of Utah SCI Institute
Foreword xiii

ITK has met its primary goals for the creation of an archival vehicle for im­
age processing algorithms and an established functioning platform to accelerate
new research efforts. New projects have connected ITK to existing software ap­
plications for medical and scientific supercomputing visualization such as "An­
alyze" from the Mayo Clinic and "SCIRun" from the University of Utah's SCI
Institute. These projects were sponsored to demonstrate that the application pro­
grammer's interface (API) for ITK was suitable for supporting large applications
in established medical and scientific research areas. Also, recent projects have
been funded to add algorithms not previously funded to the collection, proving
the versatility of the underlying software architecture to accommodate emerging
ideas in medical imaging research.
There is growing evidence that ITK is beginning to influence the international
research community. In addition to the sponsored growth of the ITK community,
the mailing list of users and developers includes over 300 members in over 30
countries. These facts may be indicators that a nascent research community may
be forming about the core ITK software. The applications for ITK have begun
to diversify including work in liver RF ablation surgical planning, vascular seg­
mentation, live-donor hepatic transplant analysis, intra-operative 3D registration
for image-guided intervention, rapid analysis of neurological function in veter­
ans, longitudinal studies of Alzheimer's patients using MR!, and even handwrit­
ing analysis. The Insight Program of the Visible Human Project is emerging as a
clear success and a significant and permanent contribution to the field of medical
imaging.
In 1 999, the NLM Office of High Performance Computing and Communi­
cations, supported by an alliance of NIH ICs and Federal Funding Agencies,
awarded six contracts for the formation of a software development consortium to
create and develop an application programmer interface and first implementation
of a segmentation and registration toolkit, subsequently named the Insight Toolkit
(ITK). The final deliverable product of this group has been a functional collection
of software components, compatible for direct insertion into the public domain
via Internet access through the NLM or its licensed distributors. Ultimately, NLM
hopes to sponsor the creation of a public segmentation and registration software
toolkit as a foundation for future medical image understanding research. The in­
tent is to amplify the investment being made through the Visible Human Project
and future programs for medical image analysis by reducing the reinvention of
basic algorithms. We are also hoping to empower young researchers and small
research groups with the kernel of an image analysis system in the public domain.

Scope of This Book

While a complete treatise on medical image processing would be an invaluable

contribution to the research community, it would be far too much hubris on our
xiv Foreword

PaIt to attempt to describe it all. The concept of Finite Element Models (FEM)
is a book-length treatise in itself, as are each of the topics of Level Set Methods
and Statistical Pattern Recognition. Indeed, we reference many valuable texts on
these topics in particular.
We therefore limit this material to methods that have been implemented in the
Insight ToolKit. All methods described herein are reflected in the source code
for ITK. We limit the perspectives in this text to the viewpoints presented in ITK,
and though there are many valid and valuable alternate approaches incorporating
many of these same ideas, we cannot treat them all, and in fairness, we claim
expertise only in those areas that we can deomnstrably express directly through
software development.
In keeping with the principles of the Insight project itself, we concentrate
only on medical image filtering, segmentation, and registration. These were the
primary areas of concentration of the software project, and thus they are the topics
described in this book. As such, we do not directly treat the ares of visualization,
identification, classification, or computer-aided detection and diagnosis. All of
the methods presented in this book are relevant to these topics, and we cannot
show our results without some form of visualization, but we do not include an
in-depth treatise of these topics or explicitly describe the integration of ITK with
visualization environments.

Suggested Uses for T h i s Book

Throughout the development of this text, we have internally described it as the

"theory book." We have avoided including source code in an attempt to focus on
the principles of the methods employed in ITK. There is a companion text for ITK
directed toward implementation and integration details, called the ITK Software
Guide. These books should be considered complementary to each other and sup­
plementary in support ofITK. If you, the reader, are familiar with the principles of
some method or approach in ITK, you only need to read the ITK Software Guide
to learn about our particular implementation of the method and our structure and
intended uses for that method. If, however, you require greater background on the
theory behind some method, you should refer to the material in this book. Thus,
we perceive this book to be an extended reference for ITK, providing more in­
depth presentation of the basic algorithms, structure, and methods in the toolkit
than can be gleaned either from the source code or from the base level descriptions
in the ITK Software Guide.
We also perceive this book to be the basis for reading and curriculum foun­
dations for a graduate course on medical image processing. ITK has been effec­
tively used to create courses on this topic. Its advanced implementations permit
the rapid explorations of complext algorithms without requiring the student or the
Acknowledgements xv

instructors and their staff to create the software infrastructure necessary to run
high-dimensional image processing software. For such a course, this book would
represent theoretical course material, and the ITK Software Guide would be an
invaluable resource for laboratory exercises and course projects.
We are extraordinarily proud of ITK and hope that you find our offering in­
structive, valuable, and a useful supporting addition to your software projects. We
humbly and gratefully present this material with hope for your future success in
research and medical software application development.

Acknowledgements

Following the consortium-based nature of this project, this book is the collec­
tive effort of many, many people. The contributing authors include in alphabet­
ical order: Brian Avants, Stephen Aylward, Celina Imielinska, Jisung Kim, Bill
Lorensen, Dimitris Metaxas, Lydia Ng, Punam K Saha, George Stetten, Tessa
Sundaram, Jay Udupa, Ross Whitaker, and Terry Yoo. Other people have made
essential contributions, including: Luis Ibanez, Will Schroeder, and others. The
Insight ToolKit would not have been possible without the selfless efforts of Bill
Hoffman, Brad King, Josh Cates, and Peter Ratiu.
We are indebted to the coalition of federal funding partners who have made
this project possible: the National Library of Medicine, the National Institute
for Dental and Craniofacial Research, the National Institute for Mental Health,
the National Eye Institute, the National Institute on Neurological Disorders and
Stroke, the National Science Foundation, the DoD Telemedicine and Advanced
Technologies Research Center, the National Institute on Deafness and other Com­
munication Disorders, and the National Cancer Institute. We would especially
like to recognize the patronage of the Visible Human Project from the Office of
High Performance Computing and Communications of the Lister Hill National
Center for Biomedical Communications of the National Library of Medicine. In
particular, we'd like to thank Drs. Donald Lindberg and Michael Ackerman for
their vision and leadership in sponsoring this project.
 Part One
Introduction and Basics
 CHAPTER

ONE

Introduction

Terry S. Yoo
National Library of Medicine, NIH

This book is an introduction to the theory of modem medical image process­

ing. Image processing is the manipulation and analysis of image-based digital
data to enhance and illuminate information within the data stream. One com­
mon approach to this topic developed by many texts is to treat the matter as a
high-dimensional form of signal processing, concenterating on filtering and fre­
quency analysis as the foundations with later chapters discussing applications and
advanced algorithms. This type of organization reflects a bias toward electrical
engineering, viewing images as similar to complicated audio or radio frequency
signals. Another approach is to start from the mathematical perspective, treat­
ing the issue as a series of exercises in statistics or applied mathematics through
equations and theorems. Our approach is to explore the topic through implemen­
tations of advanced algorithms, providing the mathematical, statistical, or signal
processing as needed for background.
This book takes a software engineer's viewpoint, and describes methods and
algorithms available in the Insight Toolkit as open source software. The goal of
this book is to explain the background specific to our implementations, enabling
the reader or student to use complex tools in medical image processing as rapidly
as possible. We do not claim an exhaustive treatment of these ideas, but rather
a working and functional introduction to these topics, partnered with working
implementations of the methods described.

3
4 Chapter 1 . Introduction

1 .1 Medical I mage P rocessi n g

The medical mission differs from the other forms o f image processing arising
from non-medical data. In satelite surveillance analysis, the purpose is largely a
screening and cartographic task, aligning multiple types of data and correspond­
ing them to a known map and highlighting possible points of interest. In computer
vision, camera views must be analyzed accounting for the perspective geometry
and photogrammetric distortions associated with the optical systems that are the
basis for robotic sensors. In many of these systems, autonomous navigation, tar­
get identification and acquisition, and threat avoidance are the primary tasks. For
the most part, the incoming information arrives as 2D images, data arrays that
can be organized using two cartesian dimensions. In addition, the tasks are to
be performed independently by the machine, relying on the development of ma­
chine learning and artificial intelligence algorithms to automatically accomplish
the tasks.
In medicine, the problem as well as the input data stream are usually three­
dimensional, and the effort to solve the primary tasks is often a partnership of
human and machine. Medicine is notably a human enterprise, and computers are
merely assistants, not surrogates nor possible replacements for the human expert.
The medical task can often be split into three areas; ( 1 ) data operations of filtering,
noise removal, and contrast and feature enhancement, (2) detection of medical
conditions or events, and (3) quantitative analysis of the lesion or detected event.
Of these subtasks, detection of lesions or other pathologies is often a subjective
and qualitative decision, a type of process ill-suited for execution by a computer.
By contrast, the computer is vastly more capable of both quantitative measurent
of the medical condition (such as tumor volume or the length of a bone fracture)
and the preprocessing tasks of filtering, sharpening, and focusing image detail.
The natural partnership of humans and machines in medicine is to provide the
clinician with powerful tools for image analysis and measurement, while relying
on the magnificent capabilities of the human visual system to detect and screen
for the primary findings.
We divide the problems inherent in medical image processing into three basic
categories;
• Filtering: These are the basic tasks involved in filtering and preprocessing
the data before detection and analysis are performed either by the machine
or the human operator.
• Segmentation: This is the task of partitioning an image (2D array or vol­
ume) into contiguous regions with cohesive properties.
• Registration: This is the task of aligning multiple data streams or images,
permitting the fusion of different information creating a more powerful di­
agnostic tool than any single image alone.
1 .2. A Brief Retrospective on 3D Medical Imaging 5

These three basic divisions represent the organization of both this book and the
Insight Toolkit for which this book was written. In order to best understand the
approaches and difficulties associated with these tasks, we being with a history of
medical imaging and a brief overview of modem imaging modalities.

1 .2 A B rief Retrospective on 3D Med ical I m ag i ng

In the broad realm of the sciences, medical imaging is a young field. The physics
enabling diagnostic imaging are barely one hundred years old. In 1 895, Wilhelm
Roentgen discovered x-rays while experimenting with a Crookes tube, the pre­
cursor to the cathode ray tube common in video applications today. It is worth
noting that he immediately recognized the potential of x-ray radiation in diagnos­
tic imaging and that one of his earliest images is of the bones of his wife's hand.
Roentgen's discovery was so noteworthy and revolutionary that he received a No­
bel prize within six years of announcing his initial work. Thus began a bountiful
and tightly-coupled one hundred year partnership between physics and medicine.
The meaning and purpose of medical imaging has been to provide clinicians
with the ability to see inside the body, to diagnose the human condition. The
primary focus for much of this development has been to improve the quality of the
images for humans to evaluate. Only recently has computer technology become
sufficiently sophisticated to assist in the process of diagnosis. There is a natural
partnership between computer science and radiology.
Early in the twentieth century, a Czech mathematician named Johann Radon
derived a transform for reconstructing cross-sectional information from a series
of planar projections taken from around an object. While this powerful theory had
been known for over fifty years, the ability to compute the transform on real data
was not possible until digital computers began to mature in the 1 970s.
Imaging in 3D emerged in 1972 when x-ray computed tomography (CT) was
developed independently by Godfrey Hounsfield and Alan Cormack. These inno­
vators later shared the 1 979 Nobel Prize in Medicine. Their achievement is note­
worthy because it is largely based on engineering, the theoretical mathematics and
the underlying science had been described decades earlier. The contribution is in
the application of mechanical engineering and computer science to complete what
had previously only been conceived on paper. Clinical systems were patented in
1 975 and began service immediately thereafter.
While techniques for using x-rays in medical imaging were being refined, or­
ganic chemists had been exploring the uses of nuclear magnetic resonance (NMR)
to analyze chemical samples. Felix Bloch and Edward Purcell were studying
NMR in the mid 1940s. Together, they shared the Nobel Prize in Physics in
1 952. Paul Lauterbur, Peter Mansfield, and Raymond Damadian were the first
to develop imaging applications from NMR phenomena. Lauterbur created tomo-
6 Chapter 1 . Introduction

graphic images of a physical phantom constructed of capillary tubes and water

in a modified spectrometer in 1 972. Damadian later was able to create animal
images in 1 975. The means of creating medical images using magnetic fields and
radio waves was later renamed Magnetic Resonance Imaging (MRI). Lauterbur
and Mansfield received the 2003 Nobel Prize in Medicine. Raymond Damadian
has not been so honored, but his contribution is significant and should not be
overlooked by us.
Radiologists were thus simultaneously presented with the possibilities for slice
(or tomographic) images in the axial plane from x-ray CT and from MRI in the
early 1 970s. Arising from the long development of x-ray technology, CT matured
first. The difficulties and expense of generating strong magnetic fields remained
an obstacle for clinical MRI scanners until engineers were able to produce prac­
tical superconducting magnets. While still expensive, MRI technology is now
available worldwide, creating precise images of deep structures within the body,
outlining the anatomy and physiology of internal objects other than bones. MRI
and CT have joined the established 3D imaging modalities of nuclear imaging and
the growing field of volumetric medical ultrasound. The creation of new modali­
ties with differing strengths has led to the need to align or register these multiple
data streams.

