...

Stroke Lesion Segmentation for tDCS Elin Naeslund LiTH-IMT/MI30-A-EX--11/502--SE Linköping 2011

by user

on
1

views

Report

Comments

Transcript

Stroke Lesion Segmentation for tDCS Elin Naeslund LiTH-IMT/MI30-A-EX--11/502--SE Linköping 2011
Stroke Lesion Segmentation for tDCS
Elin Naeslund
LiTH-IMT/MI30-A-EX--11/502--SE
Linköping 2011
Stroke Lesion Segmentation for tDCS
Examensarbete utfört i medicinsk informatik
vid Tekniska högskolan i Linköping
av
Elin Naeslund
LiTH-IMT/MI30-A-EX--11/502--SE
Handledare:
Mats Andersson
IMT, Linköpings universitet
Lucas C Parra
BME, The City College of New York
Examinator:
Hans Knutsson
IMT, Linköpings universitet
Linköping, 7 June, 2011
Avdelning, Institution
Division, Department
Datum
Date
Medical Informatics
Department of Biomedical Engineering
Linköpings universitet
SE-581 85 Linköping, Sweden
Språk
Language
Rapporttyp
Report category
ISBN
Svenska/Swedish
Licentiatavhandling
ISRN
Engelska/English
Examensarbete
C-uppsats
D-uppsats
Övrig rapport
2011-06-07
—
LiTH-IMT/MI30-A-EX--11/502--SE
Serietitel och serienummer ISSN
Title of series, numbering
—
URL för elektronisk version
xx
xx
Titel
Title
Segmentering av strokelesion för tDCS
Stroke Lesion Segmentation for tDCS
Författare Elin Naeslund
Author
Sammanfattning
Abstract
Transcranial direct current stimulation (tDCS), together with speech therapy, is
known to relieve the symptoms of aphasia. Knowledge about amount of current
to apply and stimulation location is needed to ensure the best result possible.
Segmented tissues are used in a finite element method (FEM) simulation and by
creating a mesh, information to guide the stimulation is gained. Thus, correct
segmentation is crucial. Manual segmentation is known to produce the most accurate result, although it is not useful in the clinical setting since it currently takes
weeks to manually segment one image volume. Automatic segmentation is faster,
although both acute stroke lesions and nectrotic stroke lesions are known to cause
problems.
Three automatic segmentation routines are evaluated using default settings and
two sets of tissue probability maps (TPMs). Two sets of stroke patients are used;
one set with acute stroke lesions (which can only be seen as a change in image intensity) and one set with necrotic stroke lesions (which are cleared out and filled with
cerebrospinal fluid (CSF)). The original segmentation routine in SPM8 does not
produce correct segmentation result having problems with lesion and paralesional
areas. Mohamed Seghier’s ALI, an automatic segmentation routine developed to
handle lesions as an own tissue class, does not produce satisfactory result. The
new segmentation routine in SPM8 produces the best results, especially if Chris
Rorden’s (professor at The Georgia Institute of Technology) improved TPMs are
used. Unfortunately, the layer of CSF is not continuous. The segmentation result
can still be used in a FEM simulation, although the result from the simulatation
will not be ideal.
Neither of the automatic segmentation routines evaluated produce an acceptable result (see Figure 5.7) for stroke patients. Necrotic stroke lesions does not
affect the segmentation result as much as the acute dito, especially if there is only
a small amount of scar tissue present at the lesion site. The new segmentation routine in SPM8 has the brightest future, although changes need to be made to ensure
anatomically correct segmentation results. Post-processing algorithms, relying on
morphological prior constraints, can improve the segmentation result further.
Nyckelord
Keywords
automatic segmentation, manual segmentation, stroke lesions, tDCS, SPM8, aphasia patients
Abstract
Transcranial direct current stimulation (tDCS), together with speech therapy, is
known to relieve the symptoms of aphasia. Knowledge about amount of current
to apply and stimulation location is needed to ensure the best result possible.
Segmented tissues are used in a finite element method (FEM) simulation and by
creating a mesh, information to guide the stimulation is gained. Thus, correct
segmentation is crucial. Manual segmentation is known to produce the most accurate result, although it is not useful in the clinical setting since it currently takes
weeks to manually segment one image volume. Automatic segmentation is faster,
although both acute stroke lesions and nectrotic stroke lesions are known to cause
problems.
Three automatic segmentation routines are evaluated using default settings and
two sets of tissue probability maps (TPMs). Two sets of stroke patients are used;
one set with acute stroke lesions (which can only be seen as a change in image intensity) and one set with necrotic stroke lesions (which are cleared out and filled with
cerebrospinal fluid (CSF)). The original segmentation routine in SPM8 does not
produce correct segmentation result having problems with lesion and paralesional
areas. Mohamed Seghier’s ALI, an automatic segmentation routine developed to
handle lesions as an own tissue class, does not produce satisfactory result. The
new segmentation routine in SPM8 produces the best results, especially if Chris
Rorden’s (professor at The Georgia Institute of Technology) improved TPMs are
used. Unfortunately, the layer of CSF is not continuous. The segmentation result
can still be used in a FEM simulation, although the result from the simulatation
will not be ideal.
Neither of the automatic segmentation routines evaluated produce an acceptable result (see Figure 5.7) for stroke patients. Necrotic stroke lesions does not
affect the segmentation result as much as the acute dito, especially if there is only
a small amount of scar tissue present at the lesion site. The new segmentation routine in SPM8 has the brightest future, although changes need to be made to ensure
anatomically correct segmentation results. Post-processing algorithms, relying on
morphological prior constraints, can improve the segmentation result further.
v
Acknowledgments
I would like to thank my examiner Hans Knutson for his help and input during
this thesis work, for letting me follow my dream and sometimes pulling me back
to earth. I would also like to thank my supervisors, Mats Andersson at IMT and
Lucas C Parra at The City College of New York, for answering my questions and
pointing me in the right direction when needed.
I would also like to thank the people in the Neural Engineering group at the BME
department at The City College of New York; Abhi and Davide for answering my
sometimes silly questions about Matlab and LaTeX, Asif for the discussions we
have had, and finally thank you Johnson, Belen, and Christoph for all the laughs
we shared. Thank you for making my months at The City College of New York
the best possible. I will miss every single one of you.
Finally, I would like to thank my family for always believing in me. Thank you for
all your love and support over the years, in school and in life in general. A thanks
also goes out to my friends, for making the years at Linköping University the best
possible. I love you.
Elin Naeslund
New York, May 2011
vii
Contents
1 Introduction
1.1 Problem Formulation . . .
1.2 Related Research . . . . .
1.3 Purpose and Goals . . . .
1.4 Organization of the Thesis
. . . . .
. . . . .
. . . . .
Report
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
5
5
6
6
7
2 Theory
2.1 The Human Brain . . . . . . . . .
2.1.1 A Healthy Brain . . . . . .
2.1.2 Strokes . . . . . . . . . . .
2.1.3 Stroke Lesions . . . . . . .
2.2 Stimulation of the Brain . . . . . .
2.2.1 Transcranial Direct Current
2.3 Segmentation . . . . . . . . . . . .
2.3.1 Manual Segmentation . . .
2.3.2 Automatic Segmentation . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
Stimulation
. . . . . . .
. . . . . . .
. . . . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
(tDCS)
. . . . .
. . . . .
. . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
9
9
11
11
12
13
15
15
16
16
3 Previous Work
3.1 SPM8 Original Segmentation
3.1.1 Settings . . . . . . . .
3.1.2 Problems . . . . . . .
3.2 Mohamed Seghier’s ALI . . .
3.2.1 Settings . . . . . . . .
3.2.2 Problems . . . . . . .
3.3 SPM8 New Segmentation . .
3.3.1 Settings . . . . . . . .
3.3.2 Problems . . . . . . .
3.4 Chris Rorden’s TPMs . . . .
3.5 Manual Segmentation . . . .
3.5.1 Problems . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
17
17
19
19
21
22
22
23
23
23
26
28
28
4 Method
4.1 Patients . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.1.1 Patients with Necrotic Stroke Lesions . . . . . . . . . . . .
4.1.2 Patients with Acute Stroke Lesions . . . . . . . . . . . . . .
29
30
30
32
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
ix
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
x
Contents
4.2
4.3
Automatic Segmentation vs. Ground Truth . . . . . . . . . . . . .
Regions of Interest (ROIs) . . . . . . . . . . . . . . . . . . . . . . .
33
34
5 Results
5.1 Automatic Segmentation vs. Ground Truth . . . . . . . . . . . . .
5.2 Regions of Interest (ROIs) . . . . . . . . . . . . . . . . . . . . . . .
37
37
45
6 Discussion
6.1 Interpretation of the Results . . . . . . . . .
6.1.1 Automatic Segmentation vs. Ground
6.1.2 Regions of Interest (ROIs) . . . . . .
6.2 Conclusions . . . . . . . . . . . . . . . . . .
6.2.1 SPM8 Original Segmentation . . . .
6.2.2 SPM8 New Segmentation . . . . . .
6.2.3 SPM8 Chris Rorden’s TPMs . . . .
6.2.4 Mohamed Seghier’s ALI . . . . . . .
6.3 Future Improvements . . . . . . . . . . . . .
6.3.1 SPM8 Original Segmentation . . . .
6.3.2 SPM8 New Segmentation . . . . . .
6.3.3 Mohamed Seghier’s ALI . . . . . . .
6.3.4 Manual Segmentation . . . . . . . .
49
49
49
52
56
56
56
57
57
57
57
58
58
58
Bibliography
. . . .
Truth
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
59
List of Figures
2.1
2.2
2.3
2.4
2.5
Brain regions . . . . . . . . . . . . . . . .
Brain lobes . . . . . . . . . . . . . . . . .
MRI weighting - T1 and T2 . . . . . . . .
Stroke lesion types . . . . . . . . . . . . .
Electrode configuration and corresponding
.
.
.
.
.
10
11
12
13
14
3.1
3.2
3.3
3.4
3.5
3.6
3.7
Illustration of TPMs - SPM8 original segmentation . . . . . . . . .
Segmentation example - SPM8 original segmentation and ALI . . .
Illustration of TPMs - ALI’s extra class . . . . . . . . . . . . . . .
Illustration of TPMs - SPM8 new segmentation . . . . . . . . . . .
Segmentation example - SPM8 new segmentation (default TPMs) .
Illustration of TPMs - Chris Rorden’s . . . . . . . . . . . . . . . .
Segmentation example - SPM8 new segmentation (Chris Rorden’s
TPMs) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
18
19
21
23
25
26
AC, PC, and AC-PC line . . . . . . . . . . . . . . . . . . . . . . .
Patients with necrotic stroke lesions (Keith, LeftMca, and Smalltemporal) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Patients with necrotic stroke lesions (Ela) . . . . . . . . . . . . . .
Patients with acute stroke lesions (3316, 3319, and 3322) . . . . . .
ROI example - Keith . . . . . . . . . . . . . . . . . . . . . . . . . .
29
Voxel-by-voxel plot - SPM8 original segmentation vs ground truth
Voxel-by-voxel plot - SPM8 new segmentation (default TPMs) vs
ground truth . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Voxel-by-voxel plot - SPM8 new segmentation (Chris Rorden’s TPMs)
vs ground truth . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Voxel-by-voxel plot - ALI vs ground truth . . . . . . . . . . . . . .
Confusion matrices - stroke patient . . . . . . . . . . . . . . . . . .
Confusion matrices - healthy patient . . . . . . . . . . . . . . . . .
Error rates and Total error rates . . . . . . . . . . . . . . . . . . .
ROI results (Keith) . . . . . . . . . . . . . . . . . . . . . . . . . . .
ROI results (LeftMca) . . . . . . . . . . . . . . . . . . . . . . . . .
ROI results (SmallTemporal) . . . . . . . . . . . . . . . . . . . . .
ROI results (3316) . . . . . . . . . . . . . . . . . . . . . . . . . . .
ROI results (3319) . . . . . . . . . . . . . . . . . . . . . . . . . . .
