Automated brain lesion classdification method for diffusion-weighted magnetic resonance images.
AUTOMATED BRAIN LESION CLASSIFICATION METHOD FOR
DIFFUSION-WEIGHTED MAGNETIC RESONANCE IMAGES
NORHASHIMAH BINTI MOHD SAAD
UNIVERSITI TEKNOLOGI MALAYSIA
UNIVERSITI TEKNOLOGI MALAYSIA
PSZ 19: 16 (Pind. 1/07)
DECLARATION OF THESIS/ UNDERGRADUATE PROJECT PAPER AND COPYRIGHT
Author's full name
: NORHASHIMAH BINTI MOHD SAAD
Date of birth
: 24 DECEMBER l 980
セ@
Title
: AUTOMATED BRAIN LESION CLASSIFICATION METHOD FOR
DIFFUSION-WEIGHTED MAGNETIC RESONANCE IMAGES
Academic Session
: 2014/2015-11
セ@
I declare that this thesis is classified as :
D
D
GJ
CONFIDENTIAL
(Contains confidential information under the Official Secret
Act 1972)'
RESTRICTED
(Contains restricted information as specified by the
organisation where research was done)'
OPEN ACCESS
I agree that my thesis to be published as online open access
(full text)
I acknowledged that Universiti Teknologi Malaysia reserves the right as follows :
I. The thesis is the property of Universiti Teknologi Malaysia.
2. The Library of Universiti Teknologi Malaysia has the right to make copies for the purpose
of research only.
3. The Library has the right to make copies of the thesis for academic exchange.
Certified by :
SIGNATURE OF SUPERVISOR
801224-10-5670
(NEW JC NO. /PASSPORT NO.)
Date: 3•d of June 2015
NOTES
•
PM DR. SYED ABD. RAHMAN B. SYED ABU BAKAR
NAME OF SUPERVISOR
Date: 3•d of June 2015
If the thesis is CONFIDENTIAL or RESTRICTED, please attach with the letter from
the organisation with period and reasons for confidentiality or restriction.
"We hereby declare that we have read this thesis and in our opinion this thesis is
sufficient in terms of scope and quality for the purpose of awarding the degree of
Doctor of Philosophy (Electrical Engineering)"
Signature
PM. Dr. N⦅セL@
3rd of June 20 5
Name of Supervisor I
Date
Rahman b. Syed Abu Bakar
Signature
Name of Supervisor II : PM. r. Ahmad Sobri bin Muda
3rd of June 2015
Date
Signature
:
...
セᄋ@
Name of Supervisor III: Dr. Musa b. Mohd Mokji
Date
3rd of June 2015
BAHAGIAN A - Pengesahan Kerjasama*
Adalah disahkan bahawa projek penyelidikan tesis ini telah dilaksanakan melalui ke1jasama
antara _ _ _ _ _ _ _ _ _ _ dengan _ _ _ _ _ _ _ _ __
Disahkan oleh:
Tandatangan
Tarikh:
Nama
Jawatan
(Cop rasmi)
* Jika penyediaan tesislprojek melibatkan ke1jasama.
BAHAGIAN B - Untuk Kegunaan Pejabat Sekolah Pengajian Siswazah
Tesis ini telah diperiksa dan diakui oleh:
Nama dan Alamat Pemeriksa Luar
Prof. Madya Dr. M. Igbal Saripan
Department of Computer & Communication
Systems Engineering,
Faculty of Engineering,
Universiti Putra Malaysia,
43400 Serdang,
Selangor.