1 .3 Medical I mag ing Techno logy

The majority of medical visualization involves the display of data acquired di­
rectly from a patient. The radiologist is trained through long years of education
and practice to read relatively simple presentations of the raw data. The key to
improving a diagnosis is in the careful crafting of the acquisition, applying the
physics of radiology to maximize the contrast among the relevant tissues and sup­
pressing noise, fog, and scatter that may obscure the objects of interest. This is no
less true for more complex visualizations that incorporate 3D renderings or vol­
ume projections; improving the quality of the acquired data will fundamentally
affect the quality of the resulting visualization.
This section will briefly cover some of the more common sources of three or
higher-dimensional medical imaging data. This treatment is necessarily super­
ficial, serving mostly as an annotated glossary for many of the terms taken for
granted in the radiology community. To create effective medical visualization
tools, the computer scientist requires a fundamental understanding of the source
of the image data, the technology involved, and the physical principles from which
the image values are derived. This cursory introduction will be insufficient for in­
depth research, but will serve as background for this text; the reader is encouraged
to continue the exploration of this topic. This additional command of the basics
of medical image acquisition will enhance your practice in medical visualization,
1.3. Medical Imaging Technology 7

Figure 1 . 1 . The business end of a CT scanner (circa 1998).

ease communication between you and the doctor and the technologist, improve
the quality of the source data, and smooth transitions among the many interfaces
from acquisition to display.

1.3.1 Computed Tomography

By far the most familiar form of 3D medical imaging is X-ray Computed Tomog­
raphy or cr (formerly referred to as Computer Assisted Tomography. Comput­
erized Axial Tomography, CAT scanning, Computerized Transaxial Tomography,
CfAT. Computerized Reconstruction Tomography, CRT, and Digital Axial To­
mography, DAT). The mathematics for tomographic reconstruction from multiple
views have been known for most of this century. 11 took almost fifty years be­
fore the components required for x-ray computed tomography were sufficiently
developed to make the procedure and the ensemble of instruments economically
feasible.
A cr scanner is a room sized x-ray instrument, requiring a shielded environ­
ment to protect the technologists and other clinic staff from exposure from routine
use. T h e number of manufacturers of cr scanners has been steadily decreasing
with only a handful of vendors providing them today. The cost of these devices
ranges from approximately $400,000 to well over $1,000,000. There are some
examples of portable cr machines that can be moved into a trauma care center or
into operating rooms to aid in the diagnosis and treatment of patients. Emerging
technologies in fiat-panel digital x-ray devices is enabling 3D imaging using more
conventional fluoroscopic tools, but at the time of this writing, the following is the
most common configuration for these devices.
8 Chapter 1. Introduction

An accurately calibrated moving bed to translate the patient through the scan­
ner, an x-ray tube mounted in such a way to allow it to revolve about the patient,
and an array of x-ray detectors (gas filled detectors or crystal scintillation detec­
tors) comprise the essential system components of a CT machine. The x-ray tube
and detector array are mounted in a gantry that positions the detector assembly
directly across from the x-ray source. The x-ray source is collimated by a pair of
lead jaws so that the x-rays form a flat fan beam with a thickness determined by
the operator. During the acquisition of a "slice" of data, the source-detector ring
is rotated around the patient. The raw output from the detector array is backpro­
jected to reconstruct a cross-sectional transaxial image of the patient. By reposi­
tioning the patient, a series of slices can be aggregated into a 3D representation of
the patient's anatomy. Figure 1 . 1 is a picture of a CT table and gantry.
Within the last fifteen years there have been significant advances in CT tech­
nology, allowing for faster spiral acquisition and reduced dose to the patient as
well as multi slice detector arrays permitting simultaneous acquisition of several
slices at a time. Early CT gantries were constructed with a revolving detector
array positioned directly across the patient from the moving x-ray source. Cable
lengths connecting the moving detector assembly and the x-ray tube permitted
only a single slice to be acquired at one time. The revolving assembly then had
to be "unwound", the patient advanced the distance of one slice, and the process
repeated.
As small affordable detectors have become available, scanners have been de­
signed with a fixed array of x-ray detectors. The only remaining revolving part
of the gantry is the x-ray tube. This has simplified the engineering and cabling of
the gantry. The x-ray tube can now be cabled using a slip ring, permitting con­
tinuous revolution of the tube about the patient. This type of design is capable of
helical (or spiral) CT acquisition. By simultaneously revolving the x-ray source
about the patient and continuously moving the patient through the bore of the
gantry, the data are acquired via a spiral path. These methods have enabled very
fast image acquisition, improving the patient throughput in a CT facility, reducing
artifacts from patient motion, and reducing absorbed dose by the patient. Com­
bined with multiple layers of sensors in the detection ring, these new scanners are
generating datasets of increasingly larger sizes.
A CT scanner is an x-ray modality and is subject to all of the physics as­
sociated with the generation, dispersion, absorption, and scatter associated with
all x-ray photons. The process of generating a CT scan is similar to creating a
standard x-ray film; however, while a single x-ray exposure generates a complete
film-based exam, the CT image is not acquired in a complete form. Rather, it must
be reconstructed from multiple views. The advantage of being an x-ray modality
is that laymen and clinicians alike have considerable intuition when dealing with
x-ray-based images. The concepts of dense objects like bone absorbing more
photons relative to less dense tissues like muscle or fat come naturally from our
1 .3. Medical Imaging Technology 9

experience and expectations about x-ray imaging. A typical CT scanner can gen­
erally acquire the data for a transaxial slice in a matter of seconds (within 1 to
5 seconds). An exam can include several series of slices, in some cases with
and without pharmaceutical contrast agents injected into the patient to aid in di­
agnostic reading. Slices can be spaced such that they are either overlapping or
contiguous, though some protocols call for gaps between the slices. A large study
can include well over 1 00 separate 5 1 2 x 5 1 2 pixel images. The radiation dose
from a CT scan is comparable with that of a series of traditional x-rays.
The concept of "resolution" should be divided in the researcher's mind into
spatial resolution (i.e., how much area in each dimension a voxel covers) and
sampling resolution (how many voxels in each dimension of the slice). Sampling
resolution in current scanners usually creates images that are either 256 x 256
or 512 x 5 1 2 voxels square. Sampling resolution in the longitudinal direction is
limited only by the number of slices acquired. Spatial resolution in the longitu­
dinal direction is bound by physical limitations of collimating the photons into
thin planes. The physical lower limit is approximately I mm; narrower collima­
tion requires inordinate amounts of x-ray flux to image and also leads to diffrac­
tion interference. Sampling resolution in the transaxial dimensions of a voxel is
based on the field of view selected by the operator of the CT scanner and the ma­
trix (256 x 256 or 5 1 2 x 5 1 2) yielding pixel dimensions that are generally 0.5 to
2 mm. Attempting to achieve higher spatial resolution will lead to voxels with
too little signal to accurately measure x-ray absorption. With today's diagnos­
tic equipment, if pixels smaller than 0.25 mm are attempted, low signal-to-noise
ratios become a problem.
Units of measure for the pixel values of CT imaging are standard across the
industry. Each pixel ideally represents the absorption characteristics of the small
volume within its bounds. By convention, these measurements are normalized
relative to the x-ray absorption characteristics of water. This is a unit of measure
known as Hounsfield units (HU). The Hounsfield unit scale is calibrated upon
the attenuation coefficient for water, with water reading 0 HU. On this scale air is
-1000 HU, fat tissue will be in the range of -300 to -100 HU, muscle tissue 1 0-70
HU and bone above 200 HU.
Every CT image contains artifacts that should be understood and handled
when visualizing the data. The process of reconstruction and sampling leads to
aliasing artifacts, just as any sampling procedure in computer graphics or signal
processing. Another common artifact is partial voluming, the condition where
the contents of a pixel are distributed across multiple tissue types, blending the
absorption characteristics of different materials. Patient motion while scanning
generates a variety of blurring and ring artifacts during reconstruction.
There is another class of artifacts that arises from the fact that CT is an x-ray
modality. Embedded dense objects such as dental fixtures and fillings or bullets
lead to beam shadows and streak artifacts. More common is the partial filtering
10 Chapter 1 . Introduction

Figure 1.2. The business end of an MRI scanner (circa 1998),

of the x-ray beam by dense tissue such as bone which leads to beam hardening,
a condition which causes slight shadows (0 halo the dense features of the image.
When creating visualizations, beam hardening anifacts may cause the researcher
to underestimate the volume of dense objects, making them seem smaller than
they are.

1.3.2 Magnetic Resonance Imaging

Like cr, Magnetic Resonance Imaging or MRI (fonnerly referred to as Nuclear
Magnetic Resonance imaging) was pioneered in the early 1970s. Introduced un­
der the name "Zeugmatography" (from the Greek word zeugma, meaning "that
which draws together") it remained an experimental technology for many years.
Unlike cr, MRI does not use ionizing radiation to generate cross-sectional im­
ages. MRI is considered to be a newer modality since feasible commercial de­
velopment of diagnostic MRJ scanners had to await affordable super-conducting
magnets. Today's diagnostic systems ranges from $500,000 to $2 or $3 million.
An MRJ scanner is a large magnet, a microwave transmitter, a microwave an­
tenna. and several electronic components that decode the signal and reconstruct
cross-sectional images from the data. Generally, the magnet is superconducting
and must be operated at cryogenic temperatures (4 degrees Kelvin), necessitat­
ing its immersion in a bath of liquid helium. Some MRJ scanners with relatively
poorer resolution are being constructed using fix.ed magnets without the need for
a liquid helium container, pennitting more open designs at the cost of image con-
1 .3. Medical Imaging Technology 11

trast and resolution. Figure 1 .2 is a picture of an MRI magnet and table, showing
a head coil, an antenna for imaging the head and neck.
Unlike CT scanners, the bore of an MRI scanner is often up to two meters in
length (6-8 feet). Patients are inserted into the middle of the magnetic field, often
inducing claustrophobia. The environments for MRI scanners must be shielded
for magnetic and radiofrequency interference. Large Faraday cages and substan­
tial masses of iron usually surround the magnet and sometimes the entire room
about the scanner.
In MRI scanning, the patient is placed within a high intensity magnetic field.
Field strengths vary from 0.35 Tesla to 1 .5 Tesla for most diagnostic MRI devices
(for reference, 1 Tesla = 1 0,000 Gauss, and the earth's magnetic field, though
variable, is approximate 0.5 Gauss). The induced magnetic field causes the mag­
netic moments of the hydrogen atoms within the patient to align along the prin­
cipal direction of the superconducting magnet. Low-level radio waves in the mi­
crowave frequencies (approximately 1 5 to 60 MHz) are then transmitted through
the patient, causing the magnetic moments of the hydrogen nuclei to resonate and
re-emit microwaves after each pulse. The microwaves emitted by the body are
recorded using a radio frequency antenna, filtered, amplified, and reconstructed
into tomographic slices. While all of the hydrogen nuclei typically resonate at
a frequency fixed by the strength of the induced magnetic field, different tissue
types resonate longer than others, allowing the viewer to discriminate among them
based on the magnitude of the signal from different points in space over time.
Spatial locations can be determined by varying the magnetic field about the
patient in different directions at different times. Linear gradients in the induced
magnetic field are produced by magnets that supplement the main superconduct­
ing magnet. These gradient coils are activated in a variety of sequences, altering
the phase and frequency of the microwave pulses that are received and re-emitted
by the hydrogen nuclei. The design and crafting of pulse sequences is a field of
research unto itself. The careful selection of pulse sequences can illuminate a
variety of clinical conditions including the display of swelling and bleeding, even
enhancing the blood vessels deep within the brain. Often, several sequences will
be taken of the patient during an exam to capture the variety of information avail­
able in the different pulse sequences. The visualization researcher is encouraged
to learn about these pulse sequences and their properties in order to select the best
ensemble of them to use when crafting visualizations.
The output of an MRI scanner is similar to CT. Slices representing slabs of
the object scanned are produced. However, unlike CT which always produces
transaxial slices, the slices from MRI can be oriented in any plane. The output
values at each image element are not calibrated to any particular scale. Gen­
erally they are l O-bit data samples. The values will vary depending upon the
scan parameters, and the patient's size and magnetic characteristics. Addition­
ally, the values are not constant over the entire scan space since inhomogeneity
12 Chapter 1. Introduction