ROI results (3322) . . . . . . . . . . . . . . . . . . . . . . . . . . .
38
4.1
4.2
4.3
4.4
4.5
5.1
5.2
5.3
5.4
5.5
5.6
5.7
5.8
5.9
5.10
5.11
5.12
5.13
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
current flow
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
27
31
32
33
35
39
40
41
42
43
44
45
46
46
47
47
48
Acronyms
AC
anterior commissure
CNS
central nervous system
CSF
cerebrospinal fluid
DCS
direct current stimulation
DWI
diffusion weighted imaging
FEM
finite element method
fMRI
functional MRI
FWHM full width half maximum
GM
gray matter
LOM
lesion overlap map
MOG
mixture of Gaussians
MNI
Montreal Neurological Institute
MRI
magnetic resonance imaging
PC
posterior commissure
ROI
region of interest
SPM
statistical parametric mapping
tDCS
transcranial DCS
TMS
transcranial magnetic stimulation
TPM
tissue probability map
WM
white matter
3
Chapter 1
Introduction
Stroke patients are sometimes in need of therapy, e.g. speech therapy to regain the
ability to speak again after a stroke affecting the speech center located in the left
hemisphere. This in combination with therapy involving changing the plasticity of
the brain with for example current or a magnetic field has shown positive results.
To ensure an efficient therapy, the physician has to have knowledge about on
what locations on the head to apply current and how strongly to do so. Thus,
segmentation of the brain to gain knowledge about the anatomy is needed since the
tissues present have different electrical conductivity. For example, a necrotic stroke
lesion filled with cerebrospinal fluid (CSF) has a higher conductivity of the current
within the brain. An acute lesion may also influence the electrical conductivity
since there is some change to the brain tissue. Thus, accurate segmentation is the
key.
To break down the main purpose of this thesis work, a problem formulation
was made based upon the project proposal written by professor Lucas C Parra
at The City College of New York. Related research within the field, along with
improvements already proposed and made, is also specified. Finally, the purpose
of the thesis work is summarized into a number of goals and the organization of
the report is explained for easier reading.
1.1
Problem Formulation
Segmentation of stroke lesions is of great importance during treatment of patients
with for example aphasia, which is a disease connected to the ability to formulate
words. One way of treating these patients is by applying direct current, a method
named direct current stimulation (DCS), to the patient’s brain. Recent studies have shown that transcranial DCS (tDCS), a method of applying the current
through a number of scalp electrodes, in combination with speech therapy, relieves
the symptoms of aphasia for a period of time. Another method, named transcranial magnetic stimulation (TMS), has been used within this field and involves a
depolarization in the neurons of the brain using a magnetic field [1]. Patients find
tDCS less uncomfortable with only a slight tingling at the electrodes, while TMS
5
6
Introduction
is known to cause discomfort and pain.
To know how much current to apply, and on what area of the skull to do so,
knowledge about the anatomy of the brain and how electrical signals are conducted
within it is crucial. Applying current is easier on patients with a ’normal’ brain
anatomy since the anatomy of the brain is usually known. If the patient for
example has suffered a stroke, the anatomy of the brain changes, thus changing
the way electrical signals are conducted. This is why there is a need for correct
segmentation of brains with abnormal anatomy.
Today there is a choice of performing the segmentation manually, which is very
time consuming but produce a more reliable result, or to use one of the automatic
segmentation routines available, some of which give a poor segmentation result but
is faster than the manual dito. Thus, the ideal would be a more robust automatic
segmentation routine which can handle the changes in brain anatomy that appear
after a stroke.
This thesis will start with a thorough literature study involving for example the
anatomy of the brain, possible changes in the brain after a stroke, and the theory
that the automatic segmentation routines are based upon. An investigation of a
current automatic segmentation routine named SPM8 and the improvements made
on it will also be made.
1.2
Related Research
The software that will be used during this thesis is SPM8, developed by the Wellcome Trust Centre for Neuroimaging, London, UK. During the past few years,
improvements have been made on the original segmentation routine in SPM8.
Among these improvements are Chris Rorden’s, professor at The Georgia Institute
of Technology (www.sph.sc.edu/comd/rorden/), new and improved tissue probability maps (TPMs), Mohamed Seghier’s, doctor at the Wellcome Trust Centre
for Neuroimaging, lesion identification routine, and the new segmentation routine
in SPM8 (only work in progress). These improvements along with the original
segmentation routine are explained in Section 3.
1.3
Purpose and Goals
The purpose of this thesis is to evaluate a number of automatic segmentation
routines in presence of both acute stroke lesions and necrotic stroke lesions. It can
be summarized into a number of goals.
• Perform a study of the previous work performed within this field.
• Analyze how the original segmentation routine in SPM8 works in presence
of both acute stroke lesions and necrotic stroke lesions.
• Analyze how the new segmentation routine in SPM8 works in presence of
both acute stroke lesions and necrotic stroke lesions.
• Perform and analyze the segmentation result from the new segmentation
routine in SPM8 when using Chris Rorden’s improved TPMs.
1.4 Organization of the Thesis Report
7
• Analyze and compare the segmentation results from the segmentation routines in SPM8 with Mohamed Seghier’s segmentation routine (ALI) developed to handle stroke lesions.
• Compare the results from the automatic segmentation routines with available
manual segmentation.
• Make conclusions on which automatic segmentation routine performs the
best for stroke patients.
1.4
Organization of the Thesis Report
The chapters of the thesis report include the following information.
Chapter 2 includes an overview of the human brain anatomy, some information
about what happens in the brain after a stroke, and how this affects the segmentation result. An introduction to brain stimulation is also included together with
a brief description of manual segmentation and automatic segmentation.
Chapter 3 describes previous work performed in the field of segmentation of
human brains, with and without stroke lesions present. The original segmentation
routine in SPM8 is described together with the new segmentation routine in SPM8.
The improvements made by Chris Rorden and Mohamed Seghier are also explained
along with information about the manual segmentation performed by the Neural
Engineering group at The City College of New York. The different settings are
explained together with possible problems that may arise.
Chapter 4 gives a description on how the segmentation routines will be compared to each other, along with some information on how the results will be presented and evaluated. The patients used in the evaluation are also presented.
Chapter 5 summarizes the results from the evaluation of the original segmentation routine in SPM8 and the improvements made on it. The manual segmentation
performed by the Neural Engineering group at The City College of New York is
also included as the correct way of segmenting the brain when applicable.
Chapter 6 contains a discussion and interpretation of the results, conclusions
that can be made, and ideas on possible development and improvements in the
future.
Chapter 2
Theory
The human brain has a complex structure and is dependant upon many things,
e.g. nutrition and oxygen delievered by the blood. If the blood flow to the brain is
changed in some way, there are consequences leading to a change in brain anatomy
and brain physiology. A stroke, if left untreated, will lead to a (stroke) lesion. A
lesion is defined as a change in brain anatomy and thus, also a change in electrical
conductivity of the tissue. Due to this, there is a need to ensure proper segmentation. This will be described in this chapter along with a brief introduction to
tDCS since this thesis work is part of ongoing research related to that subject.
A brief introduction to manual segmentation and automatic segmentation is also
included.
2.1
The Human Brain
This section includes a brief introduction to the brain, based upon information
found in [2]. The brain, together with the spinal cord, makes up the central
nervous system (CNS). The brain can be divided into the telencephalon, the
cerebellum, the diencephalon, and the brainstem. An illustration of the parts of
the brain can be seen in Figure 2.1. The telencephalon is made up of two areas;
the cerebral cortex (in pink) and the basal ganglia (not marked). The cerebral
cortex is the largest part of the brain, with a surface area of about 2200 cm2 . This
part of the brain is for example involved in the process of thinking, learning, and
remembering.
The cerebral cortex is built up by neurons and unmyelinated fibers giving it
a gray color, thus this part of the brain is commonly referred to as the gray
matter (GM). Below the GM is a large mass of axons that connects the cerebral
cortex with other regions of the brain. Since the axons are myelinated, the tissue
has a white color, hence this region is often referred to as the white matter (WM).
The basal ganglia has indirect connections with the cerebral cortex, thus this part
of the brain is involved in motor control.
The cerebellum is located posterior and inferior of the telencephalon and contains about 50% of the neurons found in the CNS. There is a large number of
9
10
Theory
input connections to this part of the brain, e.g. visual input and auditory input. The diencephalon can be divided into the thalamus, the subthalamus, and
the hypothalamus. These areas are involved in remembering, releasing hormones,
and controlling body temperature and hunger. The brainstem is located directly
posterior of the spinal cord and receives sensory information and sends out motor
signals through cranial nerves.
Figure 2.1: Illustration of brain regions [3].
Surrounding the brain and the spinal cord is CSF, which is a clear bodily fluid.
It acts as a ’cusion’, protecting the brain inside the skull both mechanically and
immunologically. CSF is produced at a number of locations within the brain and
after circulating the brain, it is absorbed by the venous blood. Around 500 ml of
CSF is produced every day and since the brain can only contain around 150 ml,
the rest is drained into the blood stream. Thus, the CSF turns over nearly four
times daily.
A longitudinal fissure divides the brain into two hemisheres, connected through
the corpus collosum. The hemispheres resemble each other and the structure of
each hemisphere is generally mirrored by the other. Despite their similarities, the
hemispheres have different functions, although both hemispheres are involved in
every task performed by the brain. The left hemisphere is often called the ’dominant’ hemisphere and includes the language center. The left hemisphere processes
information in a logical and sequencial manner. The right hemisphere, involved
in interpreting visual information and spatial processing, processes information
intuitively and randomly [4]. In general, the right part of the brain controls the
left part of the body and vice versa.
The brain is divided into a number of lobes, which can be seen in Figure 2.2,
each one responsible for different tasks [5]. The frontal lobe is involved in higher
functions, e.g. interpreting touch, vision, and hearing. The motor cortex is located
in the posterior part of the frontal lobe and is involved in body movements. The
temporal lobe, located inferior of the motor cortex, is involved in understanding
language, the memory process, and hearing. The parietal lobe, located posterior to
the frontal lobe, is involved in interpretation of language and sensation of touch,
2.1 The Human Brain
11
pain, and temperature. The occipital lobe, located in the posterior part of the
brain, is where visual input is interpreted.
Figure 2.2: Illustration of brain lobes [6].
2.1.1
A Healthy Brain
The anatomy of a normal, healthy brain might look like the brain seen in Figure
2.3. In an image from a magnetic resonance imaging (MRI) investigation, different
tissues show up with different image intensities [7]. In a T1 weighted image, seen
in Figure 2.3a, WM shows up the brightest with a white color, followed by GM
represented by a light gray color. CSF has the lowest image intensity and shows
up black. In a T2 weighted MRI, seen in Figure 2.3b, CSF instead shows up bright
white and WM shows up in a darker image intensity. GM is represented by a gray
color, similar to that in a T1 weighted MRI. More information about MRI can be
found in [7].
2.1.2
Strokes
A stroke is defined as ’an interruption of the blood supply to any part of the brain’
[8]. This occurs when a blood vessel transporting blood, and oxygen, to the brain
is blocked by a blood clot (ischemic stroke) or when a blood vessel is damaged,
causing blood to leak out into the brain (hemmorhagic stroke). Since the brain
is very sensitive to loss of blood supply, the patient may suffer major damages
following a stroke. The cells in the surrounding tissue either die due to lack of
oxygen or are compressed due to the leakage of blood, thus the increased pressure
later killing them. The symptoms of a stroke depend on which part of the brain
is affected, e.g. weakness or paralysis of an arm (motor cortex affected), vision
changes (visual cortex affected), and difficulty speaking or understanding spoken
words (language center affected). Although some stroke patients will not have any
symptoms at all.
There is a need to start treatment within three hours from the onset of a stroke
to ensure the best chance of recovery. Immediate treatment is for example drugs
12
Theory
(a) T1 weighted MRI.
(b) T2 weighted MRI.
Figure 2.3: Axial slices illustrating the difference in image weighting of MRIs.