Nama dan Alamat Pemeriksa Dalam
Prof. Madya Dr. Rubita binti Sudirman
Fakulti Kejuruteraan Elektrik,
UTM Johor Bahru
Disahkan oleh Timbalan Pendaftar di Sekolah Pengajian Siswazah:
Tarikh:
Tandatangan :
Nama
ASRAM BIN SUlAIMAN @ SAIM
i
AUTOMATED BRAIN LESION CLASSIFICATION METHOD FOR
DIFFUSION-WEIGHTED MAGNETIC RESONANCE IMAGES
NORHASHIMAH BINTI MOHD SAAD
A thesis submitted in fulfilment of the
requirements for the award of the degree of
Doctor of Philosophy (Electrical Engineering)
Faculty of Electrical Engineering
Universiti Teknologi Malaysia
JUNE 2015
11
I declare that this thesis entitled "Automated Brain Lesion Classification Method for
Diffusion-Weighted Magnetic Resonance Images" is the result of my own research
except as cited in the references. The thesis has not been accepted for any degree
and is not concurrently submitted in candidature of any other degree.
セ@
:
.. .......... .
Signature
: ..
Name
: Norhashimah binti Mohd Saad
Date
: 3rd
of June 2015
iii
Specially dedicated to:
My parents for their priceless supports and generous prayers.
iv
ACKNOWLEDGEMENT
First of all, I am grateful to the Almighty God for giving me strength to
endure many obstacles in life. Many times, His continuous blessings and protections
have revived my determination and hope in completing my study.
I would like to take this opportunity to thank my main supervisor, Associate
Professor Dr. Syed Abdul Rahman bin Syed Abu Bakar (UTM), my second
supervisor Associate Professor Dr. Ahmad Sobri bin Muda (Department of
Radiology, Pusat Perubatan UKM) and my third supervisor, Dr. Musa bin Mohd
Mokji (UTM) for their supports and guidance throughout the duration of my Ph.D.
research.
Thank you to all colleagues in Computer Vision Video Image Processing
(CVVIP) research group for the valuable ideas and encouragement. I am fortunate to
have received many useful comments and suggestions for my work. My deep
gratitude also goes to the colleagues in UTeM especially to Dr. Abdul Rahim bin
Abdullah, Dr. Syafeeza Ahmad Radzi for their endless help and in securing grants to
support the continuation of my research. My appreciation also extends to Ms. Aslina
binti Abdul Rahman for her moral support. Not to forget, special recognitions are
dedicated to the Study Leave Unit in UTeM and Dean of FKEKK.
Last but not the least I would like to convey my grateful appreciation to my
dear family and friends for their unceasing love and encouragement throughout the
duration of my study. I have met such inspiring people along this incredible journey.
v
ABSTRACT
Diffusion-weighted magnetic resonance imaging plays an increasingly
important role in the diagnosis of several brain diseases by providing detailed
information regarding lesion based on the diffusion of water molecules in brain
tissue. Conventionally, the differential diagnosis of brain lesions is performed
visually by professional neuroradiologists during a highly subjective, timeconsuming process. Within this context, this study proposes a new technique for
automatically detecting and classifying major brain lesions of four types: acute
stroke, chronic stroke, tumor and necrosis. An analytical framework of the brain
lesions consists of four stages which are pre-processing, segmentation, features
extraction and classification. For segmentation process, adaptive thresholding, gray
level co-occurrence matrix, region splitting and merging, semi-automatic region
growing, automatic region growing and fuzzy C-means were proposed to segment
the lesion region. The algorithm performance was then evaluated using Jaccard
index, Dice index, and both false positive and false negative rates. Results
demonstrated that automatic region growing offered the best performance for lesion
segmentation while acute stroke gave the highest rate with 0.838 Dice index. Next,
statistical features were extracted from the region of interest and fed into the rulebased classifier designed to the best suit to the lesion’s features. The performance of
the classifier was evaluated based on overall accuracy, sensitivity and specificity.