in the magnetic field causes pixels that may represent the same tissue, but located
some distance apart to give different signals. This lack of an absolute scale for a
dataset is a cause of much consternation to the researcher attempting to segment
MRI data.
Visualization researchers seldom have good intuition for the meaning of MRI
signals. Unlike x-ray-based imaging, there are no physical analogs to what the
viewer is seeing. An MRI device measures the radio signals emitted by drops of
water over time. Usually skin and fat are much brighter than bone which has vir­
tually no signal at all. Segmentation and classification are therefore significantly
harder and are the subject of much research.
As with CT, the concept of resolution must be divided into spatial resolu­
tion and sampling resolution. The matrix of the image can typically be selected
to be either 256 x 256 or 5 12 x 5 1 2, depending on the sampling resolution of the
frequency and phase decoders of the receiving equipment. Square pixels in a rect­
angular matrix are often selected since reducing the number of rows or columns
can significantly reduce the time required to complete a scan. Spatial resolution,
or field of view, is primarily dependent on the strength of the gradient magnets
and their ability to separate the slices along their gradient directions. The ra­
diofrequency receiving equipment must be able to distinguish among frequencies
and phases which have only slight variations. Stronger gradients create greater
separations and improve spatial resolution.
Remember that the MR scanner is attempting to measure the radio signal res­
onating from a drop of water that is as small as 1 mm x 1 mm x 2 mm. There
are significant trades to be made when attempting to increase spatial resolution.
Signal to noise will degrade as electronic distortion of the antenna coil and the
amplifiers begin to overcome the small signals involved in diagnostic imaging.
Imaging with larger voxels, thicker slices, or repeating and averaging multiple ac­
quisitions are solutions. The relative consequences are increased aliasing, partial
voluming, and possible patient motion artifact.
MRI data is subject to several artifacts. The issues of partial voluming, patient
motion, and aliasing are common with CT. While MRI does not have x-ray-related
artifacts, it has its own host of radio and magnetic artifacts of which the visual­
ization expert should be aware. Like any microwave used for cooking, the mass
of the patient (or food) will affect how well the material absorbs the radiofre­
quency energies. This leads to "cold" spots in the imaging volume. In addition,
the distribution of the antenna coverage for both transmission and reception and
inhomogeneity in the induced magnetic field lead to inconsistent quantitative val­
ues and even inaccurate geometry within an image. There is significant research
being conducted addressing methods to correct for these distortions.
Most MRI scanners can acquire a set of slices (30 to 50) within five to ten
minutes. An entire study of a patient generally represents two to three series of
slices, with a study time of 30 to 45 minutes. Each slice generally represents a
1 .3. Medical Imaging Technology 13

thickness of 2 to 1 0 mm and contains 256 x 256 pixels. As with CT, the pixel
dimensions are set by the image matrix and the field of view parameters.
New work in MR imaging is leading to novel capabilities in this modality.
Studies in perfusion and diffusion of various agents across the body are being
enabled by new capabilities in fast imaging. These studies can illuminate blood
flow, creating high-resolution images of vascular structure. Functional MRl has
been introduced not only to record the patient's anatomy, but also the physiolog­
ical functions of the tissues being studied, largely used today to map the cerebral
cortex of the brain. The complex pulse sequences of Diffusion Tensor Imaging (or
DTI) are beginning to reveal the direction and course of nerve bundles deep wi thin
the white matter of the brain, providing new research areas of tractography and
patient specific anatomical analysis. The pace of research in MRl is increasing as
new capabilities are enabled.

1 . 3.3 N uclear Medicine

Imaging using pharmaceuticals tagged with radioactive agents is an older tech­

nology than either CT or MRl . In nuclear medicine imaging, a radioactive source
is injected into the patient and the distribution of that source is measured using
a detector array to catch and quantify the radiation emitted from the body. The
detector array is sometimes configured similar to that for CT, while other systems
utilize a two-dimensional array of detectors. The major difference between CT
and this type of study, is that with CT the radiation source's location is external
and known, while in nuclear medicine, the source is internal and its distribu­
tion unknown. Figure 1 .3 is a picture of a two-headed Gamma Camera (Anger
Camera) used for Single Photon Emission Computed Tomography (SPECT). The
gantry has two detector arrays for gamma radiation emitted from the patient. The
detectors revolve about the patient, and tomographic images are reconstructed.
While the images generated from measuring radioactive decay are often blurry
with poor spatial resolution, nuclear medicine allows clinicians to image the phys­
iological activity (the biological activity) of the patient, rather than just the geom­
etry of the patient's anatomy. The selection of the pharmaceutical to which the
radionuclei is attached will determine the specificity and performance of the imag­
ing study. For instance, radioactive iodine generates very specific studies of the
thyroid and parathyroid glands. Radioactive tracers can be attached to hundreds
of different substances.
Nuclear medicine studies produce images with fairly low resolution and a
high amount of noise. This is due to our inability to use high radiation doses
because of the consequences of high doses of radiation to the patient. Choosing
persistent radiopharmaceuticals will cause the agent to reside longer within the
body, improving our capacity to resolve interal objects while potentially inflicting
additional damage on the patient.
14 Chapter 1 . Introduction

Figure 1.3. A two-headed Gamma Camera used in nuclear medicine (circa 1998).
1 .3. Medical Imaging Technology 15

Most nuclear medicine is still based on the 2D planar representations

generated by simply recording the activity of a detector over a period of time.
If the detectors are rotated about the patient, reconstruction is possible. The two
most common types of 3D nuclear medicine studies are Single Photon
Emission Computed Tomography (SPECT) and Positron Emission Tomography
(PET). SPECT studies utilize radiotracers that emit photons while decaying.
The radioactive agents used in SPECT have half-lives measured in hours and
can be easily generated or transported to the clinic. PET studies use radioactive
isotopes that decay by positron emission. The resulting positrons annihilate
when they encounter electrons generating two high energy photons that
conserve momentum by traveling in opposite directions. The half-lives of the
isotopes used in PET are often measured in minutes, necessitating their
production at the clinic. A particle accelerator such as a cyclotron is required
to create most of the radionuclei for PET imaging, increasing the cost of a
PET installation.
The output of a SPECT or PET scanner is typically a set of 10 to 30 transaxial
slices. Each slice is typically 5 to 1 0 mm thick and contains pixels 5 to 1 0 mm
in size. There may be gaps between the slices. Often multiple sets of scans are
taken five to fifteen minutes apart allowing for time-resolved studies. Nuclear
medicine studies produce images which have few anatomic cues. The images
mimic physiologic activity which may or may not be easily transferable to the
underlying anatomy.

1 .3.4 U ltrasou nd

Unlike the three other modalities described in this chapter, Ultrasonography uses
classical physics to perform its imaging rather than the more esoteric phenomena
of nuclear decay or x-ray quantum mechanics. No photons are used. A piezoelec­
tric quartz crystal, similar to those used to clock CPUs, wristwatches, or to start
electronic ignition lighters for barbeque grills, is used to create high frequency
acoustic energy (3 to 10 Megahertz) which is then reflected off surfaces and in­
terfaces between organs deep within the body.
The same device that is used to create the acoustic signal, a transducer, is then
used to measure the returning echo information. Partial reflections are created by
interfaces between areas of differing acoustical impedance. The result is imaging
very similar to the SONAR systems used in maritime and undersea naval imag­
ing. The sonographer places the probe against the patient and moves it to obtain
images of various parts of the body. The result is usually a 2D representation of
the area under the transducer. Most ultrasound machines consist of a linear ar­
ray of transducers, and produce an image representing a pie-shaped slice of the
body. One of the advantages of ultrasound is that it produces images in real time.
Another advantage is the absence of ionizing radiation.
16 Chapter 1 . Introduction

Ultrasound machines are fairly inexpensive compared with the other diagnos­
tic 3D modalities in common use in the health care industry. A high-end diag­
nostic ultrasound system can be purchased for approximately $250,000, while
functional, more moderately priced units can be acquired for less than $1 00,000
today.
3D images can easily be created by extruding several slices through space.
Most commercial ultrasound equipment allows for 3D imaging; however, accu­
rate spatial information is seldom provided. Spatial tracking of the transducer
array is imperative for clinical 3D visualization. Some approaches to correcting
this defect have been to use physical, optical, and electronic devices for locating
the transducer in space. Providing a rigid path for the transducer creates some
compelling results. Other techniques involve rotating the transducer assembly,
sweeping a volume in a cylindrical section similar to aeronautical RADAR imag­
ing. Finally, recent advances in transducer design have yielded demonstrable 3D
transducers that can acquire volume images from ultrasound with no mechanically
moving parts.
However, once the position and orientation of the probe are known, the data
are still often sampled in irregular intervals, and defy many image processing
techniques. Ultrasound images typically contain a large amount of noise termed
speckle, that adds to the problem of identifying structures. Object detection in
volume ultrasound is the subject of much advanced medical image processing
research.

1 .4 Acq u isiti o n , Analysis, and Visual ization

Most acquisition of medical data is not optimized for 3D presentation. Indeed, to

improve the signal to noise, thick slices are usually requested. The result is better
contrast at the cost of higher partial-voluming artifact and poor spatial resolution
in the effective z-direction. In order to reduce the dose to the patient and the
time required to perform a scan, slices are not always contiguous. Rather small
gaps are introduced between slices. Since most presentation is 2D, these practices
seldom affect the clinician; however, they can be fatal to 3D visualization.
Radiologists often must balance the need for image quality with the interests
of the patient. Many times the means to improving the diagnostic power of an
image will mean increased dose to the patient. The factors affecting the trade-off
between image quality and acceptable dose include a wide variety of concerns.
For example, since the inadvertent long range biological effects (chromosomal
damage) of ionizing radiation are most profound in children, pediatric radiol­
ogy seldom trades increased dose for improved contrast. Similar trades are made
throughout medical imaging (e.g., trading spatial resolution or increased imaging
time in MRI for an improved signal-to-noise ratio). The computer scientist inter-
1 .5. Summary 17

ested in medical visualization would be well served to learn where these trades
are being made and learn to cope with their consequences.
An even deeper understanding of the clinical findings of a case is required to
validate a 3D visualization. Knowledge of what the physician is seeking in the
patient will help to direct the acquisition so that the results not only capture the
desired pathology in high detail, but also assure the computer scientist that the
data are in an appropriate form for advanced rendering.
In his 1 993 presentation on 3D Medical Visualization, Derek Ney, an assistant
professor for the Johns Hopkins Department of Radiology, wrote:

A succinct statement of the relationship of visualization and acquisi­

tion is that the single easiest method for improving visualization is to
use better (higher resolution, higher contrast, higher signal to noise,
etc.) acquisition techniques. This implies that the method for acqui­
sition should be as optimal as possible. There is no room in most
systems for poor acquisition technique. The researcher in visualiza­
tion is warned to learn about the potential pitfalls in the acquisition
stage, so that they are presented with data that was acquired with the
best technique possible.

Indeed, as with all data processing, "if garbage goes in, only garbage comes
out." Or more precisely, if one beautifully renders garbage data, then one only
has a picture of well-dressed garbage. The result serves no one in particular and
wastes valuable resources that would be better used to serve the patient, to serve
the community, and to serve future generations of researchers and patients alike.