Images kindly provided by Julius Fridriksson, associate professor at The University
of South Carolina.
that break up blood clots if the patient has suffered an ischemic stroke. If the
patient instead has suffered a hemmorhagic stroke, surgery is often required to
remove the pool of blood in the brain and repair the damaged blood vessel. Long
term treatment involves recovering lost body functions, e.g. physical therapy to
learn how to walk again, and preventing future strokes. About 10% of all stroke
patients regain all functionality and about 50% of all stroke patients are able to
leave the hospital with some medical assistance.
2.1.3
Stroke Lesions
When a patient suffers a stroke, the brain tissue changes due to the lack of oxygen
supply or the increased pressure. The stroke lesion is initially classified as an acute
lesion, the difference can be seen only as a slight change in image intensity in an
MRI. An example of this can be seen in Figure 2.4a, where the lesion is defined
as the white oval shaped part in the right hemisphere.
After some time, the tissue becomes necrotic and is cleared out by scavenger
cells leaving just a hole in the tissue. The void is later filled with CSF. When
this has happened, the lesion is classified as necrotic or chronic. An example of a
necrotic lesion can be seen in Figure 2.4b. A part of the left hemisphere is affected,
i.e. the darker area together with the enlarged left ventricle. The enlarged left
2.2 Stimulation of the Brain
(a) Acute stroke lesion.
13
(b) Necrotic stroke lesion.
Figure 2.4: T1 weighted axial slices with different stroke lesion types (acute and
necrotic). Images kindly provided by Julius Fridriksson, associate professor at The
University of South Carolina.
ventricle is due to a shift in the brain following the clearing of tissue at the lesion
site. Instead of normal brain matter there is a bit of nothing, a ’hole’ in the brain,
filled with CSF. There is also some scar tissue, with image intensity similar to
that of GM, at the lesion site. Due to the location of the lesion, this patient
suffers from aphasia, a disorder connected to the ability to speak and understand
language. The type of aphasia depends on which part of the language center is
damaged, e.g. Broca’s area is connected to speech control and Wernicke’s area is
connected to speech interpretation.
This kind of damage is irreversible, although there are techniques used today
where some of the symptoms can be reduced by for example applying current at
numerous locations on the skull of the patient [9]. One method of interest for this
thesis work is tDCS, more information about this can be found in Section 2.2.
2.2
Stimulation of the Brain
Non-invasive brain stimulation has shown positive results when used on for example aphasia patients and is a promising technique according to Rossini et al [10].
tDCS involves passing low intensity direct current into the patient’s head via a
set of scalp electrodes [11]. Usually, one anode (stimulation) electrode and one
cathode (return) electrode are used to supply and pick up the current applied.
The skull shunts some of the current due to the high resistivity between the skull
and the scalp. The amount of shunting depends on for example the electrode con-
14
Theory
figuration. The current is conducted by the CSF and through the CSF network,
the current can later pass into the brain of the patient.
(a) 4-by-1 ring configuration.
(b) Rectangular pad configuration.
(c) Current flow for 4-by-1 ring configuration.
(d) Current flow for rectangular configuration.
Figure 2.5: Upper row: examples of different electrode configurations (anode in
red and cathode in blue). Second row: simulated current flow for the two electrode
configurations [12].
Studies have been conducted on which electrode configuration is better, e.g. a
ring configuration, seen in Figure 2.5a, or rectangular pads, seen in Figure 2.5b.
Using a ring configuration, the treatment has a higher spatial focality since the
current does not need to pass across the patient’s brain. Instead only the smaller
area of interest is stimulated compared to the rectangular electrode configuration.
Illustrations of finite element method (FEM) simulations of the current flow within
the patient’s brain using the different electrode configurations can be seen in Figure
2.5c and Figure 2.5d. The results indicate that the ring configuration is superior.
Research has also been conducted on the affect of placement of the return electrode
for the rectangular configuration, e.g. on the forehead above the contralateral
orbita or on the patient’s contralateral shoulder. Electrode montage (size and
position) determines the current flow, thus determining the neurophysiological
2.3 Segmentation
15
effects.
If the 4-by-1 ring configuration with sponges is used for stimulation, the maximum current used is 2 mA for a maximum of 22 minutes [13]. If a different
electrode setting is used or if the sponges are substituted with a conductive gel,
the strength of the field and the stimulation time change. Both direct current and
alternating current has been used with a positive result [14].
2.2.1
Transcranial Direct Current Stimulation (tDCS)
tDCS was developed in the early 19th century and involves stimulation of the brain
with constant direct current through scalp electrodes. It is non-invasive, costefficient, and the administration is non-complicated [15]. Electrodes are placed on
the scalp, usually with a sponge or a gel in between for better conductivity, and
a small amount of current is applied for a certain amount of time. Treatment of
some patients require repeated stimulation for a number of visits, e.g. once a week
for ten consecutive weeks, for optimal result.
The anode is placed on the area of interest, e.g. on the motor cortex for patients
with movement disabilities or on the speech center for patients with aphasia. When
current is applied, it induces intracerebral blood flow. This leads to a decrease
or an increase in neuronal excitability, depending on which type of stimulation
is used. An alteration in brain function will be the result. tDCS is used in
combination with behaviour therapy to help patients with aphasia to learn to speak
again. This method can for example be used on patients with a small vocabulary
resulting from the aphasia. For example one of the patients in the study was only
able to pronounce one word and was later helped with this method. The behaviour
therapy involves naming objects during a functional MRI (fMRI) scan to see which
areas of the brain are affected by the stroke. The severity of the aphasia decides
how many words the patient is able to say. Information from the fMRI is used
to place stimulation electrodes on the head of the patient. One week after the
stimulation, many patients are still able to pronounce the words practiced during
the therapy, which is a vast improvement. The improvement is mostly due to the
behaviour therapy, which is then boosted by the tDCS [16, 17, 18].
2.3
Segmentation
To know how much current to apply during tDCS and at what location to do so,
segmentation is the key. By segmenting the brain into different tissue classes, FEM
simulation can later predict how current will flow inside the head. The segmentation can for example be based on image intensity, color, or texture and all voxels
inside one segmentation region are similar in some way [19]. Since the presence
of lesions represents a change in electrical conductivity, proper segmentation is
crucial. Instead of normal brain matter, the lesioned part of the brain might only
contain CSF which is highly conductive, thus changing the current flow within the
brain. If the lesion is acute, there is only a small change in electrical conductivity
of the tissue, although this might be enough to give a bad stimulation result and
thus, this needs to be considered.
16
Theory
There are two main methods to use when segmenting the brain; either it can
be done manually or automatically. Both techniques have their pros and cons and
are described more thouroughly in Section 2.3.1 and Section 2.3.2. There is also a
combination of the two methods available, where the brain is first automatically
segmented and then some post-processing is performed by an operator, but this is
outside the scope of this thesis.
2.3.1
Manual Segmentation
Manual segmentation has long been the gold-standard for lesion identification and
according to John Ashburner ’human experts are currently believed to be able to
partition a brain image in a more accurate way than automated algorithms’ [20].
The segmentation is mouse-based and involves outlining or filling brain regions
for every 2D slice of an MRI scan. Since the user can only look at 2D slices
of the brain, it is important to be able to mentally reconstruct a 3D image to
ensure continuous tissue classes. Trained personel is required and it is very time
consuming [21], a scan currently takes about a few weeks to segment. Due to this
time factor, manual segmentation does not serve the needs for daily clinical use.
It is a laborious method that depends on the skill of the operator. When it
comes to segmentation of lesioned brains, manual segmentation is more reliable
than automatic segmentation since most automatic segmentation routines are not
implemented with an extra tissue class for the lesioned tissue.
2.3.2
Automatic Segmentation
Automatic segmentation is a fast technique giving an acceptable result if the patient has a normal brain anatomy without lesions and other abnormalities. Problems can also occur if the patient is older, since the brain shrinks with increasing
age and does not correspond to the expected brain anatomy. The automatic segmentation is based on a prior probability of finding different tissue types at a
certain location in the brain, this probability itself is based on information from
normal brains combined with an estimated likelihood of observing the present
intensity value. Due to this, automatic segmentation routines seldom produce
correct segmentation results for lesioned brains.
In the presence of stroke lesions, the segmentation routine performs badly and
misclassifies the lesion as something else believing that there should only be normal
tissue at the lesion location. There are many research teams currently conducting
studies on how to improve the automatic segmentation routines available on the
market today, more information about this can be found in Section 3.
Chapter 3
Previous Work
The previous work studied involves the SPM8 software package (developed by the
Wellcome Trust Centre for Neuroimaging, London, UK, http://www.fil.ion.
ucl.ac.uk/spm/) and the proposed changes made on it. SPM8 is the software of
choice in this thesis since it is state-of-the-art and widely used within this field. A
search in PubMed for articles related to this software was performed and articles
with a good theoretical background connected to this thesis work were chosen.
Articles from the Wellcome Trust Centre for Neuroimaging’s homepage were also
studied to get a better knowledge base on how the software works and the theory
behind it.
Modifications to the original segmentation routine in SPM8 have been made
the past years, information about these was also studied to get an idea on how the
current problems could be solved and implemented. The modifications of choice
for this thesis are Chris Rorden’s improved TPMs, the new segmentation routine
included in SPM8 (which is only work in progress), and a lesion classification
algorithm made by Mohamed Seghier. These automatic segmentation routines
were chosen since they have not yet been compared against each other.
The original segmentation routine in SPM8 and the modifications are described
more thoroughly in Section 3.1 - 3.4. A description of the manual segmentation is
included in Section 3.5 to give the reader an understanding of the technique and
possible problems that might occur.
3.1
SPM8 Original Segmentation
SPM8 is the latest version of the statistical parametric mapping (SPM) software.
The name refers to the ’construction and assessment of spatially extended statistical processes used to test hypotheses about functional imaging data’ [22]. This software combines segmentation, bias correction, and spatial normalization through
the inversion of a single unified model [23] and runs in Matlab. In brief, the unified model combines tissue classification, intensity bias, and non-linear warping
in one probabilistic model. Image intensities are modelled as a mixture of Gaussians (MOG) and a prior probability that a voxel intensity is drawn from a certain
17
18
Previous Work
tissue class is given by the mixing proportions. In the unified model, the priors of
the tissue classes are encoded by deformable TPMs illustrated in Figure 3.1.
(a) Gray matter.
(b) White matter.
(c) CSF.
Figure 3.1: Illustration of the TPMs used by the original segmentation routine in
SPM8. Image intensity values range from zero (black) to one (white). A higher
intensity value represents a higher probability of the voxel correponding to the
tissue of interest.
The tissue classification requires the image volumes to be registered to the
TPMs. After the registration, the maps represent a prior probability of voxels
at each location belonging to different tissue classes. Bayes rule is then used
to combine these priors with the probabilities for different tissue classes derived
from different voxels intensities to provide a posterior probability. However, this
procedure is circular since the registration requires an initial tissue classification
and vice versa. This problem is solved by combining both components into a
single generative model. An example of a segmentation result from the original
segmentation routine can be seen in Figure 3.2a.
The segmentation routine automatically segments the image volume into GM,
WM, and CSF based upon the probability of finding these tissue types at different locations combined with an estimated likelihood of observing the present
intensity value. This is done by using the TPMs previously mentioned, one
for each tissue class. The TPMs are made by taking the average of a large
number of normal brains. The TPMs used in the original segmentation routine are modified versions of the maps provided by the ICBM Tissue Probabilistic Atlases. These TPMs are kindly provided by the International Consortium for Brain Mapping, John C. Mazziotta and Arthur W. Toga. http:
//www.loni.ucla.edu/ICBM/ICBMTissueProb.html.
There are a number of user settings in the original segmentation routine, which
when changed will affect the segmentation result. These settings are specified and
explained more thoroughly in Section 3.1.1.
3.1 SPM8 Original Segmentation
(a) SPM8 original segmentation.
19
(b) ALI.
Figure 3.2: Example of segmentation result from the original segmentation in
SPM8 and ALI for a patient with a necrotic stroke lesion (Keith). GM is marked
with a red color, WM is represented by a blue color, and CSF is represented by
a green color. For ALI, the lesion class is marked with a pink color. These are
the same axial slices as shown in Figure 4.2a. The result is thresholded by 0.2 to
remove voxels with a low probability of belonging to the segmented tissue classes.