The overall accuracy for the classification was 81.3%. In conclusion, the proposed
automated brain lesion classification method has the potential to diagnose and
classify major brain lesions.
vi
ABSTRAK
Pengimejan magnetik resonan pemberat-resapan memainkan peranan yang
semakin penting dalam mendiagnosis beberapa penyakit otak dengan memberikan
maklumat terperinci berkenaan perbezaan jelas lesi ke atas resapan molekul air di
dalam tisu otak. Secara konvensional, perbezaan diagnosis lesi otak dilaksanakan
secara visual oleh pakar neuroradiologi profesional dengan proses subjektif serta
memakan masa yang lama. Dalam konteks ini, kajian ini mengusulkan teknik terbaru
untuk mengesan dan mengkelaskan lesi otak utama yang terdiri daripada empat jenis:
strok akut, strok kronik, tumor dan nekrosis. Analisis rangka kerja bagi lesi otak
terdiri daripada empat peringkat iaitu pra-pemproses, pengsegmenan, pengekstrakan
ciri dan pengkelasan. Untuk proses pengsegmenan, teknik ambang adaptif, matrik
gray level co-occurrence, rantau pemisahan dan penggabungan, rantau berkembang
separa automatik, rantau berkembang automatik dan fuzzy C-means dicadangkan
untuk mensegmen rantau lesi. Prestasi algoritma kemudiannya dinilai menggunakan
indeks Jaccard, indeks Dice, dan kedua-dua kadar palsu positif dan negatif.
Keputusan menunjukkan teknik rantau berkembang automatik memberikan prestasi
terbaik untuk pengsegmenan lesi sementara strok akut memberikan kadar indeks
Dice tertinggi iaitu 0.838. Kemudian, ciri-ciri statistik diekstrak daripada rantau
tarikan dan diinputkan kepada pengelas berasaskan peraturan yang telah direka untuk
disesuaikan dengan ciri lesi. Prestasi pengelas dinilai berdasarkan kejituan
keseluruhan, kepekaan dan kekhususan. Kejituan keseluruhan untuk pengkelasan
adalah 81.3%. Sebagai kesimpulannya, teknik klasifikasi automatan lesi otak yang
dicadangkan mempunyai potensi untuk mendiagnosis dan mengklasifikasi lesi otak
utama.
vii
TABLE OF CONTENTS
CHAPTER
1
2
TITLE
PAGE
DECLARATION
ii
DEDICATION
iii
ACKNOWLEDGEMENT
iv
ABSTRACT
v
ABSTRAK
vi
TABLE OF CONTENTS
vii
LIST OF TABLES
xi
LIST OF FIGURES
xiii
LIST OF ABBREVIATIONS
xvi
LIST OF SYMBOLS
xviii
LIST OF APPENDICES
xix
INTRODUCTION
1
1.1
Introduction
1
1.2
Problem Statement
2
1.3
Objectives
4
1.4
Scope of Work
5
1.5
Proposed System Design
6
1.6
Thesis Contributions
8
1.7
Thesis Outline
9
LITERATURE REVIEW
11
2.1
Introduction
11
2.2
The Human Brain
11
2.3
Neuroimaging Techniques
13
viii
2.3.1 Magnetic Resonance Imaging
15
2.3.2 Diffusion-Weighted Imaging
17
2.3.2.1 Pulse Sequence of DiffusionWeighted Imaging
18
2.3.2.2 Advantages of DiffusionWeighted Imaging
2.3.2.3 Types of Brain Lesions
2.4
2.5
2.7
3
22
Computer-Aided Detection and Diagnosis for
Brain Lesions
26
2.4.1 Stroke Lesion Analysis Techniques
26
2.4.2 Tumor Lesion Analysis Techniques
29
Diffusion-Weighted Imaging Analysis
Techniques
2.6
20
31
Brain Lesion Segmentation and Classification
Techniques
34
Summary
37
METHODOLOGY FOR LESION DETECTION,
SEGMENTATION AND CLASSIFICATION
41
3.1
Introduction
41
3.2
Data Collection
42
3.3
Image Pre-Processing
46
3.3.1 Image Normalization
47
3.3.2 Background Removal
47
3.3.3 Image Enhancement
49
3.4
3.3.3.1 Gamma-Law Transformation
49
3.3.3.2 Contrast Stretching
50
Segmentation Analysis Techniques
51
3.4.1 Adaptive Thresholding
53
3.4.2 Gray Level Co-Occurrence Matrix
59