1 .5 S u m mary

Keeping the basic concepts of medical image acquistion in mind, we tum our
attention to the processing and analysis of the generated data. The multiplica­
tion of imaging modalities with differing strengths and weaknesses in imaging
anatomy v. physiology and hard tissue v. soft tissue requires careful registration
and alignment of multimodal data, and differences in size and resolution lead to
multiscale methods as well. We refrain from covering the algorithms and methods
for generating visualizations of the data, instead concentrating on the analysis and
refinement of the data stream in preparation for the ultimate creation of rendered
images of the patient anatomy and physiology.
This book is divided into three main parts:
• Basics: These chapters describe techniques for preprocessing the image
data before complex semantic, know lege-based operations are performed
either by the machine or the human operator. Statistical, linear, and more
18 Chapter 1 . Introduction

advanced nonlinear methods are covered, matching implementations from

the software.
• Segmentation: As stated before, this is the task of partitioning an image
(2D array or volume) into contiguous regions with cohesive properties. A
variety of approaches are given including hybrid mixtures of statistical and
structural methods. We cover the use of finite element models, Voronoi
decomposition, fuzzy-connectedness, and other techniques useful for seg­
mentation implemented within ITK.
• Registration: This is the task of aligning multiple data streams or images,
permitting the fusion of different information creating a more powerful di­
agnostic tool than any single image alone. We cover basic registration
strategies as well as more complex deformable techniques, all implemented
under the umbrella of the ITK registration framework.
This book is intended as a companion to the Insight Toolkit covering the prac­
tices and principles used to develop the corresponding software. Implementation
and integration details are covered elsewhere in the ITK Software Guide. The ulti­
mate arbiter of all of this material is the source code for ITK itself, made available,
by intent and design, in the public domain. We hope that this text serves as a sum­
mary to guide and explain the many choices made and roads taken in the design
and use of the software.
Enjoy ITK.
 CHAPTER
TWO

Basic Im age Processing and

Linear Operators

Terry S. Yoo
National Library of Medicine, NIH

George D. Stetten
University of Pittsburgh

Bill Lorensen
GE Global Research

2.1 I ntroduction

The topics discussed throughout this book fall into the realm of what is usually
called image analysis. Before researchers adopted that term, they often referred
to their field as image processing, an extension of traditional signal processing
to images. An image can be thought of as a signal in space rather than time,
arrayed in two or three dimensions rather than one. This leads to fundamental
differences between conventional signal processing and image processing. While
one generally only moves forward through time, in space, one moves left or right
with equal ease, as well as up or down, forwards or backwards. The extension
to multiple dimensions raises new issues, such as rotation and other interactions
between the axes.
As with conventional signals, images can be either continuous or discrete.
Optical images, consisting of light, call for analysis in the continuous domain, as
they pass through lenses, for example. In contrast, images found on a computer

19
20 Chapter 2. Basic Image Processing and Linear Operators

generally consist of individual samples on a regular grid, and therefore require

discrete analysis. Digital sampling theory is intrinsic to image processing. Fa­
miliar artifacts such as aliasing can be understood through digital sampling and
filtering theory. Likewise, effective reconstruction, the inverse of sampling, re­
quires a clear understanding of digital sampling theory.
Both continuous and discrete image analysis often begin with linear operators
adapted from signal processing. These operators range from simple algebra to
convolution and the Fourier transform. Linear operators assume that the light or
other input adds by superposition, i.e., that individual components of the input do
not affect each other within the system. Linear operators preserve the indepen­
dence of basis functions useful in extracting information from an image. Some
linear operators affect each sample in the image, or pixel, independently from its
neighbors, while others take neighboring pixels, or even the entire image, into
consideration at once.
This chapter is a review (or introduction, depending on your background) cov­
ering the digital theories that underlie sampling and linear operators. The material
will be covered from a practical viewpoint: What should we know about the ba­
sic operators of image processing? What do they mean in terms of getting at the
underlying information in the image? What are the repercussions of digital sam­
pling theory? The presentation is cursory; we will not explore proofs or rigorous
mathematics associated with the presented theorems or tools. For example, the
presentation of the Fourier transform will be far from exhaustive, not even listing
all of the important mathematical properties of the transform. We also will not be
presenting implementations. For more in-depth discussion, we refer the reader to
standard textbooks in the field, in particular, Signals and Systems, by Oppenheim
and Willsky [2], and Digital Image Processing by Castleman [ 1 ] .

2.2 I mages

In this book, we often are concerned with 3D, or volume, images. A volume
image, cp, is a mapping from �3 to �n where n = 1 for the typical, scalar-valued
volume. More precisely:

cp : U f-t � and U C �3 , (2. 1)

where U is the domain of the volume. The image cp is often written as a function
cp(x,y,z).
Throughout this chapter, many examples will be presented as either ID or 2D,
where they may be more easily examined (at least in print), and then generalized
to higher dimensions. In general, for a specific number of dimensions, we will
use the corresponding function notation (cp(x) , cp(x,y) , or cp(x,y,z) ).
2.3. Point Operators 21

We denote the sampling of continuous images into discrete, digital images as

follows. If F is a discrete sampling of a 2D image <1> (x,y) then we can say that

(2.2)
We likewise use notation to distinguish between discrete and continuous differ­
ential operators. To find the partial derivative in the x-direction of a discretely
sampled image we can say that

(2.3)

using a method such as

oxFiI,} =
-
. Fi+l,j - Fi-l ,j Fi+ l,j - Fi- l,j (2.4)
Xi+l - Xi-l 2h
to approximate the partial derivative, where h is the grid spacing (nonnally as­
sumed to be 1 pixel). The value OxFi,j is an approximation of the instantaneous
partial derivative of <1>(Xi,Yi) at (Xi ,Yi ) in the x-direction. Equation (2.4) is the
method of central differences, one of a number of commonly used methods to
approximate the derivatives of unifonnly, discretely sampled datasets.

2 .3 Po i n t Operators

The simplest operators on an image are those that treat each pixel independently
from its neighbors. Point operators usually ignore infonnation about pixel loca­
tion, relying only on pixel intensity. A surprising number of useful things can be
done with such operators. Some examples follow.

2.3.1 Thresholding

The simplest fonn of segmentation involves thresholding the image intensity.

Thresholding is inherently a binary decision made on each pixel independent of
its neighbors. Intensity above the threshold yields one classification, below the
threshold another. This simple operation can be surprisingly effective, especially
in data such as Computerized Tomography (CT), where the pixel value has real­
world significance. The pixel value in a CT image is reported in Hounsfield units,
which are calibrated to correspond to the attenuation of x-rays measured within
the particular sample of tissue. Since bone is much more radiopaque than other
tissues, bone can be segmented effectively in a CT image by setting the threshold
between the Hounsfield value for bone and other tissues. An example is shown
in Figure 2. 1 , in which a surface was reconstructed at the threshold between bone
and other tissue in a CT scan of a fractured tibia.
22 Chapter 2. Basic Image Processing and linear Operators

Figure 2.1. A simple threshold segments the bone from other tissue in a cr scan of a
fractured tibia.

2.3.2 Adaptive Thresholding

In most applications of thresholding, the optimal threshold is not known before­
hand, but depends instead upon the particular imaging modality, tissue-type, pa­
tient. and even the individual image acquisition. For example. the imcnsity of a
given pixel in an ultrasound image will depend on the scanner's current gai n set·
tings and the intervening tissue types. In such cases, an optimal threshold might
be chosen manually by looking at the image itself or at a histogram of intensity
values from the tissues 10 be segmented. Ideally, such a histogram will look like
Figure 2.2, in which the two tissue types (A and B) have pixel intensities whose
ranges hardly overlap. In such a case, where the histogram is bimodal, it is fairly
easy to select the optimum threshold. In the real world, this is often not the case,
however, and segmentation must take other factors into account, such as neigh·
borhood statistics and expected shape.

2.3.3 Windowing
Often it is advantageous to change the range of intensity values for an image to
adjust the brightness and contrast of what will be displayed on a screen. Dis­
play is an essential pan of image analysis, because it is the portal 10 the human
visual system, whose capabilities far surpass any computer system likely to be
2.3. Point Operators 23

Figure 2.2. Histogram h( <1» of pixel intensity <I> for tissue types A and B, with optimum
threshold for segmentation (arrow).

built in the near future. The dynamic range of image display systems, between
the brightest and darkest pixel the screen can produce, is quite small compared
to the dynamic range of the human eye. Many medical imaging modalities are
also capable of a greater dynamic range than can be displayed. The display itself
therefore represents a bottleneck in the overall system. To minimize loss of in­
formation, the brightness and contrast of the image may be adjusted so that the
brightest and darkest pixels in the image exactly match the extremes of the dis­
play, as shown in Figure 2.3. This is a linear function remapping the intensity of
each pixel, with contrast being the slope of the function and brightness being the
offset. In some clinical systems, the process is called windowing.

h ( <I»

max '"

Figure 2.3. Windowing intensity histograms. Left: Histogram h(<1» of pixel intensity
<I> of original image, on a scale from zero to the maximum value of the display; Center:
windowing function 'V 1(<1» to yield a new intensity 'V; Right: new histogram
= h('I') in
which the dynamic range of the image matches that of the display.
24 Chapter 2. Basic Image Processing and Linear Operators

h ( cj» 'II = / (cj»

cj> cj> 'II

Figure 2.4. Histogram equalization. Left: Histogram h(<1» of pixel intensity <I> of origi­
nal image; Center: equalizing function 'I' f( <1» to yield a new intensity '1'; Right: new
histogram h('I') in which each intensity is equally represented.
=

2.3.4 Histogram Equ al izatio n

A further step may be taken to optimize the match between the image and the
display system, which ensures that each level of pixel intensity is equally rep­
resented in the image. This is called histogram equalization, and although it is
non-linear, we include it here as a useful point operator. As shown in Figure 2.4,
histogram equalization entails finding a monotonic function ", = f ( cj» that remaps
pixel intensity so that the histogram becomes a constant function of intensity ",.

2.3.5 Color Maps

Color video images typically have three separate channels to record the red, green,
and blue levels from the camera. Color displays are capable of communicating
these images to the human visual system, but such displays also can be used to
enhance the display of single-channel gray-scale images through the use of colors,
artificially assigned to particular pixel intensities. Such color maps are often used
to overlay additional information in anatomical images, such as Doppler flow in
ultrasound images. The color map usually follows some perceptually consistent
ordering scheme, such as the rainbow, or especially bright colors may be assigned
to emphasize pixel values above certain levels.

2.3.6 Algebraic Operators

Some point operators combine two images, keeping each pixel separate from its
neighbors, while combining it with the corresponding pixel in another image.
Algebraic point operators may be used to add images together, for example, in
averaging many images acquired from the same sample to reduce noise. Or they
2.4. Linear Filtering 25

may be used to subtract one image from another, such as in Digital Subtraction
Angiography where the difference between sequential fluoroscopic images, be­
fore and after contrast, leaves just the contrast without the anatomical structures.
Another use of algebraic point operators is to mask an image by multiplying it
by a second, binary image, containing ones where the first image is to be passed
through, and zeroes where it is to be masked.

2.3.7 Location-Dependent Point Operators

All the point operators described so far apply the same operation to every pixel
in the image. In some cases it is useful to make the operation depend upon pixel
location, for example, to correct for spatial inhomogeneity in the image acquisi­
tion. Pixels are kept independent of each other, but in a framework where location
within the image makes a difference.

2.4 Linear Fi ltering

Although many useful things can be done to an image using point operators on
individual pixels, most image processing tasks require the simultaneous consider­
ation of multiple pixels. Techniques that combine multiple pixels in a linear and
space-invariant manner are known collectively as linear filtering. Two standard
techniques for linear filtering, convolution and the Fourier transform, comprise
the rest of this chapter.

2.4.1 Convol ution

If you have ever adjusted the focus on a camera you have performed convolution
in the continuous domain. If you have run a blurring function or performed edge
enhancement using a commercial graphics program, it is likely that you were
using discrete convolution. In image processing, convolution filters are used to
make measurements, detect features, smooth noisy signals, and deconvolve the
effects from image acquisition (e.g., deblurring the known optical artifacts from
a telescope). The effect of the particular filter depends on the nature of the input
filter, or kernel.
Convolution, denoted with the operator 181 can be described in I D as a contin­
uous operation applying the filter kernel h(x) (itself an image) to an image <j>(x)
using the following integral form:

<j>(x) Q9 h(x) = f: <j>(x - 't)h('t) d't. (2.5)

Note that the expression <j>(x) 181 h(x) itself describes an image mapping. Thus
the following equality holds:
26 Chapler 2. Basic Image Processing and linear Operators

�(x) ® h(x) � (�® h)(x). (2.6)

The expressions !j)(x) 0 h(x) and ($ 0 h)(x) are used interchangeably, with the
position index x often omitted to help streamline and clarify the nOlation espe­
cially when higher dimensional (e.g., 3D) images are involved. In the practice of
digital image processing, hex) typically does nOI have infinite extent, nor is it in­
finitely dense; however, when working in the continuous domain of linear filtering
theory, considering kemels that are infinitely wide (thus avoiding the complica­
tions of truncation) is often morc convenient. We will elaborate later on discrete
functions with finite extent in the discussion on sampling.
We can generalize the I D convolution operation to 20 and 3D images. Thus,
in 20, convolution becomes

Nx,y) ® h(x,y) � [:[: �(x - ',y - v)h(', v)dvd,. (2.7)

Similarly, convolution in 3D is expressed as

$(x,y,z) ®h (x,y,z) =

1:1:1: $(x - "t,y - V,Z - ro)IJ{"t, v,ro)drodvd-t. (2.8)

2.4.2 Example-Convolution as Noise Reduction (1 D)

One of the most common uses of convolution is as a filter to suppress noise in an
image. Consider the Gaussian function as a smoothing filter kernel. If convolved
with an image containing high-frequency noise, the result is an output image that
locally averages the image intensities, reducing noise at the cost of some "sharp­
ness" of the original image. That is, the resulting image has less high-frequency
noise, but also has blurred edges.