3.1.1
Settings
The settings for the original segmentation routine in SPM8 can be divided into
three main categories; data, output files, and custom. These settings are explained
more in detail in Table 3.1.
3.1.2
Problems
The major problem with SPM8 original segmentation is that since the TPMs
are based on normal brains only, segmentation of an abnormal brain will give an
incorrect result. The segmentation routine only classifies the brain tissue into three
different tissue classes, which is a problem if you for example would like to segment
out bone tissue. Better classes for internal structures in the GM would give a more
accurate segmentation result. Since this is only a mono-spectral implementation,
it can be improved by making it multi-spectral by including the information from
for example a T2 weighted MRI. The model does not take into account that
neighbouring voxels are likely to be of the same tissue class.
20
Previous Work
Data
This setting involves choosing which image volumes
to segment. Currently, this segmentation routine
only works in a mono-spectral way, i.e. with information from only one image modality.
Output files
Gray matter
White matter
Cerebro-spinal fluid
Bias corrected
Clean up any partitions
Custom
Tissue probability maps
Gaussians per class
Affine regularisation
Warping regularisation
Warp frequency cutoff
Bias regularisation
Bias FWHM
Sampling distance
Masking image
Choose which output file to produce for GM. The
options are native space, unmodulated normalized,
modulated normalized, and a combination of them.
Same options as for GM.
Same options as for GM.
Choose to save a bias corrected version of the input
data.
Decide if there should be any cleaning up performed
on the result, this enables an extraction of the brain
from the segmented images.
Choose which TPMs to use during the segmentation.
Choose how many Gaussians that should be used to
represent the image intensity for each tissue class.
Choose which brain type to perform the affine regularisation to.
The amount of regularisation involves a tradeoff between two terms; how probable the data is given the
warping parameters and how probable the parameters are including a penalty for unlike deformations.
If the image appears distorted, it is a good idea to
increase the amount of regularisation. An increased
regularisation gives smoother deformations.
A smaller number will allow more detailed deformations to be modelled, although this gives a great increase in memory needed and thus, also computational time is increased.
Involves a spatial bias that may corrupt the image
volume and make segmentation harder. If the data
has very little intensity non-uniformity artifacts, the
bias regularisation should be increased.
Setting the full width half maximum (FWHM) of
the Gaussian smoothness of the bias. This will prevent the algorithm from trying to model out intensity
variations due to different tissue types.
Choose sampling distance for estimation of model
parameters. Smaller values use more data, thus giving a better segmentation result. Although this will
also lead to an increase in computational time.
The result can be masked to ensure that it conforms
the same space as the original image volume.
Table 3.1: Settings for SPM8 original segmentation.
3.2 Mohamed Seghier’s ALI
3.2
21
Mohamed Seghier’s ALI
Mohamed Seghier has improved the original segmentation routine in SPM8 by
adding a tissue class for the lesion. He has published an article [21] about this
work and was kind enough to provide the software code for evaluation. This has
enabled testing and evaluation of the segmentation algorithm on both necrotic
stroke lesions and acute stroke lesions.
ALI is a modification of the original unified segmentation already present in
SPM8. The purpose of the segmentation routine is to minimize misclassification
of abnormal tissue in GM and WM segments by segmenting these abnormal voxels as CSF and/or as members of a fourth extra tissue class. An example of a
segmentation result from ALI can be seen in Figure 3.2b.
With ALI the user can identify abnormal/lesioned brain tissue and generate
lesion overlap maps (LOMs) of the population/group. Basically, when lesions
are identified, a LOM can be generated and voxels that are frequently lesioned
in the population/group can be identified (e.g. this is useful for lesion-symptom
mapping). The identification of abnormal brain tissue is divided into four steps.
• Segmentation of the anatomical images with the modified unified segmentation (with an ’extra’ lesion class).
• Smoothing of the segmented GM and WM images.
• Detection of abnormal voxels in both GM and WM images (compared to
normal values of a control population).
• Grouping and definition of the lesion.
The prior probability for the lesion class is made up by a combination of the
TPMs for CSF and WM. The default setting is the mean of these two TPMs.
This simple prior has a low probability in the GM to avoid including normal GM
voxels in the lesion class. Similarly, the prior has a low probability in the WM
segment. The prior has a higher probability in the WM area than in the GM area,
thus forcing the segmentation to reclassify mis-classified tissue. An illustration of
the TPM for the extra tissue class can be seen in Figure 3.3.
Figure 3.3: Illustration of the TPM for the extra class used by ALI. Image intensity
values range from zero (black) to one (white). A higher intensity value represents
a higher probability of the voxel correponding to the tissue of interest.
22
3.2.1
Previous Work
Settings
There are not a lot of settings for the user to change in ALI, the settings involving
the unified segmentation are set by the program itself. The settings that the user
can change are specified in Table 3.2.
Data
Involves choosing which image volumes to segment.
Important to remember is that ALI is only tested for
T1 weighted MRIs.
Tissues
Prior extra class
Other settings
Number of iterations
Threshold probability
Threshold size
Coregister to MNI space
The prior of the extra class is chosen. The default
TPM is the mean of the TPMs for WM and CSF.
This is a good estimate.
Choose how many iterations to run, a good value is
around two to three. A lower number of iterations
may lead to lesioned tissue not being included in the
extra/lesion class. An increased number of iterations
may lead to healthy tissue being classified as lesioned
tissue. Information about this can be found in the
complementary material to [21].
Before using the result as a prior probability in the
next iteration, the result is cleaned up by removing
voxels below a user defined probability value. This
step will help to limit the search to only abnormal
voxels with a high probability in the lesion class.
Before using the result as a prior probability in the
next iteration, the result is cleaned up to remove
smaller areas segmented as lesion. Only regions with
a relatively big size (> threshold) will be considered
for the definition of the prior for the lesion class.
Choose to coregister to Montreal Neurological Institute (MNI) space or not. Coregistration will help the
accuracy of the segmentation routine.
Table 3.2: Settings for ALI.
3.2.2
Problems
ALI is a temporary version and is only tested for T1 weighted MRIs. For some
patients the lesion is not classified as lesion; if the lesion class is empty after
segmentation, either the lesion area is classified as part of another tissue class or
not at all.
3.3 SPM8 New Segmentation
3.3
23
SPM8 New Segmentation
The new segmentation routine in SPM8 is an extension of the default unified
segmentation. It is essentially equal to the algorithm described in Section 3.1,
although there are four major changes made. First, there is a slightly different
treatment of the mixing proportions. Second, the registration model used is improved. Third, a multi-spectral mode is implemented, making it possible to use
information from different image modalities (e.g. T1 weighted MRIs, T2 weighted
MRIs, and diffusion weighted imaging (DWI)) in the segmentation. Fourth, there
is an extended set of TPMs, illustrated in Figure 3.4, allowing voxels within the
brain to be treated differently.
(a) Gray matter.
(b) White matter.
(c) CSF.
Figure 3.4: Illustration of the TPMs used by the new segmentation routine in
SPM8. Image intensity values range from zero (black) to one (white). A higher
intensity value represents a higher probability of the voxel correponding to the
tissue of interest.
The algorithm can segment the brain into a total of six tissue classes; GM,
WM, CSF, bone, soft tissue, and air/background. An example of a segmentation
result from this routine can be seen in Figure 3.5.
3.3.1
Settings
The settings for this segmentation routine are similar to the ones mentioned in
Section 3.1.1. The settings are presented and explained in Table 3.3.
3.3.2
Problems
Since this is only an extension of the original segmentation routine in SPM8, the
problem with lesioned brains lingers. The new segmentation routine is included in
SPM8, although it is only work in progress and there is hardly any documentation
on it. Another question is if it is rigourously tested. The multi-spectral mode
does not produce a correct segmentation result, e.g. the segmented CSF is not a
continuous layer which is anatomically incorrect.
24
Previous Work
Data
Volumes
Bias regularisation
Bias FWHM
Save bias corrected
Tissues
Tissue probability maps
Number of Gaussians
Native tissue
Warped tissue
Warping
Warping regularisation
Affine regularisation
Sampling distance
Deformation fields
This setting involves choosing which image volumes
to segment. Firstly, the user decides whether to use
mono-spectral mode or multi-spectral mode. The
different channels can be treated in different ways
when it comes to the bias, although it is only the
first channel’s data that is used for the initial affine
regularisation with the TPMs.
Involves a spatial bias that may corrupt the image
volume and make segmentation harder. If the data
has very little intensity non-uniformity artifacts, the
bias regularisation should be increased.
Setting of the FWHM of the Gaussian smoothness of
the bias. This will prevent the algorithm from trying to model out intensity variations due to different
tissue types.
Choose whether to save a bias corrected version or
not.
Choose which TPMs to use, in the same way as in
Section 3.1.1.
Choose which number of Gaussians to use to describe
the image intensity of the tissue classes.
Decide whether to save the result in native space or
not. By saving the segmentation result in the native
space, the result can later be overlayed on top of the
original anatomical image.
Decide to save the result in warped space or not. If
so, decide which warping to use.
The amount of regularisation involves a tradeoff between two terms; how probable the data is given the
warping parameters and how probable the parameters are including a penalty for unlike deformations.
If the image appears distorted, it is a good idea to
increase the amount of regularisation. An increased
regularisation gives smoother deformations.
Choose which brain type to perform the affine regularisation to.
Choose sampling distance for estimation of model
parameters. Smaller values use more data, thus giving a better segmentation result. Although this will
also lead to an increase in computational time.
Select to save the deformation field as a .nii-file or
not.
Table 3.3: Settings for SPM8 new segmentation.
3.3 SPM8 New Segmentation
(a) SPM8 new segmentation monospectral.
25
(b) SPM8 new segmentation multispectral.
Figure 3.5: Example of segmentation results from the new segmentation routine in
SPM8 using default TPMs for mono-spectral mode and multi-spectral mode on a
patient with a necrotic stroke lesion (Keith). GM is marked with a red color, WM
is represented by a blue color, and CSF is represented by a green color. These are
the same axial slices as shown in Figure 4.2a. The result is thresholded by 0.2 to
remove voxels with a low probability of belonging to the segmented tissue classes.
26
3.4
Previous Work
Chris Rorden’s TPMs
Chris Rorden, professor at the Georgia Institute of Technology (www.sph.sc.edu/
comd/rorden/), has improved segmentation performed with SPM8 by providing
improved TPMs. There are two different versions of TPMs available, either a set
including only the head region or a set that also includes a part of the neck and
the spine. There are TPMs available for classification of six tissue classes; GM,
WM, CSF, bone, soft tissue, and air/background. In this thesis work, the TPMs
for the head region only are used. These TPMs are illustrated in Figure 3.6.
(a) Gray matter.
(b) White matter.
(c) CSF.
Figure 3.6: Illustration of Chris Rorden’s TPMs. Image intensity values range
from zero (black) to one (white). A higher intensity value represents a higher
probability of the voxel correponding to the tissue of interest.
The TPMs are used in the new segmentation routine in SPM8 to evaluate in
what way segmentation result is improved. Both mono-spectral mode and multispectral mode will be used and evaluated. Examples of segmentation results from
the new segmentation routine in SPM8 using Chris Rorden’s TPMs are illustrated
in Figure 3.7.
3.4 Chris Rorden’s TPMs
(a) SPM8
spectral.
Chris
Rorden
27
mono-
(b) SPM8 Chris Rorden multi-spectral.
Figure 3.7: Example of segmentation results from the new segmentation routine
in SPM8 using Chris Rorden’s TPMs for mono-spectral mode and multi-spectral
mode on a patient with a necrotic stroke lesion (Keith). GM is marked with a red
color, WM is represented by a blue color, and CSF is represented by a green color.
These are the same axial slices as shown in Figure 4.2a. The result is thresholded
by 0.2 to remove voxels with a low probability of belonging to the segmented tissue
classes.