3.4.2.1 Co-Occurrence Histogram
62
3.4.2.2 Minimum and Maximum
Threshold
65
ix
3.4.2.3 Region and Boundary
Information
3.4.2.4 Optimal Threshold Values
67
3.4.3 Region Splitting and Merging
69
3.4.4 Automated Region Growing
76
4.2.5 Fuzzy C-Means
80
Features Extraction
85
3.5.1 First Order Statistic Features
86
3.5.2 Statistical Features Analysis
87
3.6
Rule-Based Classifier
90
3.7
Performance Assessment Matrices
93
3.5
4
66
RESULTS AND PERFORMANCE ANALYSIS
97
4.1
Introduction
97
4.2
Pre-processing Stage
98
4.3
Segmentation Stage
101
4.3.1 Adaptive Thresholding
101
4.3.2 Gray Level Co-occurrence Matrix
103
4.3.3 Automatic Region Growing
104
4.3.4 Fuzzy C-Means
107
4.4
Performance Evaluation of Segmentation
Techniques
4.5
108
Performance Evaluation of Lesion
Classification
118
4.5.1 Classification Results Validated with
Manual Reference Segmentation
119
4.5.2 Classification Results for Automated
System
4.6
Performance Benchmarking with Other
122
125
Techniques
4.6.1 Lesion Segmentation Benchmarking
125
4.6.2 Lesion Classification Benchmarking
128
x
5
CONCLUSIONS AND FUTURE WORK
130
5.1
Conclusions
130
5.2
Contributions
133
5.3
Recommendations
133
REFERENCES
135
Appendices A-D
149-159
xi
LIST OF TABLE
TABLE NO.
TITLE
PAGE
2.1
Major parts of the brain
12
2.2
Comparison of reliabilities in neuroimaging modalities
21
2.3
Lesion in DWI and ADC image
24
2.4
Description of brain lesions, types, symptoms and
pathological findings
2.5
25
Summary of methods for DWI segmentation and
classification
39
3.1
Summary of dataset for segmentation
43
3.2
Summary of training and validation dataset for features
extraction and classification
44
3.3
Gray level description from DWI raw data
46
3.4
Types of intensity label according to hyperintensity or
hypointensity lesions
46
3.5
Average optimal threshold values for hyperintense lesions
58
3.6
Range of brain regions from GLCM
65
3.7
Intensity of lesions based on histogram
71
3.8
Threshold value of selected homogeneity criteria
72
3.9
Statistical features for region level 2
73
3.10
Statistical features for region level 3
74
3.11
Feature extraction comparison with medical diagnosis
85
3.12
Performance evaluation for classification of Class A
96
4.1
Segmentation results using adaptive thresholding
102
4.2
Performance evaluation of each lesion
103
4.3
Segmentation results using GLCM
104
4.4
Segmentation results using region splitting and merging
105
4.5
Segmentation results using automatic region growing
106
xii
4.6
Segmentation results using FCM
108
4.7
Classifier input with 30 % training data
118
4.8
Classifier input with 50 % training data
118
4.9
Classifier input with 70 % training data
119
4.10
Classifier input with 90 % training data
119
4.11
Confusion matrix for 30 % training data
120
4.12
Confusion matrix for 50 % training data
120
4.13
Confusion matrix for 70 % training data
120
4.14
Confusion matrix for 90 % training data
120
4.15
Overall classification accuracy with different training data
validated with manual reference segmentation
4.16
Confusion matrix for 50 % training data for fully
automatic classification system
4.17
121
123
Evaluation results for proposed fully automatic
region growing
125
4.18
Performance comparison of stroke segmentation
127
4.19
Classification results using rule-based classifier
128
4.20
Comparison with previous researchers
129
xiii
LIST OF FIGURES
FIGURE NO.