Figure 2.5. A 1D noisy step edge: an example input signal q,(x).

2.4. Linear Fillering 27

Figure 2.5 shows a noisy 1 D step discontinuity, or edge. The function $(x) is
shown from x = 0 to x = 255 with a step function at x = 128. The signal to noise
ratio is approximately 4: I.
Figures 2.6. 2.7. and 2.8, show a progression of filtered versions of the input
from Figure 2.5 with Gaussian filter kernels of increasing aperture or width a.
The Gaussian function g(x) is defined as

, �
g {x,a) = p;=:: e-2a! (2.9)
v 27ta

0.'"

• ,.
.,
(,) (b) (c)

Figure 2.6. Gaussian filtering of input �(x) where (I = 16. (a) Input 4I(x); (b) a ID
Gaussian kernel, 8(X,(I) where (I = 16; (e) output of �(X)0g(x, 16).

'''1..

:��J��� �
.,
•
�
., ,. ••

.,.. .M '''1 -.. ,..

(a) (b) (c)

Figure 2.7. Gaussian filtering of input 4I(x) where (I = 24. (a) Input 4I(x); (b) a ID
Gaussian kernel, g(x,(I) where (I = 24; (e) output of 4I(x) <8> g(x, 24).

'"

(,) (b) (c)

Figure 2.8. Gaussian filtering of input 4I(x) where (I 32. (a) Inpul ¢I(x); (b)
= a lD
Gaussian kernel. g(x,(I) where (I = 32; (e) Output of,(x) <8>g(x, 32).
28 Chapter 2. Basic Image Processing and Linear Operators

Notice how increasing the aperture of the Gaussian decreases the noise but
also blurs the edge. This trade-off of resolution for noise reduction is one of the
many considerations in the design of linear filter systems. An entire approach
to image analysis, called "scale space," depends upon the Gaussian to produce a
broad range of blurred versions of a given image.

2.4.3 Properties of the Convolution Operation

The nature of a linear filtering operation depends on the properties of the filter
kernel. For instance, the "shape" of a 2D or 3D kernel will determine whether
the operation remains invariant with respect to rotation. Independent of kernel
shape, however, the convolution operation has many useful properties, including
the following:

Convolution is linear:

(Ap + Bq) 0 r = A(p 0 r) + B(q 0 r) . (2. 1 0)

Convolution is commutative:

p 0 q = q 0 p. (2. 1 1)

Convolution is associative:

(p 0 q) 0 r = p 0 (q 0 r) . (2. 1 2)

Convolution is distributive over addition:

p 0 (q + r) = p 0 q + p 0 r. (2. 1 3)

These combined properties create the justification for using convolution as the
principal operation in linear filtering.

2.4.4 Differentiatio n by Convolution

Differentiation may be accomplished by using convolution. We will see how this

works in the discrete domain a little later, but first let us consider differentiation
in the continuous domain. We can explicitly denote differentiation as convolution
using the 0 operator:

(2. 14)

Equation (2. 14) depicts the differential operator as a kernel by which convolution
accomplishes differentiation. This is hard to illustrate using the ideal differential
2.4. Linear Filtering 29

� r\��� ..•
•
.-

.-

i\�� ..
.. ..,

jk - -
....., .• •

(a) (b) (e)

Figure 2.9. Taking derivatives of a noisy input by convolution with the derivative of a
Gaussian kernel. (a) I-D input 41(-,); (b) tg(-" a) where a = 3; (c) 41(-,) 0 tg(-" O").

kemel, because such a kernel has infinitesimal width and infinite height. In the
real world such a kernel is impossible. In any event, because differentiation en­
hances high-frequency noise, it is often necessary to regularize (smooth) a noisy
image before computing its derivative, by first convolving the function with a
smoothing kernel h{-,). It follows from the associative and commutative proper­
ties of convolution thaI

(2.15)

In other words, the derivative of a function $(x) convolved with a filter kernel
h(x) is equivalent to convolving 1/I(x) with the derivative of h(x). This suggests
that one of the easiest ways to compute the derivative of a function is through con­
volution with the derivative of some smoothing kernel. We have already demon­
strated the use of the Gaussian g(x) as a smoothing kernel. The Gaussian's in­
finitely differentiable propenies make it attractive for taking derivatives as well as
smoothing.
Figure 2.9 depicts a noisy 10 input signal for which derivative infomJation is
desired. When convolved with the derivative of a Gaussian, the resulting output
repons the derivative of a smoothed version of the input. By the commutative
and associative properties of convolution, it can just as well be described as a
smoothed version of the derivative of the input. This technique for differentiating
functions can be extended to higher order derivatives, and to higher numbers of
dimensions.

2.4.5 Convolution of Discretely Sampled Data

Convolution of discretely sampled data is similar to the continuous form, except
that discrete summation is substituted for integration. Also, since discrete filter
kernels cannot be implemented with infinite extent, the limits of the summations
do not range from _00 to 00 but are rather constrained to the size of the filter
30 Chapter 2. Basic Image Processing and Linear Operators

kernel. In 2D, a discrete convolution of image P with finite kernel Q looks like
domainx[Q] domailly[Q]
Px,), (9 Qx,), = L L Px-i,y-j Qi,j. (2. 16)
i j
The discrete version of convolution shares all the above-mentioned attributes
of its continuous cousin. Convolution of discretely sampled data is linear, commu­
tative, associative, and distributive over addition. As in the continuous domain,
convolution in the discrete domain can be used to smooth noisy data or detect
boundaries, depending on the choice of kernel. An example of discrete convo­
lution for differentiation in 2D has already been seen in the central differences

[
operation (Equation (2.4)). The kernel in that case would be
0 0
-0.5 0 (2. 17)
o 0
which has the effect of subtracting the pixel to the left from the pixel to the right,
just as Equation (2.4) specifies, to represent the x-component of the gradient.
The following section shows more examples using convolution for smoothing and
taking derivatives in 2D.

2.4.6 The Bi nomial Kernel

One particularly useful kernel for discrete convolution is the binomial kernel, gen­
erated by repeated convolutions with a simple box of identical values, as shown
here:

[: :H� �n
[ 11 , (2. 1 8)

With successive iterations, the binomial kernel approaches a Gaussian shape (by
the Central Limit Theorem). The scale of the Gaussian is determined by the
number of iterations. (The kernels shown in Equation (2. 1 8) would be normalized
by 114 with each iteration.)

2.4.7 Difference of Offset Gaussians

Boundaries in an image represent areas of high gradient magnitude, i.e., the image
intensity increases or decreases rapidly across the edge. The gradient vector,

(2. 19)
2.4. Linear Filtering 31

is oriented in the direction of the steepest change in image intensity, normal to the
implied boundary. The gradient magnitude,

(2.20)

represents an orientation-independent measure of boundary strength. How can we

determine the individual components of the gradient vector?
We have seen the central difference operator already. Another way to rep­
resent a gradient component is to use a kernel known as the difference of offset

[ 1
gaussians (DoOO).
-1 -2 0 2 1
-1 o 1 ], -2 -4 0 4 2 . (2.21 )
-1 -2 0 2 1
The two DoOO kernels shown in Equation (2.2 1 ) correspond to iterations 1
and 3 in Equation (2. 18). Each kernel is the difference between two copies of
the corresponding binomial kernels displaced along the x-axis. Convolution with
these DoOO kernels measures the gradient component in the x-direction. Simi­
lar kernels can be constructed by displacing copies of the binomial kernel in the
y-direction to measure the y-component of the gradient. DoOO kernels can be
constructed for images in 3D or higher dimensions. The number of iterations of
the underlying binomial kernel determines the scale of the DoOO.
Figure 2.10 shows the results of detecting the individual gradient components
as well as the gradient magnitude of a simple image containing a rectangular
object. Notice the orientations of the edges detected by the individual gradient
components, and the orientation-independence of the gradient magnitude.

2.4.8 20 Example-Higher-Order Differentiation

We have seen how the gradient vector represents the strength and orientation of
the boundary. But what of higher order differentials? Where the first derivative
(gradient) is represented as a vector, the second derivative is a matrix, known as
the Jacobian

[ � a1
dYdx
a2 e!> a2 e!>
ay
•

The elements of this matrix are useful in a number of ways. For example, the
(2.22)

Laplacian,

(2.23)
32 Chapter 2. Basic Image Processing and linear Operators

Figure 2.10. Gradient components and gradient magnitude: (A) original image; (8) x­
component of gradient; (C) y-component of gradient; (D) gradient magnitude.

sums the diagonal tenns of the Jacobian to yield a rotationally invariant repre­

[
sentation of the second derivative of image intensity. A common 3 x 3 kemel
representing the Laplacian in 20 is

1
\ -\ \
-\ 8 -\ . (2.24)
-\ -\ -\

Convolution with this matrix yields zero along a boundary, no matter what
orientation the boundary has. For example, convolution with the following 3 x 3
patch of an image containing a diagonal boundary (between regions with intensi­
ties of I and 3 respectively),

[: ; n
yields zero for the center pixel (which is directly on the boundary),
(2.25)

For an intuitive understanding of why the second derivative should be zero on

the boundary let us examine the ID case.
2.4. Linear Filtering 33

af
f
ax

x x

Figure 2.1 1 . ID case of the intensity function across a boundary, its first derivative, and
its second derivative showing a zero crossing at the boundary location.

As shown in Figure 2 . 1 1 , the boundary is located at an inflection point (arrow)

in the intensity curve
<I> (x)
corresponding to a maximum in the first derivative

and a zero crossing in the second derivative

In two or more dimensions, we use the Laplacian to capture the "total" second
derivative at a pixel. Mathematically, the Laplacian is the divergence of the gra­
dient. Divergence is a common concept in fluid dynamics, describing how much
more fluid enters a region than leaves it. In our case, it describes how much more
gradient is "entering" a pixel than "leaving" it from any direction (making it ro­
tationally invariant). A boundary exists where there is no net change in gradient
(i.e., the gradient is at a maximum on the boundary), making the Laplacian zero,
just as in the ID case. The difference of Gaussian (DOG) kernel, which we will
discuss in Section 2.4.9, yields results similar to the Laplacian, because the DOG
kernel has a similar shape to the Laplacian.
Let us now consider the terms off the diagonal of the Jacobian matrix (Equa­
tion (2.22» . These can be used to measure the curvature of boundaries. The
individual components of the gradient at a curved boundary change as one moves
orthogonally to that particular component. This type of change shows up as partial
second derivatives of intensity off the diagonal of the Jacobian. For illustration,
Exploring the Variety of Random
Documents with Different Content
Clive, I am positive that the man was none other than the original of
this likeness, and——"
He was interrupted by a passionate cry of pain and anger, and
Ralph, snatching the photograph from his hand, stood confronting
him with blazing eyes.
"It is false!" he cried. "You know it is false! I believe that you are
responsible for my father's disappearance!"
 CHAPTER III
MR. ST. CLIVE PROVES HIMSELF A TRUE
FRIEND

"I believe that you are responsible for my father's disappearance."