28
3.5
Previous Work
Manual Segmentation
Manual segmentation is, as mentioned earlier, considered gold-standard and will
be treated as the correct way of segmenting the brain in this thesis, i.e. the ground
truth. The segmentation is performed by the Neural Engineering group at The
City College of New York using a software named ScanIP, a feature of Simpleware.
It is performed by a skilled engineer and it currently takes a couple of weeks to
segment one image volume into six tissue classes; GM, WM, CSF, skin, bone, and
air/sinus.
The operator manually marks voxels corresponding to different tissue classes by
clicking on the screen. Each tissue is a different mask and each mask can be set to
a different color. This is repeated for all axial slices of the image volume, making it
important to have a good spatial thinking while segmenting to for example ensure
continuous tissue layers. Another challenge is the quality of the anatomical images
which varies between scans. In some cases it can be hard to distinguish between
tissue types since there is not a clear line identifying the boundary. In those cases,
the segmentation is based more on experience of the operator and the surrounding
image slices.
The general procedure involves setting the shape of the head by coloring the
skull. The skin is then traced and any overlapping skin is subtracted from the
skull. This is followed by a similar procedure for the brain. The folds and shapes
of the brain is outlined by marking the CSF and subtracting the brain matter
(usually GM) from the CSF.
3.5.1
Problems
Since manual segmentation involves one person going through all slices of a image
volume, it is very time consuming. It is also user dependant, hence the segmentation result depends on the skill of the operator. Due to the operator dependancy,
segmentation result of the same patient might give a different result when done
by someone else.
Since only information from T1 weighted MRIs is used during the manual segmentation, there can be an error in segmentation result since not all structures are
clearly visible. One idea would be to use information from T2 weighted image volumes also to get a more reliable result. For acute stroke lesions, the segmentation
may be helped by using information from different image modalities, for example
DWIs.
Chapter 4
Method
To evaluate the performance of the automatic segmentation routines, a number
of stroke patients are used as an input. The influence of the settings was tested,
although the default settings gave the best results and are used to produce the
results in this report. The patients are presented in Section 4.1 which is followed
by an introduction to the evaluation technique in Section 4.3.
The image volumes were shifted using SPM8 to ensure that the location of
the anterior commissure (AC) is coordinate [0 0 0] and that the line between the
AC and the posterior commissure (PC) is horizontal. This is important since the
automatic segmentation routines assume this. An illustration of AC, PC, and the
line connecting them (AC-PC line) can be seen in Figure 4.1. The segmentation
routines also depend on the patient being coregistered to the TPMs. If the segmentation result was poor, a coregistration to the TPMs was performed and the
patient was segmented again. This was only the case using some of the automatic
segmentation routines on the patient named Ela, presented in Section 4.1.1.
Figure 4.1: Illustration of AC, PC, and AC-PC line [24].
29
30
4.1
Method
Patients
Automatic segmentation routines are known to perform well in presence of necrotic
stroke lesions, since the only difference from a normal brain is the extra presence
of CSF at the lesion site. Thus, necrotic stroke lesions are usually just segmented
as CSF which is correct. Problems occur when there is scar tissue present at the
lesion site, which is usually segmented as GM since the image intensity is similar
to that of GM. Since the segmentation is based on the image intensity, this is
correct. There will be a problem when using this segmentation result in the FEM
simulation though since scar tissue does not have the same electrical conductivity
as GM. Brains with acute stroke lesions are known to cause more problems and
give a poor segmentation result due to only a slight change in image intensity in
the lesion area. Depending on what type of lesion present, it can be segmented as
either GM, WM, or CSF depending on the image intensity at the lesion site.
First, the automatic segmentation routines are evaluated for patients with
necrotic stroke lesions. After analyzing the segmentation results, the automatic
segmentation routines are again evaluated for patients with acute stroke lesions.
4.1.1
Patients with Necrotic Stroke Lesions
A set of four patients with necrotic stroke lesions is used to initially evaluate
the segmentation routines. The lesions that are somewhat different from each
other, regarding size, location and image intensity, are illustrated in Figure 4.2
and Figure 4.3. Both mono-spectral segmentation and multi-specral segmentation
is performed on three of these patients; on Ela only mono-spectral segmentation
is performed since only a T1 weighted MRI is available for her.
Keith
Keith has a large lesion, which can be seen in Figure 4.2a and Figure 4.2d, affecting
the left hemisphere including the speech center. The lesion is labeled as an open
lesion, since it affects the outer part of the cerebral cortex it is impossible to know
the outer border of it. There is a mass of scar tissue in the lesion area, seen as a
darker image intensity of the tissue surrounding the lesion area (the paralesional
area). Since the scar tissue has an image intensity similar to that of GM, there is
a high probability of the scar tissue being segmented as GM. Since scar tissue is
known to have a slightly different electrical conductivity than GM, the result from
the FEM simulation will not be entirely correct.
LeftMca
This patient has a stroke lesion similar to Keith’s, although it affects a larger part
of the left hemisphere. This stroke lesion is also classified as open since there is
no clear outer border of it. There is a lot of scar tissue present at the lesion site,
following earlier reasoning it will probably be segmented as GM. Axial slices of
this patient can be seen in Figure 4.2b and Figure 4.2e.
4.1 Patients
31
(a) Keith T1.
(b) LeftMca T1.
(c) SmallTemporal T1.
(d) Keith T2.
(e) LeftMca T2.
(f) SmallTemporal T2.
Figure 4.2: Axial slices from T1 weighted MRIs (first row) and T2 weighted MRIs
(second row) of patients with necrotic stroke lesions (lesion area marked with red
arrow). Images kindly provided by Julius Fridriksson, associate professor at The
University of South Carolina.
SmallTemporal
The stroke lesion, which can be seen in Figure 4.2c and Figure 4.2f, is located in
the left hemisphere along the edge of the cerebral cortex in the temporal lobe. It
is not as big as the lesions previously mentioned, although there is a change in
anatomy and image intensity. The image intensity of the lesion is similar to that
of GM, thus there is a high probability of the lesion being segmented as GM.
32
Method
Ela
Ela has a lesion in the visual cortex, located in the posterior part of the brain. The
lesion, illustrated in Figure 4.3, has well defined borders and is filled with CSF.
There is a high probability of this stroke lesion being segmented as CSF only.
Figure 4.3: Axial slice from a T1 weighted MRI of Ela with a necrotic stroke lesion
(lesion area marked with red arrow). Image kindly provided by The City College
of New York.
4.1.2
Patients with Acute Stroke Lesions
A set of three patients, illustrated in Figure 4.4, with acute stroke lesions is also
used to evaluate the automatic segmentation routines. The lesions are all different
from each other regarding size, location, and image intensity. The MRIs from the
patients with acute stroke lesions are of low resolution compared to those with
necrotic stroke lesions. The mono-spectral mode was used on these patients since
there were no T2 weighted MRIs available.
3316
This patient has a large lesion, seen in Figure 4.4a, in the posterior part of the
right hemisphere. The actual lesion is the dark matter surrounded by a white
border. This is a result of a hemmorhagic stroke when blood leaks out into the
brain. It is hard to know what the white border actually is and if it is a part of
the lesion. Since the lesion has a low image intensity similar to that of GM inside
the white border, it will probably be segmented as that. The white ring has a very
bright image intensity and will probably be segmented as WM.
4.2 Automatic Segmentation vs. Ground Truth
33
3319
This patient’s lesion, seen in Figure 4.4b, is marked by the white oval shaped area
in the right hemisphere. By looking at results from another image modalities than
the T1 weighted image (e.g. DWI), it is clear that the lesion site is larger than
what can be seen in the T1 image. Due to the intensity of the lesion being similar
to that of WM, it will probably be segmented as WM.
3322
The lesion in this patient is a bit different from the other two patients with acute
stroke lesions since it is on the outer part of the cerebral cortex. The lesion is
illustrated in Figure 4.4c by the area with a lower image intensity. Due to the
image intensity of the lesion area, it will probably be segmented as GM.
(a) T1 3316
(b) T1 3319
(c) T1 3322
Figure 4.4: Axial slices from T1 weighted MRIs of patients with acute stroke
lesions (lesion area marked with red arrow). Images kindly provided by Julius
Fridriksson, associate professor at The University of South Carolina.
4.2
Automatic Segmentation vs. Ground Truth
Ground truth (manually segmented GM, WM, CSF, bone, skin, and lesion) is
available for one stroke patient (Ela). Ground truth (manually segmented GM,
WM, CSF, bone, and skin) is also available for one healthy patient. A comparison
will be made between the automatic segmentation routines and the ground truths
for both patients. The original segmentation routine in SPM8, the new segmentation routine in SPM8 using the default TPMs, the new segmentation routine in
SPM8 using Chris Rorden’s TPMs, and Mohamed Seghier’s ALI will be evaluated
and the corresponding segmentation results compared to each other.
A confusion matrix is made for each automatic segmentation routine and each
patient. A confusion matrix represents the accuracy of the predicted result (from
the automatic segmentation routine) compared to the ground truth. An Error
34
Method
rate and a Total error rate of how well the segmentation routines perform are
calculated based on the elements in the confusion matrix. The diagonal elements
are considered correct and all off-diagonal elements are considered part of the
error.
The Error rate, see Equation 4.1, calculated for each tissue type (T) and segmentation routine (M), is a measurement of how well the automatic segmentation
routine performs for each tissue class. The Total error rate, see Equation 4.2, which
is calculated for each automatic segmentation routine (M), is a measurement of
how well the segmentation routine performs over all. The error measurements are
calculated as
Error rate(T,M) =
error(T, M )
error(T, M ) + correct(T, M )
Total error rate(M) =
4.3
error(M )
error(M ) + correct(M )
(4.1)
(4.2)
Regions of Interest (ROIs)
To compare the results from the automatic segmentation routines, a region of
interest (ROI) is defined for each patient (except for Ela who is only used in the
comparison between automatic segmentation and ground truth). The ROIs are
defined as cubes with a side length of 30 voxels. Since the image volumes for each
patient have different image resolution, the size of the ROI in millimeter (mm)
differs between the patients. This can be seen in Table 4.1. Lesioned tissue is
included in the ROIs, along with normal tissue and scar tissue, to evaluate the
performance of the routine at the lesion site. An example of a ROI can be seen in
Figure 4.5.
Patient
Image resolution [mm]
Keith
LeftMca
SmallTemporal
3316
3319
3322
1*1*1
1*1*1
1*1*1
0.45 * 0.45 * 0.9
0.98 * 0.98 * 1
0.98 * 0.98 * 1
Table 4.1: Image resolution for image volumes/patients used to evaluate the automatic segmentation routines.
4.3 Regions of Interest (ROIs)
35
Voxels corresponding to different tissue classes within the ROI are summarized
as a measure of how well the automatic segmentation routines performs. For the
patient with manual segmentation results available (Keith), this will be considered
ground truth and the automatic segmentation routines will be compared to these
values. For the other patients, there is no ground truth available and a comparison
between the automatic segmentation routines will be made instead.
Figure 4.5: Keith’s ROI illustrated by a white square. Top left corner coronal
slice, top right corner sagittal slice, and bottom left corner axial slice. The ROI is
a cube with a side length of 30 voxels (30 mm).
Chapter 5
Results
This section includes results from the segmentation routines for each patient. The
manual segmentation (ground truth) for Ela is compared to the results from the
automatic segmentation routines in voxel-by-voxel plots. Confusion matrices are
made and error rates are calculated for one stroke patient (Ela) and one healthy
patient. The segmentation results from the healthy patient will be considered
ideal. Finally, the results from the ROIs are summarized for each patient (three
acute and three necrotic) and each segmentation routine. Further analysis and
conclusions of the results can be found in Section 6.
5.1
Automatic Segmentation vs. Ground Truth
A comparison is made between the automatic segmentation results and the manual
segmentation results (ground truth) for Ela. Illustrations of the results can be seen
in Figure 5.1 - 5.4, where the difference in segmentation result is plotted voxel-byvoxel.