TITLE
PAGE
1.1
Proposed system design
6
2.1
Major parts of the human brain
12
2.2
Electromagnetic spectrum in medical imaging
14
2.3
Schematic diagram of an MRI machine
16
2.4
CT and conventional MRI of patient with tumor
17
2.5
DWI pulse sequence
18
2.6
Acute stroke
20
2.7
Average sensitivity by different modalities for
neuroimaging of acute stroke
23
3.1
Flow chart of DWI image analysis
43
3.2
DWI images of the brain and lesions indicated by
neuroradiologists
45
3.3
Intensity description from image dataset
45
3.4
DWI of solid tumor
49
3.5
Image response for gamma-law transformation
50
3.6
Input and output response for contrast stretching
51
3.7
Flowchart of the proposed adaptive thresholding
technique
54
3.8
32 × 32 pixels per macro-block region
55
3.9
Histogram distribution of each macro-block region
55
3.10
Normal and abnormal histogram of macro-block regions
56
3.11
Superimposed histogram
57
3.12
Gradient function to determine the optimal threshold
57
3.13
Brain lesion segmentation using adaptive thresholding
59
3.14
Eight nearest-neighbor pixel cells to pixel
60
xiv
3.15
Segmentation process using GLCM
61
3.16
GLCM for DWI
62
3.17
Plot of co-occurrence histogram
64
3.18
Image cluster based on COH
65
3.19
GLCM cross-section and gradient function
66
3.20
Gray level region and boundary information
66
3.21
Example of GLCM segmentation
68
3.22
Flow chart of the splitting and merging algorithm
69
3.23
Example of the split and merge segmentation process
70
3.24
Region splitting at first stage shows split image of
region 3 and 4
3.25
Histogram of region 3 and 4, the arrow shows the
histogram of hyperintense lesion
3.26
71
Region splitting at second level with lesion apprearance
in region 4-1 and 4-3
3.27
71
73
Region splitting at third level with lesion apprerance in
region 4-1-4
74
3.28
Hyperintense lesions and their segmentation results
75
3.29
Segmentation results for hypointense lesion and normal
76
3.30
Flowchart of the proposed method
77
3.31
Region splitting and merging
79
3.32
Histogram from region slitting and merging
79
3.33
Hyperintense lesions and their segmentation results for
acute stroke
3.34
80
Hypointense lesions and their segmentation results for
chronic stroke
80
3.35
Flow chart of FCM segmentation
83
3.36
Flowchart of FCM clustering in this study
84
3.37
FCM segmentation
84
3.38
Features extraction process
88
3.39
Distributions of median and mean
89
3.40
Distributions of mode and standard deviation
89
3.41
Distributions of compactness and mean of boundary
89
3.42
code for a rule-based classifier
90
xv
3.43
Rule-Based classification process
91
3.44
Code for the classification process
92
3.45
Segmentation assessment indices
94
4.1
Flowchart of the analysis
98
4.2
DWI of acute stroke and histograms for pre-processing
stage
100
4.3
Jaccard index of the segmentation results
110
4.4
Average Jaccard index for each technique
110
4.5
FPR of the segmentation results
111
4.6
Average FPR for each technique
112
4. 7
FNR of the segmentation results
113
4.8
Dice index for the segmentation results
114
4.9
Average Dice index
114
4.10
Average simulation time
115
4.11
Comparison of the segmentation techniques in
terms of similarity indices
4.12
Comparison of the segmentation techniques in
terms of error rates
4.13
117
Sensitivity of the lesion classification validated with
manual reference segmentation
4.14
117
121
Sensitivity of the lesion classification
for fully automatic classification system
122
4.15
Sensitivity of the lesion classification for 50 % training data 123
4.16
Specificity for 50 % training data
124
4.17
Overall classification accuracy for 50 % training data
124
4.18
Automatic region growing performances for 28 acute
stroke cases
126
xvi
LIST OF ABBREVIATIONS
ADC
-
Apparent Diffusion Coefficient
ANFIS
-