So did Ralph Rexworth cry in his anger; and Lord Elgert started, and
his face grew dark with rage.
"You impudent young dog!" he shouted, raising his stick; and the
blow would have fallen, had not Mr. St. Clive stopped it with his arm.
"Lord Elgert," he said sternly; for he was shocked at the callous way
in which the charge had been made, "I cannot stand by and allow
that. You have made a very serious charge——"
"Nothing so serious as that young rascal has made. I am surprised
that you stand by and listen to it, St. Clive; but you always were
antagonistic to me! I assert what is fact. My place was broken into
——"
"Did any one but yourself see this man?"
"An absurd question! Who was there to see him? By the time the
alarm was given he was gone. I shall have to tell the police of that
photograph; it will be wanted to help in tracing him. I expect this
story is all nonsense; and upon inquiry it will be found that the
farthest these two have travelled is from London. Most probably this
boy, who makes such unfounded charges, knew well the business
which brought his father here. The story of what happened in the
woods is really too romantic. If two people were there, the second
was most likely an accomplice; and they have gone off, leaving the
boy here to see what he can learn, or pick up. You are easily
deceived, St. Clive." And Lord Elgert turned upon his heel with a
mocking laugh.
But ere he could go, Ralph stood in his path, regarding him with a
fixed stare.
"I do not know you," he said. "I never saw you before; but I can tell
friend from enemy, and you are an enemy. I am only a boy; but one
day I will bring your words back to you, and make you prove them."
"Out of my way, you young rascal!" came the answer, "or I will have
you in prison before long. St. Clive, I wish you joy of your young
friend. Take my advice, and keep a sharp eye on the silver, if you
suffer him to enter your house."
Ralph would have surely been provoked into some foolish action had
not Mr. St. Clive laid a gentle hand upon his shoulder, and led him
back into the inn; and then the boy quite broke down.
"Oh, sir! Oh, sir!" he cried. "To say such things about my dear father
—my dear, kind father! But he shall prove them," he added fiercely.
"I will make him prove them. I believe that he knows something."
"Ralph," answered Mr. St. Clive quietly, "because Lord Elgert has
been both unkind and foolish, that is no reason why you should talk
wildly. To say that Lord Elgert has had anything to do with your
father's disappearance, seems to me to be the very height of folly.
He is a rich man, and one of our justices——"
"Where does he live, sir?" queried Ralph suddenly.
"At Castle Court, near Great Stow. Ah," he added, as he saw Ralph's
look, "I know what you are thinking—that it is in the direction
whither your father was going! But remember, that will be equally
applicable to Lord Elgert's story that your father was going there. It
is most likely that some one in a measure resembling your father, did
break into Castle Court—we have not the slightest reason for
discrediting Lord Elgert's statement—and in the confusion of the
struggle, he did not clearly distinguish his opponent, and so says
that he resembles this photograph. Mistaken identity is a common
occurrence, and——"
"You do not believe his story, sir? I could not bear to think that."
"I do not, Ralph. If I did so, I should still feel my debt of gratitude to
you; but I do not believe it. I am not so foolish as to mistake
between a gentleman and a thief; and though I have not seen your
father, I think that I can see him in you and your manner. Now be
brave, and do not trouble about what his lordship said. He was
angry because you spoke as you did; and though it was natural,
your language was not very polite." And Mr. St. Clive smiled slightly.
"Now let us talk sensibly. First, you cannot stay here by yourself;
therefore, disregarding the warning I have received, I invite you to
be my guest for the time, until we can see what is best to be done.
What money have you of your own?"
"Only a few shillings, but there is the purse, sir." And Ralph opened
the purse which they had picked up in Stow Wood. "Here are five
sovereigns, and two five-pound notes, sir."
"Then we had better pay the innkeeper and make a start. Simon"—
as the old fellow came in answer to the bell—"I am going to take
this young gentleman home with me. If his father should return, or if
letters arrive, you will let us know. Make out your bill. And, Simon, I
suppose that you did not recognize Mr. Rexworth at all?"
"Why, no, sir; I cannot say that I did! But he knew the place, sir;
and when I told the girl to show him up to No. 10, sir, he just went
straight up to it. He knew the Horse and Wheel, sir."
"Well, get your bill ready."
The old man went out. It was something of a relief to know that he
was going to be paid; for he had begun to have some doubts about
the matter.
So it came about that Ralph Rexworth was taken home by Mr. St.
Clive; and there he was received with kindness and warmth by that
gentleman's wife, while little Irene smiled shyly, and put out one
dainty little hand for him to take in his brown palm.
"I thank you very much," the little lass said. "I think that horrid bull
would have killed me if it had not been for you." And Mrs. St. Clive
shuddered as she listened; for her husband had told her how great
was the peril from which Irene had been rescued.
Leaving the two young people to make friends, Mr. St. Clive took his
wife aside and told her of the strange position in which their young
guest was placed.
"The boy does not seem to have a friend in the world," he said. "And
he is undoubtedly a gentleman, Kate. What is to be done? His father
may return; but I confess that it looks as if a tragedy had taken
place. It was wonderful how the lad pieced together traces which
were invisible to me. I fear that something bad has occurred. As to
Lord Elgert's idea, I do not put much faith in it. Elgert is too fond of
thinking evil of people—he is one of the most merciless men on the
bench. What shall we do, Kate?"
"Do?" replied his wife, with a fond smile. "Why, Hubert, you have
already determined what to do!"
Her husband laughed pleasantly.
"I confess that I have. Still, I like to have your desire run with my
own. You want this lad to stay here?"
"Yes, Hubert. If he is lonely and friendless, let us be his friends; for
had he not rescued her, our dear little daughter would have been
killed."
So husband and wife agreed; but when they went to Ralph they
found that he was not quite willing to accept the invitation.
"I know how kind it is of you," the boy said. "And it is true that I
have no friends, and nowhere to go; but I—I cannot live on your
charity. I want to earn my living somehow."
"That is good, Ralph," was the hearty reply of Mr. St. Clive; "but you
must be reasonable. There is such a thing as unreasonable pride.
You cannot earn your living in any calling as a gentleman, without
you are fitted for it. Your life on the plains, and life here, or in
London, would be very vastly different. If you had friends in Texas
we might send you back again——"
"No, no, sir!" cried Ralph, interrupting him. "I could not go back.
Here I must stay for two reasons. I must live to find out what has
become of my father, and I must clear his name from the accusation
that man made."
"Your first reason is good; your second I do not think that you need
worry over. Then you will stay? Well, then, you must certainly let the
wish of my wife and of Irene conquer your pride. I want to help you
all I can; and if presently it is better for you to go, I promise you
that I will not seek to detain you."
"Do stop, Ralph," added Irene, who, pet as she was, had stolen into
her father's study, and heard what was said. "I want you to stay;
and I want you to teach me how to throw a rope like that, though I
should never dare to throw it at a bull. Please stay."
And somehow Ralph looked down into that upturned little face, and
he could not say "no."
"It is very good of you, sir," he murmured, to Mr. St. Clive,
"especially after what Lord Elgert said——"
"My lad, do not be so sensitive concerning that."
"But I cannot help it, sir. He first called my father a thief; and he—he
—you know what he said about your silver?"
And Ralph turned very red.
Mr. St. Clive understood, and sympathized. He liked Ralph all the
better for being keenly sensitive about it.
"There, let it go, my dear boy. Now, once more, business. Have you
any luggage, save these two handbags?"
"In London, sir. Two great trunks. Father left them at the station.
Here are the papers for them." And the boy took a railway luggage
receipt from his pocket-book.
"This is important. We may find something to help us in those
trunks," cried Mr. St. Clive. "Of course, I am not legally justified in
touching them, Ralph; but, under the circumstances, I think that I
might do so. We must have them here, and examine their contents.
We may then discover what brought your father to Stow Ormond;
and that, in its turn, might give us some clue as to what may have
happened."
"I do not think there is much doubt as to what has happened,"
sighed the boy. But Mr. St. Clive would not listen to that.
"Never look at the darkest side, lad. There is a kind Providence over
all, and we must never despair. Now, our very first task must be to
obtain your travelling trunks without delay."
Mr. St. Clive lost no time in putting this resolution into practice. The
trunks were got down from London, and opened; but, to their
disappointment, their contents revealed nothing which tended in any
way to throw a light upon the mystery—clothing, a few mementoes
of their Texan home, and—and in view of Ralph's future welfare this
was most important—banknotes and gold to the amount of £3,000!
"No need to feel yourself dependent upon any one now, Ralph," was
the remark of Mr. St. Clive, as they counted this money; "and no
need to give another thought to Lord Elgert's suspicions. People
possessed of so much money do not go breaking into houses, risking
their liberty for the sake of what they may be able to steal."
Now, though Irene St. Clive was delighted, and would have been
quite content for Ralph to have stayed as her companion, her father
did not look at matters in that way; and he had a serious talk with
Ralph, having first quietly questioned him in order to ascertain his
acquirements.
"You see, Ralph," he said, "what a man needs in England is quite
different from what he may need abroad. You can ride, shoot, and
round up cattle; but that is no good here. Your father has given you
a general education, so that you are not a dunce; but it is nothing
like what you will need as a gentleman here. Knowledge is power
and your desire to clear up the matter of your father's disappearance
demands that you should acquire all the power obtainable. My
advice—I have no right to insist, remember—but my advice is that
you should spend a couple of years at a first-class school—we have
a splendid one here—and if you work honestly during that time, with
your intellect you ought to have made a good headway. What do you
say?"
The boy knit his brows. To one who had passed his days in a wild,
free life, such a prospect did not hold out many charms; but then
Ralph was fond of learning, and had sometimes sighed that he could
not learn more. Besides, his one object in life was to solve the
matter of his father's disappearance, and clear his name from any
foul charge. In his heart, Ralph had resolved ever to live under
honour's flag. He looked up, and answered frankly—
"I will be guided entirely by you, sir, unless my father comes back;
then, of course, I should do whatever he directed."
"My feeling is, that had your father elected to remain in England he
would certainly have sent you to school. Now, Ralph, I am going to
be frank with you. We have, as I have said, a splendid school near
here; but amongst its pupils is Horace Elgert. I fear that he takes
after his father somewhat; and if Lord Elgert has said anything, or
does say anything to him when he knows you are there, young
Horace may try to make it unpleasant for you. Do you understand?"
"Perfectly, sir," replied Ralph.
"And will you go there?"
Ralph looked Mr. St. Clive in the face, and he answered firmly:
"Yes, sir. The boy's being there is nothing to me. I will go."
"Good!" replied Mr. St. Clive, with a nod of appreciation. "We will go
over and see the Headmaster to-morrow."
 CHAPTER IV
RALPH'S FIRST DAY AT SCHOOL