Confusion matrices are made for each automatic segmentation routine and each
patient (one stroke patient (Ela) and one healthy patient). The confusion matrices
represent the accuracy of the automatic segmentation results compared to ground
truth. The confusion matrices are illustrated in Figure 5.5 and Figure 5.6. All
confusion matrices for Ela are normalized in the same manner. This is also true
for the healthy patient, although the confusion matrices are normalized with a
different value than for Ela.
An Error rate is calculated for each segmented tissue class and each automatic
segmentation routine. It is summarized in plots which can be seen in Figure 5.7a
and Figure 5.7b. The error rate is normalized in the same manner for all tissue
types for Ela. This is also true for the error rate for the healthy patient, although
it is normalized with a different value than for Ela. A Total error rate is calculated
for each automatic segmentation routine. It is summarized in plots which can be
seen in Figure 5.7c and Figure 5.7d.
The image intensity of the confusion matrices and the error plots is gamma
corrected (γ = 0.5) for both the healthy patient and the stroke patient. This
37
38
Results
means that the image intensity is not linear, but non-linear, thus increasing the
resolution in the areas with lower (darker) image intensity. This is important to
remember since the difference in image intensity is a sign of the performance of the
automatic segmentation routines, i.e. areas with a lower image intensity appear
brighter than they really are.
Figure 5.1: Illustration of voxels segmented incorrectly by the original segmentation routine in SPM8. The first line of plots corresponds to GM, the second line
of plots corresponds to WM, and the third line of plots corresponds to CSF. The
first column illustrates an axial slice, the second line illustrates a coronal slice, and
the third column illustrates a sagittal slice. Green voxels correspond to ground
truth that the original segmentation routine failed to classify as the current tissue
class and orange voxels correspond to voxels segmented incorrectly as the current
tissue class by the original segmentation routine.
5.1 Automatic Segmentation vs. Ground Truth
39
Figure 5.2: Illustration of voxels segmented incorrectly by the new segmentation
routine in SPM8 using default TPMs. The first line of plots corresponds to GM, the
second line of plots corresponds to WM, and the third line of plots corresponds to
CSF. The first column illustrates an axial slice, the second line illustrates a coronal
slice, and the third column illustrates a sagittal slice. Green voxels correspond to
ground truth that the new segmentation routine failed to classify as the current
tissue class and orange voxels correspond to voxels segmented incorrectly as the
current tissue class by the new segmentation routine using default TPMs.
40
Results
Figure 5.3: Illustration of voxels segmented incorrectly by the new segmentation
routine in SPM8 using Chris Rorden’s TPMs. The first line of plots corresponds
to GM, the second line of plots corresponds to WM, and the third line of plots
corresponds to CSF. The first column illustrates an axial slice, the second line
illustrates a coronal slice, and the third column illustrates a sagittal slice. Green
voxels correspond to ground truth that the new segmentation routine failed to classify as the current tissue class and orange voxels correspond to voxels segmented
incorrectly as the current tissue class by the new segmentation routine using Chris
Rorden’s TPMs.
5.1 Automatic Segmentation vs. Ground Truth
41
Figure 5.4: Illustration of voxels segmented incorrectly by ALI. The first line of
plots corresponds to GM, the second line of plots corresponds to WM, the third
line of plots corresponds to CSF, and the fourth line of plots corresponds to the
lesion class. The first column illustrates an axial slice, the second line illustrates
a coronal slice, and the third column illustrates a sagittal slice. Green voxels
correspond to ground truth that ALI failed to classify as the current tissue class
and orange voxels correspond to voxels segmented incorrectly as the current tissue
type by ALI.
42
Results
Figure 5.5: Confusion matrices for the stroke patient (Ela). Top left corner SPM8
original segmentation, top right corner SPM8 new segmentation using default
TPMs, lower left corner SPM8 new segmentation using Chris Rorden’s TPMs, and
lower right corner ALI. The ideal case is higher values (brighter) for the diagonal
elements and low values (darker) for the off-diagonal elements, this corresponds
to a correct automatic segmentation result. The new segmentation routine using
Chris Rorden’s TPMs performs the best for this patient, although it is far from
ideal.
5.1 Automatic Segmentation vs. Ground Truth
43
Figure 5.6: Confusion matrices for the healthy patient. Top left corner SPM8 original segmentation, top right corner SPM8 new segmentation using default TPMs,
lower left corner SPM8 new segmentation using Chris Rorden’s TPMs, and lower
right corner ALI. The ideal case is higher values (brighter) for the diagonal elements and low values (darker) for the off-diagonal elements, this corresponds to a
correct automatic segmentation result. The new segmentation routine using Chris
Rorden’s TPMs is almost ideal and performs the best for this patient.
44
Results
(a) Error rate (stroke patient).
(b) Error rate (healthy patient).
(c) Total error rate (stroke patient).
(d) Total error rate (healthy patient).
Figure 5.7: Error rates and Total error rates for the stroke patient (left) and the
healthy patient (right). Top row illustrates the Error rate for each tissue class
and automatic segmentation routine compared to ground truth. A lower image
intensity (darker) represents a better automatic segmentation result compared to
a higher image intensity (brighter). Bottom row illustrates the Total error rate
for every automatic segmentation routine. A lower percentage represents a better
segmentation result and vice versa. ’Chris Rorden’ indicates the usage of the new
segmentation routine in SPM8 with Chris Rorden’s TPMs.
5.2 Regions of Interest (ROIs)
5.2
45
Regions of Interest (ROIs)
ROIs are defined for a set of six stroke patients (three with acute stroke lesions
and three with necrotic stroke lesions) and includes lesioned tissue, healthy tissue,
and scar tissue. The areas are used to evaluate the performance of the automatic
segmentation routines where they are known to produce poor segmentation results.
The results are summarized for each patient. Each segmentation routine is
represented by a stack and each segmented tissue class is represented as a part
of the stack with a certain color. The manual segmentation is considered ground
truth, i.e. the correct way of segmenting the tissue. For patients without ground
truth (i.e. all patients except Keith), a comparison is made between the automatic
segmentation routines instead. The results can be seen in Figure 5.8 - 5.13. Since
the new segmentation routine in SPM8 should perform better than the other
automatic segmentation routines, the results from this routine is considered the
best way of segmenting the patient’s brain.
Figure 5.8: Segmentation results for Keith. Each segmentation routine is represented by a stack and each segmented tissue is represented by a color according
to the legend. Multi-spectral segmentation was possible for this patient. Manual
segmentation was available from the Neural Engineering group at The City College
of New York. ’Chris Rorden’ indicates the usage of the new segmentation routine
in SPM8 with Chris Rorden’s TPMs.
46
Results
Figure 5.9: Segmentation results for LeftMca. Each segmentation routine is represented by a stack and each segmented tissue class is represented by a color
according to the legend. Multi-spectral segmentation was possible for this patient.
’Chris Rorden’ indicates the usage of the new segmentation routine in SPM8 with
Chris Rorden’s TPMs.
Figure 5.10: Segmentation results for SmallTemporal. Each segmentation routine
is represented by a stack and each segmented tissue class is represented by a color
according to the legend. Multi-spectral segmentation was possible for this patient.
’Chris Rorden’ indicates the usage of the new segmentation routine in SPM8 with
Chris Rorden’s TPMs.
5.2 Regions of Interest (ROIs)
47
Figure 5.11: Segmentation result for 3316. Each segmentation routine is represented by a stack and each segmented tissue class is represented by a color according to the legend. ’Chris Rorden’ indicates the usage of the new segmentation
routine in SPM8 with Chris Rorden’s TPMs.
Figure 5.12: Segmentation results for 3319. Each segmentation routine is represented by a stack and each segmented tissue class is represented by a color according to the legend. ’Chris Rorden’ indicates the usage of the new segmentation
routine in SPM8 with Chris Rorden’s TPMs.
48
Results
Figure 5.13: Segmentation results for 3322. Each segmentation routine is represented by a stack and each segmented tissue class is represented by a color according to the legend. ’Chris Rorden’ indicates the usage of the new segmentation
routine in SPM8 with Chris Rorden’s TPMs.
Chapter 6
Discussion
This section includes an interpretation and evaluation of the results presented in
Section 5, along with conclusions made, and a list of future improvements of the
segmentation routines evaluated.
6.1
Interpretation of the Results
The results presented in Section 5 will be discussed and evaluated in this chapter. The automatic segmentation routines vs. ground truth is evaluated for two
patients, one healthy patient and one stroke patient (Ela), in Section 6.1.1. The
results from the ROIs for a total of six stroke patients, three with acute stroke
lesions and three with necrotic stroke lesions, are evaluated in Section 6.1.2 for
each available segmentation routine and each tissue class.
6.1.1
Automatic Segmentation vs. Ground Truth
This section includes an evaluation and discussion of the difference in segmentation
result from the automatic segmentation routines compared to ground truth for Ela.
The difference is plotted voxel by voxel as an illustration.
An interpretation of the confusion matrices for the stroke patient (Ela) and the
healthy patient is also included together with a discussion about the error rates
based on the elements in the confusion matrices.
Difference in Segmentation
An illustration of the difference in segmentation results from the automatic segmentation routines compared to ground truth can be seen in Figure 5.1 - 5.4.
What is clear in the illustration of the error of the original segmentation routine
in SPM8, seen in Figure 5.1, is that the segmentation result is incorrect for all tissue
classes presented. The segmentation result follows the shape of the head instead
of the shape of the brain. This result is common if there is a problem with the
coregistration to the TPMs, although in this case coregistration does not seem to
49
50
Discussion
produce a more accurate result. If no coregistration is used, the result is even
worse than the one presented. Due to this problem, a large number of voxels of for
example WM are not segmented correctly, which can be seen in the second line in
the illustration. A large number of voxels are incorrectly segmented as GM and
CSF, seen as orange voxels in the illustration.
The new segmentation routine in SPM8, regardless of TPMs used, performs
well for the WM, since the number of voxels wrongly classified is small. The routine
has more problems with the other tissue classes, especially with CSF voxels located
further out in the brain, i.e. closer to the skull. What can be seen clearly in the
second line of Figure 5.2 and Figure 5.3 is that the segmentation routine does
not segment the entire spine included in the original image volume. This is since
the TPMs only cover the head region. The lower part of the spine is included in
the ground truth, making it turn up green in the illustration. When using Chris
Rorden’s TPMs, the eyes are included in the segmented CSF. This is not the case
for the manual segmentation, making the eyes turn up orange in the illustration.
There is a medium sized error in the results from ALI, mostly for the GM and
CSF, which can be seen in Figure 5.4. ALI fails to classify a number of GM voxels
in the outer part of the GM layer, i.e. closer to the skull. It also classifies a thicker
layer of CSF than what is anatomically correct. There is an error in the lesion
class, since ALI does not classify the lesion site as a part of the lesion class. Thus,
the lesion class turns up green in the illustration.
Confusion Matrices
A confusion matrix is an overall measurement of the performance of the automatic
segmentation routine, divided into the tissue classes used. For the stroke patient,
a total of seven tissue classes are used in the comparison: GM, WM, CSF, bone,
skin, lesion, and other (including non-segmented voxels and voxels corresponding
to air/background). For the healthy patient, a total of six tissue classes are used in
the comparison: GM, WM, CSF, bone, skin, and other (including non-segmented
voxels and air/background). The healthy patient is used as a comparison since
there should not be any errors when segmenting a brain with a normal anatomy.
In this way, the ’ideal’ segmentation result is shown. This will be compared to the
segmentation results for the stroke patient.
The ideal confusion matrix has high (bright) values for the diagonal elements,
corresponding to a high percentage of correctly segmented voxels for that tissue
class. Following similar reasoning, lower (darker) values in the off-diagonal elements is ideal, corresponding to a lower percentage of wrongly segmented voxels
in the automatic segmentation routine. The image intensity of the confusion matrices is gamma corrected (γ = 0.5) to increase the resolution for matrix elements
with a lower (darker) image intensity. This results in a non-linear treatment of
the image intensity as explained earlier.