Adaptive Network Based Fuzzy Inference System
AO
-
Area Overlap
AS
-
Acute Stroke
BG
-
Background Image
CAD
-
Computer Aided Detection and Diagnosis
CS
-
Chronic Stroke
CSF
-
Cerebral Spinal Fluid
CT
-
Computed Tomography
DICOM
-
Digital Imaging and Communications in Medicine
DWI
-
Diffusion-Weighted Imaging
FCM
-
Fuzzy C-Means
FDA
-
Food and Drug Administration
FNR
-
Fast Negative Rate
FPR
-
Fast Positive Rate
GLCM
-
Gray Level Co-Occurrence Matrix
IBSR
-
Internet Brain Segmentation Repository
kNN
-
k-Nearest Neighbors
k-SOM
-
Kohonen Self-Organizing Map
MA
-
Misclassified Error
MAPE
-
Mean Absolute Percentage Error
MLP
-
Multi-Layer Perceptron
MRI
-
Magnetic Resonance Imaging
MS
-
Mean-Shift
xvii
NC
-
Necrosis
PD
-
Proton Density
PET
-
Positron Emitted Tomography
RBF
-
Radial Basis Function
ROI
-
Region of Interest
S
-
Sensitivity
SNR
-
Signal to Noise Ratio
Sp
-
Specificity
SPM
-
Statistical Parametric Mapping
ST
-
Solid Tumor
SVM
-
Support Vector Machine
TE
-
Time Echo
TN
-
True Negative
TP
-
True Positive
TR
-
Time Repetition
WHO
-
World Health Organization
WM/GM
-
White Matter/Gray Matter
xviii
LIST OF SYMBOLS
B0
-
Magnetic Field Strength
b
-
Diffusion Gradient
D
-
Diffusion Coefficient
E
-
Energy
eV
-
Electron Volts
Hz
-
Hertz
T
-
Tesla
r err
-
Absolute Error Ratio
γ
-
Gamma
µ
-
Mean
σ
-
Standard Deviation
xix
LIST OF APPENDICES
APPENDIX
TITLE
PAGE
A
Enhancement Performance Analysis
150
B
Segmentation Performance Evaluation
151
C
List of Publications
158
D
List of Awards
160
1
CHAPTER 1
INTRODUCTION
1.1
Introduction
Diffusion-weighted magnetic resonance imaging (DW-MRI or DWI) is
increasingly having an important role in the diagnosis of many brain diseases. This
medical imaging technique provides higher pathologic or lesion contrast based on
diffusion of water molecules in brain tissues, compared to conventional MRI. This
unique information of diffusion properties plays a key role in evaluating multiple
neurologic diseases, especially for stroke detection (Mukherji et al., 2002,
Holdsworth and Bammer, 2008). It also gives additional information for cerebral
diseases such as stages in neoplasm (cancer, tumor, and necrosis), infections, and
others.
DWI is considered as the most sensitive technique in detecting acute
infarction and is useful in giving details of the component of brain lesions (Mukherji
et al., 2002). In DWI, image intensity and contrast only depend on the strength of
diffusivity of tissue. Tissue with altered diffusion rates may appear with either
hyperintense or hypointense on a pixel basis, which is absent in healthy tissue. Such
information forms vital image characteristics that may lead to the classification of
several brain-related diseases.
Automatic brain lesion detection and classification is necessary to implement
successful therapy and treatment planning. Computer-aided detection and diagnosis
system (CAD) has become a major research subject in diagnostic radiology to assist
DIFFUSION-WEIGHTED MAGNETIC RESONANCE IMAGES
NORHASHIMAH BINTI MOHD SAAD
UNIVERSITI TEKNOLOGI MALAYSIA
UNIVERSITI TEKNOLOGI MALAYSIA
PSZ 19: 16 (Pind. 1/07)
DECLARATION OF THESIS/ UNDERGRADUATE PROJECT PAPER AND COPYRIGHT
Author's full name
: NORHASHIMAH BINTI MOHD SAAD
Date of birth
: 24 DECEMBER l 980
セ@
Title
: AUTOMATED BRAIN LESION CLASSIFICATION METHOD FOR
DIFFUSION-WEIGHTED MAGNETIC RESONANCE IMAGES
Academic Session
: 2014/2015-11
セ@
I declare that this thesis is classified as :
D
D
GJ
CONFIDENTIAL
(Contains confidential information under the Official Secret
Act 1972)'
RESTRICTED
(Contains restricted information as specified by the
organisation where research was done)'
OPEN ACCESS
I agree that my thesis to be published as online open access
(full text)
I acknowledged that Universiti Teknologi Malaysia reserves the right as follows :
