Scribd
Upload
57 views

Segmentation of Anatomical Structures: Frithjof Kruggel, M.D

Lecture 05: Segmentation of Anatomical Structures BME 234: Neuroimaging Data Analysis, Spring 2016

Uploaded by

popstick2000
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
57 views

Segmentation of Anatomical Structures: Frithjof Kruggel, M.D

Lecture 05: Segmentation of Anatomical Structures BME 234: Neuroimaging Data Analysis, Spring 2016

Uploaded by

popstick2000
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 19

Lecture 05:

Segmentation of Anatomical Structures


BME 234: Neuroimaging Data Analysis, Spring 2016, Course Code 14310

Frithjof Kruggel, M.D.


Department of Biomedical Engineering, University of California, Irvine
Office: REC 204, Phone: 4-3729, Email: [email protected]

First Prev Next Last Go Back Full Screen Close Quit

Review: Why process anatomical data?


Remember several reasons for analyzing anatomical data:
1. Morphometry: studying macroscopic properties of structures in groups of (healthy) subjects,
2. Structure-function relation: comparing structural properties with other variables (from
physiology, cognitive neuroscience, biochemistry, genetics),
3. Anatomical Reference: relating functional activation to an anatomical locus,
4. Describing pathology: studying macroscopic tissue changes under the condition of CNS
diseases (e.g., focal or diffuse lesions, regional or global atrophy, developmental disorders),
5. Visualization: mapping physical, physiological or functional quantities on structures for
displaying results.
6. Constraining the solution space: restricting the search space in signal detection to specific
sub-compartments.
Items 1, 2, 4, 5, 6 require a segmentation of brain structures under interest. Items 1, 2, 3
require atlas mapping methods. Both will be discussed in more detail now.

First Prev Next Last Go Back Full Screen Close Quit

Brain segmentation: Overview


Segmentation means to split an image into disjunct regions. For neuroimaging datasets of the head, such regions are neuroanatomically defined regionsof-interest, i.e., extracting the brain, the temporal lobe or the thalamus.
There major strategies are available for segmentation:
1. Discrimination of a structure due to some unique properties (e.g., intensity, size, location,...). Here, methods operate in the individual space of
a dataset.
2. Mapping the individual dataset onto a reference atlas and segmenting
(or measuring) structures in the normalized space of the atlas.
3. Mapping a reference atlas onto an individual dataset and segmenting (or
measuring) structures in the individual space.
Note that strategies 2 and 3 are not equivalent and will yield different results.

First Prev Next Last Go Back Full Screen Close Quit

Atlas-based segmentation I
Atlas-based approaches require an annotated reference atlas and an algorithm for (nonlinear)
registration of the individual dataset with the reference.
The following annotated volumes are available:
The ICBM Single Subject MRI Anatomical Template
(see:
http://www.loni.ucla.edu/ICBM/Downloads/Downloads ICBMtemplate.shtml).
This is a labeled dataset of a single subject in stereotaxic space.
The ICBM 452 T1 Atlas
(see: http://www.loni.ucla.edu/ICBM/Downloads/Downloads 452T1.shtml). This atlas
is an average of T1-weighted MRIs of normal young adult brains. It does not contain
annotations.
The LPBA40 Atlas
(see: http://www.loni.ucla.edu/Atlases/Atlas Detail.jsp?atlas id=12). This is a probabilistic atlas built from manual annotations of 54 structures in 40 datasets. This atlas
contains the per-voxel chance by which a specific structure is found at that location.
The Voxel-Man Atlas
(see: http://www.voxel-man.de/3d-navigator/brain and skull/) contains an annotated
brain from the Visible Human Project.
Refer to a closer explanation about the difference of Talairach coordinates and the different
reference brains to: http://imaging.mrc-cbu.cam.ac.uk/imaging/MniTalairach.
First Prev Next Last Go Back Full Screen Close Quit

Atlas example

First Prev Next Last Go Back Full Screen Close Quit

Atlas-based segmentation II
Advantages and disadvantages of atlas-based segmentation:
Conceptually simple: segmentation is performed by registration with a
reference.
Relies on the correctness of the annotation and the quality of registration.
Improvements may be achieved if multiple atlases are used, and labels
are fused for a specific object.
Example-based atlases do not reflect the variance in the (healthy) population.
Probabilistic atlases contain per-voxel probabilities, and will thus lead
to an imprecise segmentation.
Due to the variability of the human brain, results are better for the core
structures (e.g., basal ganglia), and worse for the cortex.
Approaches are not stable in the presence of lesions.
For certain structures that are (currently) not discriminatable by MR
imaging (e.g., subthalamic nuclei) atlas-based registration is the only
method for localization.
Note that atlases and their use for brain segmentation will be discussed in
First Prev Next Last Go Back Full Screen Close Quit

Segmentation of individual datasets


Developing methods for segmenting individual datasets is one of the major
topics in biomedical imaging. Structures are segmented based on unique
characteristics (e.g., location, intensity, texture, shape) and recognized by
comparison with generic (symbolic) models.
The advantage of this approach is that the individual space is retained. Methods are generally stable against the presence of lesions, or were developed
to segment lesions.
One of the most commonly applied processing chains is used to extract the
brain from MRI data of the human head, often called brain shelling or
peeling. An example procedure is discussed in the following.