The confusion matrices for the stroke patient, seen in Figure 5.5, are far from
the ideal case. The original segmentation in SPM8 and ALI are way off since there
is hardly any high values in the diagonal. The CSF seems to be the only tissue
that is segmentated in an acceptable manner. A better result is gained when using
6.1 Interpretation of the Results
51
the new segmentation routine in SPM8 and the result can be improved further by
using Chris Rorden’s TPMs, since the diagonal elements have a higher value and
the off-diagonal elements have a lower value for that combination. It is clear that
the lesion is segmented as CSF for all automatic segmentation routines, which can
be seen in the high values in the matrix elements corresponding to truth lesion predicted CSF.
The confusion matrix for the healthy patient, seen in Figure 5.6, is far better
than the one for the stroke patient following earlier reasoning. The results from the
original segmentation routine in SPM8 and ALI are still not perfect, despite the
high quality T1 weighted MRI used in the segmentation. The new segmentation
routine in SPM8, regardless of which set of TPMs used, produces an almost perfect
segmentation result, which can clearly be seen in the confusion matrix.
For both patients it is clear that the original segmentation routine and ALI do
not perform as well as the new segmentation routine. One thing to keep in mind
though is that neither the original segmentation routine nor ALI segments bone or
skin as own tissue classes. This explains the low values for these two classes. Also,
there is not an own class for the lesion implemented in the original segmentation
routine or the new segmentation routine. Hence, the voxels corresponding to
these tissue classes are included in the ’other’ class, thus creating an almost black
intensity at those matrix elements.
Error Rates
Two patients are used for the calculated error rates, one stroke patient and one
healthy patient as the ideal case. An Error rate for each automatic segmentation
routine and tissue is calculated and illustrated as the first line in Figure 5.7. A
low value corresponds to a low error, thus a correct segmentation of the tissue
type. The image intensity of the Error rate plots is gamma corrected (γ = 0.5)
to increase the resolution for matrix elements with a lower (darker) image intensity. This results in a non-linear treatment of the image intensity as explained
earlier. The second line in the same figure represents the Total error rate for each
automatic segmentation routine. This rate will be used to determine which automatic segmentation routine performs the best for each patient. A lower percentage
corresponds to a smaller error, thus a better segmentation result.
The new segmentation in SPM8 produces a better segmentation result than
the original segmentation routine in SPM8 and ALI for both patients. The original
segmentation routine and ALI appear to have problems with the WM and GM
classes, but the result for the CSF class is very good. The high error rate in the
skin class, the bone class, and the lesion class of the original segmentation routine
is due to the routine not having separate classes for these tissue types. This is
also true for ALI, except that ALI has an own tissue class for the lesion. The
brightness of the lesion class is instead due to ALI not classifying the lesion voxels
as part of the lesion class, but instead as CSF. The values for GM and WM are
almost perfect for the new segmentation routine, regardless of which set of TPMs
used. The results for skin tissue and bone tissue are also good which can be seen
by the gray color of the matrix elements. Since the new segmentation routine does
52
Discussion
not have an own tissue class for the lesion, there is a high error for this tissue class
and segmentation routine also.
It is quite clear that the segmentation result is better when the new segmentation routine in SPM8 is used with Chris Rorden’s TPMs, since about 50% of all
voxels are correctly segmented for the stroke patient. The high Total error rate of
the original segmentation routine in SPM8 and ALI are due to non-existing classes
for skin tissue and bone tissue, along with the abscense of a lesion class (original
segmentation routine in SPM8) and incorrect classification of the lesion (ALI).
The best results for both patients are gained when using the new segmentation
routine in SPM8 with Chris Rorden’s TPMs.
6.1.2
Regions of Interest (ROIs)
This section includes an evaluation of the performance of the segmentation routines
for six stroke patients, three with acute stroke lesions and three with necrotic
stroke lesions. A comparison is made between the automatic segmentation routines
against ground truth when possible. If not, a comparison between the available
segmentation routines is made. The new segmentation routine in SPM8 is known
to produce a more accurate result and the other automatic segmentation routines
will therefore be compared against it.
Keith
A comparison between the results from the segmentation routines can be seen in
Figure 5.8. The manual segmentation for Keith is not entirely correct since the
lesion is not segmented as an own class for this patient. Because the lesion is
classified as open and the outer border of it is unknown, a manual segmentation of
the lesion only is impossible to perform. Hence, the manual segmentation cannot
be used as ground truth for Keith. ALI, whose segmentation results can be seen
in Figure 3.2b, classifies a part of the lesion site as lesion, namely the enlarged left
ventricle along with some tissue within the left hemisphere. The rest of the lesion
site is classified as a mixture of GM, WM, and CSF. In this case, the segmentation
result is correct since the segmentation routine is based upon image intensity and
the image intensity of the voxels within lesion site is a mixture of the tissue classes
previously mentioned.
The result from the original segmentation routine in SPM8, seen in Figure
3.2a, is a bit different from the other automatic segmentation routines, since a
larger percentage of CSF is segmented compared to the others. The CSF is segmented as a thicker layer and includes some tissue that is not CSF in reality but
bone, which can clearly be seen by visual inspection of the outer part of the right
hemisphere. The result from the new segmentation routine in SPM8 using default
TPMs and Chris Rorden’s TPMs is similar compared to each other, although the
result using the later is considered more accurate since a larger part of the CSF is
segmented as CSF. Another difference is that the eyes are segmented as CSF when
using Chris Rorden’s TPMs. An illustration of the result from these segmentation routines using the mono-spectral mode can be seen in Figure 3.5a and Figure
6.1 Interpretation of the Results
53
3.7a. The multi-spectral mode in the new segmentation routine produces a similar
result as the mono-spectral mode, although by looking at the segmented CSF it
is clear that the segmentation routine performs badly since the CSF layer is not
continuous. The CSF layer in reality is continuous since the fluid covers the entire
region between the brain and the inside of the skull. The discontinuity appears
in the results using both default TPMs and Chris Rorden’s TPMs, although the
segmentation result using the later is better. An example of a segmentation result
can be seen in Figure 3.5b and Figure 3.7b.
The manual segmentation does not sum up to 100% which can be explained
by voxels in the ROI not corresponding to any of the tissue classes used in the
comparison. Those voxels can for example be segmented as bone instead since a
tiny part of bone is included in the ROI.
Looking at all segmentation results, the multi-spectral mode in the new segmentation routine using Chris Rorden’s TPMs produces the best segmentation
result for this patient, although there is a problem with the discontinuity of the
CSF layer. It is recommended to use this routine but with a slight post-processing
of the result afterwards. Unfortunately ALI, which is designed to handle stroke
patients, does not perform well enough since it segments some tissue incorrectly
and misses to classify a large part of the left hemisphere at the lesion site.
LeftMca
A comparison between the automatic segmentation routines can be seen in Figure
5.9. There is no ground truth available for this patient, thus only a comparison
between the automatic segmentation routines will be made. All segmentation
routines produce a similar segmentation result regarding the amount of GM, WM,
and CSF. The original segmentation routine and ALI are the ones that differ
from the others. ALI does not perform well since it fails to classify the lesion area
as part of the lesion class. This can be seen by visual inspection of the lesion
class which is empty after the segmentation is performed. It seems like most of
the lesion voxels instead are classified as CSF, which would explain the increased
percentage of CSF. Unfortunately, ALI does not perform well here either since a
part of the left hemisphere is not segmented at all, similar to the case with Keith.
The original segmentation routine in SPM8 also differs a bit from the others,
this is since the CSF layer is segmented a bit thicker than it is in reality. This
explains the increased percentage of CSF for this patient also. The result from the
new segmentation routine using the default TPMs and Chris Rorden’s TPMs is
similar, which is expected since the only real difference in the segmentation is the
probability of finding different tissues at different locations within the brain. The
problem with discontinuity of the CSF layer still lingers, making the mono-spectral
mode a better choice than the multi-spectral mode.
Comparing the results from the different automatic segmentation routines, the
multi-spectral mode in the new segmentation routine using Chris Rorden’s TPMs
produces the best result for this patient also. This routine together with postprocessing to ensure a continuous layer of CSF is recommended.
54
Discussion
SmallTemporal
A comparison between the automatic segmentation routines can be seen in Figure
5.10. There is no ground truth available for this patient, thus only a comparison
between the automatic segmentation routines will be made. The results from the
segmentation routines are similar, the real difference is when using the original
segmentation routine in SPM8 and ALI. The increase in CSF percentage for ALI
is due to the wrong classification of GM and WM as CSF, which can be seen by
visually inspecting the segmentation result (not included in this report). The same
error is also present in the result from the original segmentation routine in SPM8.
Unfortunately, ALI does not perform well here either since it fails to classify the
lesion as lesion, which can be explained by the lesion being classified as necrotic.
The small amount of CSF in the new segmentation routine using the default
TPMs and Chris Rorden’s TPMs is due to the discontinuity of the CSF mentioned
before. Only a fraction of the total amount of CSF at the lesion site is segmented
as it, although using Chris Rorden’s TPMs seems to produce a better result. The
result for both mono-spectral modes does not sum upp to 100%, since there are
some voxels in the ROI that are not segmented as any of the three tissue classes
used in the comparison.
A comparison between the automatic segmentation routines points towards the
use of the multi-spectral mode in the new segmentation routine using Chris Rorden’s TPMs, although post-processing is needed to ensure a correct segmentation
result.
3316
A comparison between the automatic segmentation routines can be seen in Figure
5.11. There is no ground truth available for this patient, thus only a comparison
between the automatic segmentation routines will be made. Since only a T1
weighted MRI was available for this patient, multi-spectral mode could not be
used for the new segmentation routine in SPM8. As before, the new segmentation
routine in SPM8 using default TPMs and Chris Rorden’s TPMs produce almost
equal results, in this case the results are almost identical. Some parts of the lesion
site are wrongly segmented though and in the tissue posterior of the lesion site a
part of the WM is wrongly segmented as GM. This can be seen by visual inspection
of the segmentation result (not included in this report).
The original segmentation routine in SPM8 produces a better result at the
lesion site, although parts of the CSF layer does not correspond to what is seen in
the T1 weighted MRI. ALI classifies the white border of the lesion as part of the
lesion class since the image intensity is abnormal at that location. The inside of
the white ring is classified as GM, which is correct since the segmentation is based
on image intensity, although it should be included in the lesion class as it is a part
of the lesion. ALI and the original segmentation routine classifies the posterior
part of the brain in a better way than the new segmentation routine. The tiny
amount of CSF according to ALI is more accurate than the amount according to
the other routines, since there should only be a small amount of CSF in the ROI.
6.1 Interpretation of the Results
55
The best segmentation routine for this kind of lesion is ALI, since the segmentation of the lesion site is more correct and the routine segments the posterior part
of the brain in a better way than the other routines. ALI also segments the lesion
as an own class which is the whole idea. A small amount of post-processing should
be applied to be able to use the segmentation results in a FEM simulation later.
3319
A comparison between the automatic segmentation routines can be seen in Figure
5.12. There is no ground truth available for this patient, thus only a comparison
between the routines will be made. Since only a T1 weighted MRI was available
for this patient, multi-spectral mode could not be used for the new segmentation
routine in SPM8. By visual inspection of the segmentation results (not included
in this report), almost all automatic segmentation routines have problems with
this patient. This might be since the patient seems to have had strokes before
this acute one, making the segmentation a bit harder since the brain anatomy is
very different from a normal brain. The new segmentation routine in SPM8, using
either of the sets of TPMs, produces a poor segmentation result since it does not
segment the posterior part of the brain correctly. Otherwise this routine performs
decent regarding the percentage of tissues in the ROI. The entire lesion area is
segmented as WM, which is correct since the segmentation routine is based upon
image intensity. To be able to use the segmentation results in a FEM simulation
though, the lesioned needs to be marked as an own class and not as WM due to
the difference in tissue properties between the two.
The original segmentation routine in SPM8 segments a large part of the brain
tissue as CSF which is incorrect. There is a quite high percentage of CSF in the
ROI according to the original segmentation routine, which is incorrect since the
lesion site is mostly made up of brain tissue. ALI performs well enough on this
patient since the entire lesion is included in the lesion site, although parts of the
brain tissue is wrongly segmented as CSF.