"He is a fine young fellow, but his past life has been spent amidst
very different scenes, and he is far from having a fitting education.
But he is very intellectual and will acquire knowledge quickly. His
father must have been a gentleman, and he has taught his son to be
one also."
It was Mr. St. Clive who spoke, and his words were addressed to Dr.
Beverly, the principal of Marlthorpe College—the best school in all
the county.
A fine-looking man was the doctor, tall, erect, dignified, with firm
face and piercing eyes—eyes which could look terribly severe when
their owner was angry, but which otherwise were gentle, and even
mirthful.
Dr. Beverly was proud of his school, but prouder still of his work. He
did not labour to make scholars only, but also to build up men—
good, noble men—who should be a credit to the old school, and a
blessing to their country. Work or play, the doctor believed in
everything being done as well as it could be, for his watchword was
"Whatever you do, do it to the glory of God," and nothing can be
done to God's glory that is not done as well as it possibly can be.
Mr. St. Clive had explained how Ralph came to be under his care,
and had told the doctor how much he owed to him; and he finished
by mentioning the cruel statement which Lord Elgert had made, and
the angry way in which Ralph had answered it.
"I tell you this," he said, "that you may know everything. I attach no
weight to Elgert's statement myself—it is too absurd, but you must
exercise your own discretion," and the doctor smiled slightly.
"Lord Elgert is rather prone to make rash statements," he said. "I
shall be quite willing to receive your young friend, and I will do my
best to turn him into a good man."
"That I am sure of," was the hearty reply, "and I am also sure that
you will have good material to work upon. Then I will bring Ralph
over."
"And do you propose that he shall board here entirely, or return to
you every Saturday, as most of the lads do?"
"Oh, come home. That is how I did in my day—you know I want to
watch the boy. Good-day, doctor," and Mr. St. Clive came away.
Marlthorpe College was a splendid old building, with large playing
fields at the back, and a great quadrangle in front, to which entrance
was gained through a pair of great iron gates, against which the
porter's lodge was built.
The school itself was at the other side of the quadrangle, directly
facing the gates—a two storey building, with the hall, in which the
whole school assembled upon special occasions, below, and with the
classrooms above. It had two wings; the one to the right being the
doctor's own residence, and that on the left the undermaster's
quarters.
At the back there were again buildings on the right and left—on the
left junior dormitories, the dining-hall, and matron's rooms; and on
the right senior dormitories and studies.
Mr. St. Clive drove home and told Ralph the result of his visit.
"I am sure that you will like the doctor," he said, "and you will find
your companions a nice lot of fellows. Of course there will be some
unpleasant ones; and Ralph, if things are as they used to be, you
will find that there are two sets of fellows—those who mean to work
honestly, and those who never intend to take pains. I need not ask
which set you will belong to," and Mr. St. Clive smiled. "But now," he
added, "I want you to try and be brave. You have a very terrible
sorrow, I know; and it is hard to put it from my mind——"
"It is never from my mind, sir," interrupted Ralph sadly. "I am always
thinking of it."
"But you must not brood over it. To do that, will unfit you for all else.
Leave it with God, Ralph, and do not let even so great a grief
interfere with life's duties. Will you promise me to try and remember
this?"
"I will indeed, sir," answered Ralph. "If I have lost father, I mean to
try and think that he knows, and just do that which would please
him."
"That is good; but still better is it to remember that we have to try
and do that which shall please our Heavenly Father. Now, Ralph, I
suppose that out where you made your home, blows often were the
only way of settling troubles. I do not say that blows are never
justifiable, for sometimes we are placed in such circumstances as
warrant fighting, but do not be too ready to quarrel, or to avenge
every fancied insult with your fist. But there, I am sure that I can
leave that to you. Now come to lunch, and then we must see about
starting."
"I am so glad that you are coming home every week, Ralph," so said
Irene St. Clive, when she heard of the arrangements which her
father had made. "My own lessons are finished on Friday, and we
can have all Saturday to ourselves. I shall count all the days until
each Saturday comes."
So with kindly words to cheer him on his way, Ralph started off with
Mr. St. Clive, and was introduced to Dr. Beverly; and Ralph felt that
he liked the doctor from the very first moment that he saw him; and
he determined that he would do all that he could to get on and
prove to Mr. St. Clive that he meant to keep his word.
Then when his friend had gone, the doctor questioned Ralph to see
just what he knew; and at the conclusion of the examination he laid
his hand on his shoulder.
"My boy," he said, "it is my desire always to have the fullest
confidence in my scholars, and also to enjoy their confidence. I want
you to remember that I desire to be your friend as well as your
master, and that out of school hours I am always glad to see any of
my boys who want to talk with me. I do not mean who want to
come tale-bearing," he added, and Ralph smiled as he answered—
"Thank you, sir. I think I understand."
"You will have to be in the Fourth Form at first, that is the lowest
Form in the Senior House," the doctor continued. "But if you work
well, you will soon be in the Fifth. Now, if you will come with me I
will introduce you to your master, Mr. Delermain, and I think you will
find him ever ready to help you in any way he can."
Ralph thanked the Head again, and followed him, with more of
curiosity than of nervousness, to make the acquaintance of the boys
with whom he was to study; and twenty pairs of eyes glanced up as
the Head opened the door, and then dropped as quickly when they
saw who had entered.
But the master rose from his seat and came forward to meet the
doctor, who said, patting Ralph on the shoulder—
"I have brought you a new scholar, Mr. Delermain. This is Ralph
Rexworth, and he is the young gentleman of whom you have heard
—the one who saved Mr. St. Clive's daughter." Hereat the eyes were
stealthily raised, and glances of something like respectful awe
followed. Of course every one there had heard of the incident about
the bull, and of the disappearance of Mr. Rexworth.
"Rexworth is rather backward," the Head continued. "His life has
been spent abroad, and he has not had the opportunities for study;
but I believe that he will soon pick up." And with this Dr. Beverly
went, and Mr. Delermain, having spoken a few words of welcome,
beckoned to a boy to come forward.
"Warren, let Rexworth sit beside you this afternoon, and give him a
set of the sums we are doing. If you find them too difficult," he
added to Ralph, "do not hesitate to come to me."
But Ralph did not need to ask for aid, he could do the sums and the
exercises that followed. Indeed, he did better than some who had
been there longer, notably one big lad with a sickly flabby face, who
was seated at the bottom of the class, and who received a
reprimand from his master for his indolence.
"It is shameful, Dobson! Here, a new boy has done better than you
have. Your idleness is disgraceful."
A writing exercise followed; and Ralph was bending over his book,
when flop!—a wad of wet blotting-paper hit him in the cheek. He
looked up, but every one seemed busy with their work, so wiping his
cheek he put the wet mass on one side, and went on with his task.
Flop! A second wad came. Ralph noted the direction, and saw that at
the end of the form Dobson was seated, and Ralph had his
suspicions. Pretending to be absorbed in his work, he kept a covert
watch; and presently he was rewarded by seeing Dobson extract a
third wad from his mouth, where he had been chewing it into a
convenient pellet, and under cover of the boy in front of him prepare
to fire it by a flick of his thumb. Ralph raised his eyes and looked
him full in the face, and, somehow, Dobson seemed confused. He
turned red, and bent over his work hastily; and no more pellets were
fired at Ralph that afternoon.
It seemed rather a wearisome afternoon to the boy, used as he was
to his open-air life, but he worked away with all his might; and
presently the bell rang and work was over; and then Warren, the
boy beside whom he had sat, came to him and held out his hand.
"I am first monitor of our form," he said, "and I hope that we shall
be friends. If you come with me I will take you round the school."
"Rexworth."
Ralph turned as his name was called; his master stood there.
"I want you a few minutes. Warren, you can take him round
afterwards. I want to arrange about his study."
"We have only got one vacant, sir," the monitor said. "Charlton has
that."
"I know," was the quiet answer; and then, when Warren ran off, the
master turned to Ralph.
"Rexworth," he said, "I must explain that in our form every two boys
have one study between them, and as you heard Warren say, we
have only one study that is not fully occupied. A lad named Charlton
has it, and you must chum with him. It is about him I want to speak
to you."
"Yes, sir," said Ralph, wondering why his master spoke so gravely.
"Rexworth, I am sorry to say that Charlton is not quite in favour with
his schoolmates. His father got into some trouble and has
disappeared—it is supposed that he is dead—and the boy managed
to gain a scholarship at another and poorer school, and has come
here. He is a real nice lad, but very weakly and timid, and the others
put upon him, partly on that account, partly because of his father's
disappearance, and partly because he is poor—a sad crime in the
eyes of many. It would have been wiser, I think, if he had not come
here, but Dr. Beverly wished him to do so. I wish, Rexworth, that
you would try to be his friend, for he needs one; some of the lads
are nice enough to him, but he seems so very much alone."
"I would like to help him, sir," was the ready answer. And the master
smiled.
"I thought that I was not mistaken in you," he said. "Look, there the
lad is. Charlton, come here."
The lad came up. He was a pale boy, very delicate in appearance,
and with a sad, wistful face.
"Yes, sir," he said.
"Charlton, there is only one vacancy in our studies, and that is with
you. Rexworth will have to chum with you." The boy cast a startled
glance at Ralph. "Take him and show him where it is, and try to
make him feel at home."
"Yes, sir." The boy beckoned to Ralph. "Please come with me," he
said, in troubled tones, as if he doubted whether Ralph would care
about sharing the study with him.
"Have we got to be chums?" asked Ralph; and the other boy
nodded.
"Yes. That is what we call it. It means sharing studies; but you need
not speak to me if you don't want to, and I will not be in the study
much. I am not as it is, for they are always disturbing me and
spoiling my things."
"They! Who?" demanded Ralph; and the lad answered—
"The other chaps and the Fifths. Dobson, in ours, and Elgert of the
Fifth, are the worst. They go in and spoil my things."
"They have no business to, of course?"
"Go in? No, of course not—only the two who chum have any right in
it. Here we are, and—there, they are in now!"—as a scuffling and
burst of laughter came from the inside of the study before which the
boy had halted. "Oh, what are they doing! Will you stop until they
have gone?"
"Not I," answered Ralph grimly. "That study is mine as well as yours,
and I mean to see that we have it to ourselves, Charlton. Come on,
and we will see what is up." And saying this, Ralph threw open the
door and walked into the little room, followed by his companion.
 CHAPTER V
MAKING THINGS STRAIGHT

A burst of laughter greeted Ralph's ears as he opened the study

door, and some one said:
"Look sharp. Here he comes! Hurry up there, Elgert!"
But the laughter died away somewhat awkwardly when the boys saw
that Charlton was not alone, and one or two of the boys came up to
Ralph.
"Hallo, you new fellow! They surely haven't put you to chum with
Charlton, have they? What a shame! I should kick against it. Some
one else must make room for you."
Such were the remarks of those who had taken a fancy to Ralph, but
he paid no heed to it all. He just calmly gazed round, as if counting
the number of boys there and taking their measure; and then he
quite as calmly shut the door, locked it, and put the key in his
pocket. Those present looked in surprise for a moment—some
laughed, and one, a tall, handsome boy, came haughtily up to him.
"What do you mean by that?" he demanded. "How dare you lock
that door?"
Ralph regarded him with the utmost coolness. No one had told him
who the boy was, and yet he seemed to know—he felt sure that this
was none other than Horace Elgert himself.
"Wait a bit," he said calmly. "So far as I understand, this study
belongs to Charlton and myself. We have a perfect right to lock the
door."
"But not to lock us in," retorted Elgert. "Open it at once, and think
yourself lucky that you don't get a licking for your impudence!"
"Steady!" was Ralph's answer. "It seems to me that if you had not
been where you have no right to be, you would not have got locked
in; and now that you are here, you must wait my pleasure as to
going out."
This was beginning school life with a vengeance, but Ralph believed
in settling things once and for all, and his indignation was hot as he
saw what these half dozen lads had been doing.
But Horace Elgert was not a boy to be spoken to like that, and he
came striding up to Ralph to take the key by force.
"I will soon settle you," he began, and he aimed a blow at this
impertinent new boy's head, only somehow the blow did not get
there. Ralph adroitly stepped aside, and the Honourable Horace
Elgert stumbled to the ground violently.
"A fight! A fight!" cried the rest; but Ralph smiled and shook his
head.
"Oh, no, my friends. I have something better to do, and this is not
the place for fighting."
They were staggered. They could not understand this coolness and,
moreover, they had all heard about Ralph having tackled the bull,
and the story had grown somewhat. They stood considerably in awe
of this boy from the Western plains, and they began to wish that
they were anywhere else than in his study.
Horace Elgert got up, his face white with passion but he made no
more attempts to take the key from Ralph.
"You are right," he said, in suppressed tones; "this is not the place
to fight. Open the door, and we will soon settle things."
"Presently," was all the answer he got. "Now, then, let us see what
you have been up to."
He glanced round at the books tumbled on the floor, at a desk upset,
at an ink-bottle on its side, and then turned to his chum.
But Charlton was standing, looking very white, and staring at a
picture on the wall—the picture of a lady, and beneath it some one
had written—
"This is Charlton's mammy. But where is his daddy? Puzzle—Find
daddy, and tell the police."
Ralph felt his nerves tingle. He felt sure that Elgert had done that,
and he remembered the words of Lord Elgert respecting his own
father.
"Who did that?" he said, and no one answered. He went up to
Elgert. "Did you do it?"
"Well, if I did, what is it to do with you? Mind your own business!"
"Take that scrawl down. Quick, or I shall lose my temper, and then I
fancy some one will get hurt! Down with it! That is right"—as the
other, considerably startled, pulled the writing down. "Give it to me."
It was remarkable how the daring of the one lad held the half dozen
in check. Elgert handed him the paper, and Ralph tore it up and
threw the fragments into his face.
"Now then, you have upset this room. Just put it straight again, and
look sharp about it!" he said. "And please to understand that
Charlton and I are chums, and mean to stick together. Oh, and I
want a word with you"—and he walked up to Dobson, who turned a
trifle more pasty-looking than before. "Do you know what these
are?"
Ralph produced two wads of chewed blotting-paper from his pocket
as he spoke, and Dobson blustered—
"You keep to your chum, since you are so thick with him. I don't
want anything to do with you. I say, you chaps, are you going to let
him crow over you like this? Rush him!"
"Good advice; only, why don't you do the rushing first?" said Ralph.
"I asked you if you recognized these. If you don't, I will tell you what
they are—they are pieces of blotting-paper, which you chewed and
then threw at me. They came out of your mouth, and they are going
back there again—when I have mopped up this ink which you have
spilt." Ralph suited the action to the word, and presented the two
unpalatable-looking objects to Dobson, who was at once a coward
and a bully. "Now, then, open your mouth!"
"I won't! Who do you think that you are? I—— Oh!"
For Ralph did not argue. He grabbed hold of Dobson, and with a
quick jerk sent him backwards across the little study table.
"Oh, oh! You are breaking my back!" howled the bully.
"Open your mouth!"
"I won't! Oh, help me, you fellows—he will break my back! Oh! Ugh!
Ow! I am choking!" For, just as he opened his mouth to yell, Ralph
had pushed both those pieces of blotting-paper in.
"Now, then, take them," he said. "Quick, or it will be the worse for
you!"
Dobson, with many queer grimaces, had to comply—it was the most
unsavoury morsel which he had tasted for many a day.
 "Dobson, with many queer grimaces, had to comply." p. 49
"Now! Ah, I see that you have straightened things!" Ralph went on.
"Now you chaps can go, and the next time you want to come into
our study take my advice and ask leave, or there will be more
trouble. Clear out!"
And he unlocked the door and flung it open.
And out those half dozen boys went, looking considerably crestfallen
and stupid, and knowing also that they were cowards—they were all
frightened by Ralph, so greatly does one of dauntless bearing affect
a number.
But one boy turned, and that one was Horace Elgert, and he came
back and gave Ralph look for look.
"Look here, you new fellow!" he said, "you have been very clever,
but you have done a bad day's work for yourself. You have made
one enemy at least. As for that insult which you offered me, you will
have to fight me for it; and as for you, you miserable cub"—and he
turned towards Charlton, who cowered back before his raised fist
—"as for you, I will——"
"Hold hard—you will do nothing!" answered Ralph, with the utmost
good humour. "You are talking tall, that is all about it. Now, take my
advice, and go; and when you are calmer, you will see things
differently. And then, as to fighting—well, I shall not run away in the
meantime. Clear!"
And with that he shut the door and locked it behind his discomfited
foes. Then, seating himself, he looked at the bewildered Charlton,
and laughed again as he saw the look of admiration in his face.
"There, I think that has taught them a lesson! We shall not have
them upset our study again," he said. "One must maintain one's
rights, and we may as well begin as we mean to go on. So this is our
study, is it?"
"Yes, if you will share it with me," the other boy said. And Ralph
answered—
"Share it? Of course I shall share it with you! Did not you hear Mr.
Delermain say that we were to share it?"
"But most fellows don't like me, because—because——"
"Never mind why," interrupted Ralph, anxious to spare the boy's
feelings. "I heard something about your father being gone; well, my
father is gone, you know"—and Ralph's voice shook a little—"and so
we two ought to be chums, and help each other. Then, I suppose
that you know more than I do; for, except at roping a steer or
rounding up a herd of cattle, I am afraid that I am not of much use.
You will be able to help me on no end."
"What! I help you?" gasped Charlton. "How can I do that?"
"You know Greek and Latin, and goodness knows how much more,
that I am only just at the beginning of, and you will be able to give
me a hand with it. I want to get on and pick up things as quickly as
I can."
"I might help you that way, if you would let me," the boy said
doubtfully. And Ralph laughed.
"What a chap you are! Have I not told you that I shall be downright
thankful: and there you keep on about if I will let you. Come, shake
hands upon it! Charlton, we two are chums, and we are going to
stick together and help each other. Is that so?"
"Yes, if you will. I shall be so glad to have a chum, because it has
been rather lonely sometimes; and then, you see, I am not very
strong, and I am not brave like you, and the fellows know it, and
they try to play all sorts of tricks upon me. Do you really mean to be
my chum, Rexworth?"
"Really and truly! Now, let us go down, and then you can show me
what the place is like," was Ralph's answer. And the two, descending
to the playground were met by Warren, who stopped and looked
from Ralph to Charlton, and then asked—
"I say, Rexworth, what have you been up to so soon? There is
Dobson declaring that he will do all manner of things to you. You
seem to have been having some fun already."
So Ralph explained what had happened, and the monitor laughed
until the tears ran down his cheeks.
"Well, all I can say is that you are a cool hand," was his comment,
"and I am not sorry that you have taught Dobson a lesson. You have
not much to fear from him, but you will find that Elgert, for all he is
an Honourable, has precious little honour about him. He will pay you
back if he gets the chance, be sure of that. However," he went on, "I
am glad that you two are chums, for I think you will like each other;
but there is the bell for tea. Come on, or we shall be late."
The rest of that day passed without further incident and at last the
boys—evening preparation and supper over—went trooping to their
dormitories, there to laugh and chat as they undressed; and many
glances were bestowed upon Ralph. His exploit of that afternoon
had been spoken of, and there was no attempt to play any jokes
upon one who was prepared to take his own part so vigorously.
But presently the laughing suddenly stopped, and something like a
hush of surprise succeeded the noise. Warren seated on the edge of
his bed, looked round to see what had happened—he thought that
one of the masters had come in unexpectedly; but he saw his
companions standing glancing across towards the spot where
Ralph's bed was, and he, following their gaze, saw that the boy who
was ready to face half a dozen of his companions, was down on his
knees, his head bent upon his hands in prayer.
Warren felt a thrill of shame. He was a real good lad at heart, but
somehow he did not do that—none of them did—they thought that
public prayers were enough; and yet he had promised his mother
that each night he would kneel alone in prayer.
Some of the boys were tittering, some looked grave. Warren
suddenly found himself resolved. "If a thing should be done, do it at
once," was his motto. He gave one hasty glance round, half
ashamed, half defiant, and then, in the sight of all his companions,
the Fourth Form monitor also knelt down by his bed, following the
brave example set by Ralph Rexworth.
 CHAPTER VI
AN EARLY MORNING SPIN