First Prev Next Last Go Back Full Screen Close Quit

Brain peeling I
After inhomogeneity correction, the image is segmented into three intensity
classes. Voxels of the third class are selected, and the largest connected
component is selected as a raw white matter segmentation.

First Prev Next Last Go Back Full Screen Close Quit

Brain peeling II
Next, a raw surface of this segment is computed, and optimized to the
WM/GM interface using the intensity-corrected dataset. The BG/GM interface can be computed by expanding this interface. Finally, this surface
can be used to extract the brain from the imaging data.

First Prev Next Last Go Back Full Screen Close Quit

Cortical thickness
The local distance between the WM/GM and BG/GM interfaces can be interpreted as the neocortical thickness. Values can be integrated over a surface
patch as the neocortical volume.

First Prev Next Last Go Back Full Screen Close Quit

Neocortical parcellation
The neocortical surface can further be subdivided into sulci and gyri, and
basins are identified as substructures of individual sulci.

Note that the individual variability is retained in these results. The segmentation of individual structures is a topic in Seminars 2 and 4.

First Prev Next Last Go Back Full Screen Close Quit

Segmentation of focal lesions


A focal lesion is a circumscribed region of damaged tissue in the body.
Restrictring the discussion on the brain, focal lesions are a consequence of
head trauma, cerebral infarcts or intracerebral hemorrhages.
In a wider sense, an increase of tissue due to a pathological process (e.g., a
tumor, a cyst) may also be denoted as a lesion.

First Prev Next Last Go Back Full Screen Close Quit

Properties of focal brain lesions


A lesion:
may affect different adjacent tissue types,
has a more or less well defined border,
undergoes changes with time (due to restorative processes or continued
action of a pathological process),
induces a loss of function,
is detectable by a neuroimaging method (e.g., MRI, CCT, SPECT, PET).

First Prev Next Last Go Back Full Screen Close Quit

Approaches for lesion detection


Lesion segmentation may be based on:
manual delineation of lesion borders by an expert
local intensity changes
( classification-based methods, boundary-based methods),
local changes in image texture,
a geometrical or physical model
( pixon-based methods, deformable models),
comparison with an anatomical atlas.
The sensitivity and specificity of lesion detection is improved if multichannel information is included.

First Prev Next Last Go Back Full Screen Close Quit

Problems of lesion detection


Lesion detection is complicated by the following facts:
A lesion must be discriminated from different surrounding compartments, some of which may have intensity distributions similar to the
lesion.
Lesions are generally not homogeneous, often with completely damaged
core parts and minor damage in peripheral portions.
In some parts, the lesion border may be sharp, while in others it is faint.
Multichannel images typically do not have the same voxel dimensions
and patient position (i.e., registration and interpolation is required).

First Prev Next Last Go Back Full Screen Close Quit

Segmentation approach
Approximation: the brain is a mirror-symmetric organ.
Lesions considered here are confined to a single hemisphere.
Tissue on the contralateral side is generally healthy.
Find compact areas with an intensity statistic that differs significantly
from the contralateral side.

First Prev Next Last Go Back Full Screen Close Quit

Algorithm summary
The process chain consists of the following steps:
affine registration of T1- and T2-weighted MRI data,
comparison of the signal statistics in two subregions that are mirrorsymmetric w.r.t. the mid-sagittal plane,
thresholding of the statistical map and determination of connected components,
comparison of the component size with the size distribution of natural
asymmetries,
characterization of detected lesions.
Natural asymmetries are expected to be smaller than pathological lesions.

First Prev Next Last Go Back Full Screen Close Quit

Lesion segmentation: Example


Using the symmetry criterion offers the advantage that the true lesion
size is more correctly estimated in comparison with a simple intensity- or
boundary-based algorithm.

First Prev Next Last Go Back Full Screen Close Quit

Lesion description
The detected lesion may further be characterized using an intensity- or
texture-based segmentation. Besides parameters such as size, position and
compactness the severity of tissue damage may be quantified by an index
and followed up in time-series evaluation.

First Prev Next Last Go Back Full Screen Close Quit

You might also like