By comparing the segmentation routines to each other, the best automatic
segmentation routine for this patient is ALI (at least at the lesion site) since it
classifies the entire lesion as a part of the lesion class. At other locations, the
new segmentation routine performs better. The best solution would be to use
the new segmentation routine with Chris Rorden’s TPMs and do some light postprocessing, including marking the lesion area by hand.
3322
A comparison between the automatic segmentation routines can be seen in Figure
5.13. There is no ground truth available for this patient, thus only a comparison
between the routines will be made. Since only a T1 weighted MRI was available
for this patient, multi-spectral mode could not be used for the new segmentation
routine in SPM8. ALI is not able to classify the lesion area as a part of the
lesion class and instead classifies the area as GM. Following the reasoning about
image intensity earlier, this is correct although it would be nice to have an own
tissue class for the lesion due to the change in electrical conductivity in the tissue
56
Discussion
following a stroke. The original segmentation routine in SPM8 and ALI segments
a high percentage of CSF in the ROI. This is wrong since the ROI mostly includes
brain tissue, which can be seen by visual inspection of the ROI.
Following the reasoning presented, the new segmentation routine in SPM8,
regardless of which set of TPMs used (default or Chris Rorden’s), produces a
more accurate segmentation result than the other segmentation routines. By visual
inspection the result also looks more accurate. This setup, together with some light
post-processing afterwards, including a manual marking of the lesioned area as an
own tissue class to ensure proper segmentation result, is recommended.
6.2
Conclusions
There are a number of conclusions that can be made about the segmentation
routines evaluated. The conclusions made are summarized for each segmentation
routine in Section 6.2.1 - 6.2.4. The main conclusion is that the new segmentation routine in SPM8 is superior compared to the original segmentation routine
in SPM8 and ALI. If used together with Chris Rorden’s improved TPMs, the segmentation result is better and more reliable. Therefore, the recommendation is to
use the new segmentation routine in SPM8 together with Chris Rorden’s TPMs.
6.2.1
SPM8 Original Segmentation
The original segmentation routine in SPM8 does not always produce a correct
segmentation result. This can for example be seen for Keith where a part of the
skull is wrongly segmented as CSF. Since this segmentation routine does not even
produce an acceptable segmentation result for the healthy patient, the conclusion
is that the original segmentation routine is not good enough to use on patients with
stroke lesions. Instead, the improvements made on it should be used, including
the new segmentation routine in SPM8 and Chris Rorden’s TPMs.
6.2.2
SPM8 New Segmentation
The mono-spectral mode produces an acceptable segmentation result, although
there are errors present. For example, the CSF layer is not continuous which is
totally wrong from an anatomical point of view. The new segmentation routine
is better than the original segmentation routine, although there are possible improvements that can be made. Since this routine is only work in progress, the
recommendation is to continue working on it to improve it even more.
The multi-spectral mode is a nice implementation, using information from more
than one image weighting during the segmentation should help the segmentation
routine to produce a more accurate result. The segmentation result is better in
some ways, although the major problem here is the discontinuity of the CSF layer
(which is also present when using mono-spectral mode). It is hard to use the results
from this segmentation routine to simulate tDCS using FEM since the tissues are
not correctly segmented.
6.3 Future Improvements
57
If this routine is used together with manual post-processing, this will be the
best choice for patients with stroke lesions, especially if it is used together with
Chris Rorden’s improved TPMs. The post-processing can for example include
improvement of the CSF layer or manual marking of the lesion area. This combination will be fast and produces a reliable result. Although the ideal would be to
improve this automatic segmentation routine until it provides a reliable result for
both healthy patients and stroke patients.
6.2.3
SPM8 Chris Rorden’s TPMs
The result when using Chris Rorden’s TPMs is similar to when using the default
dito. The quality of the segmentation result is higher thanks to the improved
TPMs. The segmentation result will be better in general when using Chris Rorden’s TPMs. Thus, using these TPMs is a first step towards a more reliable
segmentation result.
6.2.4
Mohamed Seghier’s ALI
Since ALI is developed for patients with stroke lesions, a better segmentation result
was expected. The segmentation result is slightly better for patients with acute
stroke lesions, but for necrotic lesions the lesion site is often classified as CSF. The
definition of a lesion in this segmentation routine is simply just abnormal voxels
in GM and WM segments, perhaps that is why some lesions are not segmented as
part of the lesion class. Paralesional areas are often poorly segmented, if segmented
at all, and there is a large amount of voxels incorrectly segmented. For example, a
large number of GM and WM voxels are incorrectly segmented as CSF which is a
massive error. For patients with larger necrotic stroke lesions (Keith and LeftMca)
a large part of the hemisphere where the lesion if located is often left out of the
segmentation or is segmented incorrectly.
The final conclusion for this segmentation routine is that it is a good idea to
implement an own class for the lesion, although the performance of this routine is
often bad. This idea should be improved and a new segmentation routine should
be developed which can handle a wider range of stroke lesions.
6.3
Future Improvements
There are numerous possible improvements for all segmentation routines evaluated,
some proposed by the maker and some proposed based on the conclusion in this
thesis. No future improvements will be suggested for Chris Rorden’s TPMs. The
suggested improvements are listed in Section 6.3.1 - 6.3.4.
6.3.1
SPM8 Original Segmentation
There are numerous things that could be improved in the original segmentation
routine, for example the TPMs used, the mixture proportion of the image intensities for each tissue, a multi-spectral mode to ensure that more information is used
58
Discussion
in the segmentation, and so on. These improvements have already been made and
are a part of the new segmentation routine. Therefore, no more improvements
will be proposed for this routine since the new segmentation routine in SPM8 is
assumed to perform better than the original segmentation routine in SPM8.
6.3.2
SPM8 New Segmentation
Both the mono-spectral mode and the multi-spectral mode in the new segmentation routine produce a poor segmentation result, e.g. since the segmented CSF is
not a continuous layer and since some voxels in the outer part of the brain tissue
are incorrectly segmented. When Chris Rorden’s TPMs are used, the segmentation result is improved, although the result is still not entirely correct. A new and
improved version, including an improved mono-spectral mode and an improved
multi-spectral mode, is needed to ensure a correct segmentation result.
Other possible improvements for the new segmentation routine in SPM8 are
solutions to the problems presented in Section 3.3.2.
6.3.3
Mohamed Seghier’s ALI
ALI can be improved by implementing a multi-spectral version of it or by evaluating it with MRIs of other image weightings than just the T1 weighting. Since more
information about the lesion site can be found in for example DWIs, this could be
one solution for patients with acute stroke lesions. Also, a new definition of the
lesion class can be implemented, for example including more than just abnormal
voxels in the GM and WM segment.
Other possible improvements for ALI are solutions to the problems presented
in Section 3.2.2.
6.3.4
Manual Segmentation
By implementing a certainty measure for each manually segmented voxel, corresponding to how certain the operator is of the current voxel corresponding to the
current tissue class, the manual segmentation can be improved. This will be an
improvement since errors in automatically segmented voxels where the operator is
more certain is more critical than errors where the operator is less certain.
A lower certainty can for example be implemented for voxels corresponding
to the bottom of the skull cavity where there is a possible mix-up between what
should be air and skull and voxels corresponding to the nasal cavity where there
is a possible mix-up between bone, soft tissue, and air. A higher certainty can be
implemented for voxels corresponding to the middle and top portion of the head,
since the only thing that really changes between axial slices is the CSF.
Other possible improvements for the manual segmentation are solutions to the
problems presented in Section 3.5.1.
Bibliography
[1] Wassermann E, Epstein AM, and Ziemann U. The Oxford handbook of transcranial stimulation. Oxford University Press, 2008.
[2] Boron WF and Boulpaep EL. Medical Physiology. 2nd ed. Philadephia:
Saunders Elsevier, 2009.
[3] iChristian School. Anatomy & physiology week 13. Available at: http://www.
ichristianschool.org/anatomy/2ndQuater/anatomy13.htm, May 6 2011.
[4] Mayfield Clinic. Anatomy of the brain. Available at:
mayfieldclinic.com/PE-anatbrain.htm, January 21 2011.
http://www.
[5] Tortora G and Derrickson B. Principles of Anatomy and Physiology. 12th ed.
Wiley, 2009.
[6] UIC Center for Cognitive Medicine. Traumatic brain injury. Available at:
http://ccm.psych.uic.edu/PatientInfo/TBIInfo.aspx, January 21 2011.
[7] Mattson J and Simon M. The pioneers of NMR and magnetic resonance in
medicine: The story of MRI. Bar-Ilan University Press, 1996.
[8] Darling D. The internet encyclopedia of science. Available at: http://www.
daviddarling.info/encyclopedia/S/stroke.html, January 21 2011.
[9] Datta A, Baker JM, Bikson M, and Fridriksson J. Individualized model predicts brain current flow during transcranial direct-current stimulation treatment in responsive stroke patient. Brain Stimulation, In press, 2011.
[10] Rossini PM, Rossini L, and Ferreri F. Brain-behavior relations transcranial
magnetic stimulation: A review. IEEE Eng Med Biol Mag, 29(1):84–95, 2010.
[11] Datta A, Elwassif M, Battaglia F, and Bikson M. Transcranial current stimulation focality using disc and ring electrode configurations: Fem analysis.
Journal of neural engineering, 5(2):163–174, 2008.
[12] Datta A, Bansal V, Diaz J, Patel J, Reato D, and Bikson M. Gyri-precise head
model of transcranial direct current stimulation: Improved spatial focality
using a ring electrode versus conventional rectangular pad. Brain stimulation,
4(2):201–207, 2009.
59
60
Bibliography
[13] Minhas P, Bansal V, Patel J, Ho J, Diaz J, Datta A, and Bikson M. Electrodes
for high-definition transcutaneous dc stimulation for applications in drug delivery and electrotherapy, including tdcs. Journal of Neuroscience Methods,
190(2):188–197, 2010.
[14] Reato D, Rahman A, Bikson M, and Parra LC. Low-intensity electrical stimulation affects network dynamics by modulating population rate and spike
timing. Journal of Neuroscience Methods, 30(45):15067–15079, 2010.
[15] Bikson M, Bulow P, Stiller JW, Datta A, Battaglia F, Karnup SV, and Postolache TT. Transcranial direct current stimulation of major depression: a
general system for quantifying transcranial electrotherapy dosage. Current
Treatment Options in Neurology, 10(5):377–385, 2008.
[16] Fridriksson J, Richardson JD, Baker JM, and Rorden C. Transcranial direct current stimulation improves naming reaction time in fluent aphasia: A
double-blind, sham-controlled study. Stroke, 42(3):819–821, 2011.
[17] Baker JM, Rorden C, and Fridriksson J. Using transcranial direct currentstimulation to treat stroke patients with aphasia. Stroke, 41(6):1229–1236,
2010.
[18] Bolognini N, Pascual-Leone A, and Fregni F. Using non-invasive brain stimulation to augment motor training-induced plasticity. Journal of NeuroEngineering and Rehabilitation, 6(8), 2009.
[19] Eklund A, Andersson M, and Knutson H. MRI, fMRI, Image Registration,
Image Segmentation. TBMI02 Medical Image Analysis: Course Compendium,
2010.
[20] Ashburner J. Computational anatomy with spm software. Magnetic resonance
imaging, 27(8):1163–1174, 2009.
[21] Seghier AM, Ramlackhansingh A, Crinion J, Leff AP, and Price CJ. Lesion
identification using unified segmentation-normalisation models and fuzzy clustering. NeuroImage, 41(4):1253–1266, 2008.
[22] The Wellcome Trust Centre for Neuroimaging. Statistical parametric mapping. Available at: http://www.fil.ion.ucl.ac.uk/spm/, January 31 2011.
[23] Ashburner J and Friston KJ. Unified segmentation. NeuroImage, 26(3):839–
851, 2005.
[24] UCLA 3T MRI lab. The ac-pc line. Available at: http://airto.hosted.
ats.ucla.edu/BMCweb/HowTo/AC-PC.html, May 10 2011.
Fly UP