I. The thesis is the property of Universiti Teknologi Malaysia.
2. The Library of Universiti Teknologi Malaysia has the right to make copies for the purpose
of research only.
3. The Library has the right to make copies of the thesis for academic exchange.
Certified by :
SIGNATURE OF SUPERVISOR
801224-10-5670
(NEW JC NO. /PASSPORT NO.)
Date: 3•d of June 2015
NOTES
•
PM DR. SYED ABD. RAHMAN B. SYED ABU BAKAR
NAME OF SUPERVISOR
Date: 3•d of June 2015
If the thesis is CONFIDENTIAL or RESTRICTED, please attach with the letter from
the organisation with period and reasons for confidentiality or restriction.
"We hereby declare that we have read this thesis and in our opinion this thesis is
sufficient in terms of scope and quality for the purpose of awarding the degree of
Doctor of Philosophy (Electrical Engineering)"
Signature
PM. Dr. N⦅セL@
3rd of June 20 5
Name of Supervisor I
Date
Rahman b. Syed Abu Bakar
Signature
Name of Supervisor II : PM. r. Ahmad Sobri bin Muda
3rd of June 2015
Date
Signature
:
...
セᄋ@
Name of Supervisor III: Dr. Musa b. Mohd Mokji
Date
3rd of June 2015
BAHAGIAN A - Pengesahan Kerjasama*
Adalah disahkan bahawa projek penyelidikan tesis ini telah dilaksanakan melalui ke1jasama
antara _ _ _ _ _ _ _ _ _ _ dengan _ _ _ _ _ _ _ _ __
Disahkan oleh:
Tandatangan
Tarikh:
Nama
Jawatan
(Cop rasmi)
* Jika penyediaan tesislprojek melibatkan ke1jasama.
BAHAGIAN B - Untuk Kegunaan Pejabat Sekolah Pengajian Siswazah
Tesis ini telah diperiksa dan diakui oleh:
Nama dan Alamat Pemeriksa Luar
Prof. Madya Dr. M. Igbal Saripan
Department of Computer & Communication
Systems Engineering,
Faculty of Engineering,
Universiti Putra Malaysia,
43400 Serdang,
Selangor.
Nama dan Alamat Pemeriksa Dalam
Prof. Madya Dr. Rubita binti Sudirman
Fakulti Kejuruteraan Elektrik,
UTM Johor Bahru
Disahkan oleh Timbalan Pendaftar di Sekolah Pengajian Siswazah:
Tarikh:
Tandatangan :
Nama
ASRAM BIN SUlAIMAN @ SAIM
i
AUTOMATED BRAIN LESION CLASSIFICATION METHOD FOR
DIFFUSION-WEIGHTED MAGNETIC RESONANCE IMAGES
NORHASHIMAH BINTI MOHD SAAD
A thesis submitted in fulfilment of the
requirements for the award of the degree of
Doctor of Philosophy (Electrical Engineering)
Faculty of Electrical Engineering
Universiti Teknologi Malaysia
JUNE 2015
11
I declare that this thesis entitled "Automated Brain Lesion Classification Method for
Diffusion-Weighted Magnetic Resonance Images" is the result of my own research
except as cited in the references. The thesis has not been accepted for any degree
and is not concurrently submitted in candidature of any other degree.
セ@
:
.. .......... .
Signature
: ..
Name
: Norhashimah binti Mohd Saad
Date
: 3rd
of June 2015
iii
Specially dedicated to:
My parents for their priceless supports and generous prayers.
iv
ACKNOWLEDGEMENT
First of all, I am grateful to the Almighty God for giving me strength to
endure many obstacles in life. Many times, His continuous blessings and protections
have revived my determination and hope in completing my study.