It was quite a common thing for new boys at Marlthorpe College to

be made the victims of practical jokes during their first night in the
school; but such was the impression which Ralph Rexworth had
made, that no tricks were attempted with him. A boy who could take
his own part so vigorously was not the sort that it was safe to take
liberties with.
Nor was that the only reason. With Dobson and his friends it was
quite sufficient, but with the better boys, that quiet kneeling down to
pray had not been without effect. Some of them recognized that to
do that might require more courage than to deal as he had done
with those who had invaded his study—a moral courage, far greater
and better than a physical; and they realized that a boy who
possessed that courage was not a fit subject for stupid jokes.
So Ralph slept peacefully until the morning, when, used to early
rising all his life, he opened his eyes before any of the other boys
were awake.
At first he felt puzzled with his surroundings, but he soon
remembered; and propping himself upon his elbow he lay watching
the faces of the others, wondering what sort of lads they would
prove to be, and how he should get on with them, and whether he
would be able to master the lessons which they were engaged upon.
Then he looked at Charlton, and thought how sad he looked, even in
his sleep; and he noted how often he sighed. Perhaps he was
dreaming of his father.
That sent him thinking of his own father, and the mystery of his fate;
and he pondered whether it would ever be possible for him—a lonely
boy in this strange land—to find out the truth concerning his parent's
disappearance. But he was not altogether alone; it was wrong to
think of himself in that light. God had given him a friend in Mr. St.
Clive, and another in Mrs. St. Clive, and yet a third—a very nice,
lovable third—in Irene! Ralph, who had never had anything to do
with girls, thought Irene the sweetest, dearest little friend that it
would be possible to find.
A bell rang, and his companions stretched and yawned and opened
their eyes; and though some grunted and turned over again,
determined to have every minute they could, several jumped up at
once, and hastily pulling on their clothes began sluicing and
splashing in good, honest, cold water.
"Hallo! Awake? Slept well?" queried Warren seeing that Ralph was
preparing to follow the example of these last boys. "Any one try any
games with you in the night?" And he came and sat down on Ralph's
bed, and grinned when the new boy answered that he had not been
disturbed.
"I suppose they thought better of it. That is your basin!" he added,
pointing to one washstand. "Mind that they don't take all the water,
or you will either have to sneak another fellow's, or go and get some
more for yourself. Look sharp, and then we will go and have a turn
with the bells, and a spin afterwards, I like to get all I can before
breakfast; it seems to set a fellow up for the day."
Ralph nodded, and began vigorously sluicing and polishing; and the
boys, too busy about their own business, paid no attention to him.
He was quite capable of looking after himself, in their opinion. At
last, all ready to accompany the monitor, he quietly repeated his
action of the previous night—he knelt down in prayer.
That staggered even Warren. As a whole, the boys were good lads,
but even those who had been accustomed to evening prayers in
their homes did not seem to think that morning prayers were quite
as important. They wanted to scramble off to play as quickly as
possible. The Head always read prayers in school, and that was
enough; and here was this new fellow wasting precious time in this
way!
A few sneered and giggled; some shrugged their shoulders, and ran
off; some looked grave; and Warren sat nursing his foot, and
pondering; while Charlton turned red.
But they made no remarks; and when Ralph rose from his knees, the
three went out together. Warren was turning over a decidedly new
leaf. If he had not annoyed Charlton before, he had left him pretty
much alone, and now he was admitting him to his company. Well,
Charlton was Rexworth's chum, and if he wanted Rexworth he must
have the chum as well.
Charlton hardly expected the monitor to be friendly to him, but he
waited for his chum, and Warren waited, too.
"Let us get down and have a try at the bells," suggested the
monitor, leading the way. And Ralph inquired innocently—
"Ringing bells, do you mean?"
Whereat Warren stared, and felt just a little less respect for the new
boy. What sort of a fellow could he be if he didn't know what dumb-
bells were?
"Ringing bells?" he repeated. "No; dumb-bells—exercises, you know!
Come on, I will show you."
"I never saw bells like those," was Ralph's comment, when a pair
was produced. "How do you use them?"
Warren went through a set of exercises, and then handed them to
Ralph, who laughed, and said—
"Why, they don't weigh anything! I don't see much exercise in this!"
"They are six-pounders," was the answer; "quite as heavy as you will
want. Now try this exercise—do it a dozen times."
Warren showed Ralph the right way, and off he went; Charlton, who
had also got a pair of bells, doing the same. And, to Ralph
Rexworth's surprise, he found that those weights at which he had
laughed soon made him feel tired, and that Charlton could keep on
longer than he could. He could not understand that.
"I don't see why it should be," he said.
And a voice replied—
"Because you are exercising muscles which you have not tried much
before, my lad." And he turned, to see Mr. Delermain watching him.
"Try again," said the master. "Only once; this sort of thing must be
done gradually. Go slow, and take time."
Ralph obeyed: but dumb-bells certainly made his arms ache. And
then Warren suggested Indian clubs.
"Indian clubs," repeated Ralph, "and what are they? I never saw the
Indians use clubs. They have knives and hatchets, and spears and
bows, and some of them use guns, too, and shoot wonderfully well;
but I never saw them use clubs."
Now that speech caused a smile, but it was a very respectful smile;
for here was a boy who had actually seen real Indians. That was
something, even if he did not know what Indian clubs were!
However, the clubs were produced, and Ralph was shown how to
swing them. And, as a natural result of his first attempt, he hit his
head a smart crack, evoking a burst of laughter thereby.
"Slow and steady," he answered; "I shall get it in time. I don't
understand these things; but if you get me a coil of rope, I will show
you one or two little things that I do not think any of you can do."
"A coil of rope—that is easily supplied," said Mr. Delermain; and
when it was brought, he said: "Now, Rexworth, let us see what you
can do." And all the boys stood round while Ralph took the rope and
made a running noose at one end.
"Give me plenty of room," he said, and he commenced to whirl the
noose round and round his head, letting the rope run out as he did
so; until at last he held the very end in his hand, and the rest was
twirling round and round him in a perfect circle.
"One of you try to do that," he said.
And try they did, in vain. They could not even get it to go in a circle,
and it made their arms ache dreadfully.
Then he made the circle spin round him on its edge just as if that
rope was a hoop; and afterwards he actually jumped through it as it
was going, explaining that the cowboys on the ranches frequently
indulged in such tricks as these, and were experts at it—far more so
than the Indians themselves.
Then nothing would do but that he must show them how a lasso
was thrown. And though several, including the master, essayed to
try, not one of them was able to send the noose over Ralph's
shoulders, though he caught them, one after the other, without the
slightest trouble.
"It is what one is used to," he said laughing. "I have not had much
to do with bells and clubs—nothing to do with them, indeed—but I
have played with a rope all my life."
Dobson had come in with his friends, and he stood and glared.
Elgert came in, and looked angry. This new boy was evidently on the
way to become a favourite in the school, and, unless something was
done, he might rival them. Though just then they did not speak to
each other about it, both Dobson and Elgert arrived at the same
conclusion—namely, that something should be done, and that Ralph
Rexworth should be humbled and disgraced.
Then Warren suggested a spin, and of course Charlton went, and
two or three other boys—who found Ralph very good company—had
Welcome to Our Bookstore - The Ultimate Destination for Book Lovers
Are you passionate about books and eager to explore new worlds of
knowledge? At our website, we offer a vast collection of books that
cater to every interest and age group. From classic literature to
specialized publications, self-help books, and children’s stories, we
have it all! Each book is a gateway to new adventures, helping you
expand your knowledge and nourish your soul
Experience Convenient and Enjoyable Book Shopping Our website is more
than just an online bookstore—it’s a bridge connecting readers to the
timeless values of culture and wisdom. With a sleek and user-friendly
interface and a smart search system, you can find your favorite books
quickly and easily. Enjoy special promotions, fast home delivery, and
a seamless shopping experience that saves you time and enhances your
love for reading.
Let us accompany you on the journey of exploring knowledge and
personal growth!

ebookgate.com