I would like to take this opportunity to thank my main supervisor, Associate
Professor Dr. Syed Abdul Rahman bin Syed Abu Bakar (UTM), my second
supervisor Associate Professor Dr. Ahmad Sobri bin Muda (Department of
Radiology, Pusat Perubatan UKM) and my third supervisor, Dr. Musa bin Mohd
Mokji (UTM) for their supports and guidance throughout the duration of my Ph.D.
research.
Thank you to all colleagues in Computer Vision Video Image Processing
(CVVIP) research group for the valuable ideas and encouragement. I am fortunate to
have received many useful comments and suggestions for my work. My deep
gratitude also goes to the colleagues in UTeM especially to Dr. Abdul Rahim bin
Abdullah, Dr. Syafeeza Ahmad Radzi for their endless help and in securing grants to
support the continuation of my research. My appreciation also extends to Ms. Aslina
binti Abdul Rahman for her moral support. Not to forget, special recognitions are
dedicated to the Study Leave Unit in UTeM and Dean of FKEKK.
Last but not the least I would like to convey my grateful appreciation to my
dear family and friends for their unceasing love and encouragement throughout the
duration of my study. I have met such inspiring people along this incredible journey.
v
ABSTRACT
Diffusion-weighted magnetic resonance imaging plays an increasingly
important role in the diagnosis of several brain diseases by providing detailed
information regarding lesion based on the diffusion of water molecules in brain
tissue. Conventionally, the differential diagnosis of brain lesions is performed
visually by professional neuroradiologists during a highly subjective, timeconsuming process. Within this context, this study proposes a new technique for
automatically detecting and classifying major brain lesions of four types: acute
stroke, chronic stroke, tumor and necrosis. An analytical framework of the brain
lesions consists of four stages which are pre-processing, segmentation, features
extraction and classification. For segmentation process, adaptive thresholding, gray
level co-occurrence matrix, region splitting and merging, semi-automatic region
growing, automatic region growing and fuzzy C-means were proposed to segment
the lesion region. The algorithm performance was then evaluated using Jaccard
index, Dice index, and both false positive and false negative rates. Results
demonstrated that automatic region growing offered the best performance for lesion
segmentation while acute stroke gave the highest rate with 0.838 Dice index. Next,
statistical features were extracted from the region of interest and fed into the rulebased classifier designed to the best suit to the lesion’s features. The performance of
the classifier was evaluated based on overall accuracy, sensitivity and specificity.
The overall accuracy for the classification was 81.3%. In conclusion, the proposed
automated brain lesion classification method has the potential to diagnose and
classify major brain lesions.
vi
ABSTRAK
Pengimejan magnetik resonan pemberat-resapan memainkan peranan yang
semakin penting dalam mendiagnosis beberapa penyakit otak dengan memberikan
maklumat terperinci berkenaan perbezaan jelas lesi ke atas resapan molekul air di
dalam tisu otak. Secara konvensional, perbezaan diagnosis lesi otak dilaksanakan
secara visual oleh pakar neuroradiologi profesional dengan proses subjektif serta
memakan masa yang lama. Dalam konteks ini, kajian ini mengusulkan teknik terbaru
untuk mengesan dan mengkelaskan lesi otak utama yang terdiri daripada empat jenis:
strok akut, strok kronik, tumor dan nekrosis. Analisis rangka kerja bagi lesi otak
terdiri daripada empat peringkat iaitu pra-pemproses, pengsegmenan, pengekstrakan
ciri dan pengkelasan. Untuk proses pengsegmenan, teknik ambang adaptif, matrik
gray level co-occurrence, rantau pemisahan dan penggabungan, rantau berkembang
separa automatik, rantau berkembang automatik dan fuzzy C-means dicadangkan
untuk mensegmen rantau lesi. Prestasi algoritma kemudiannya dinilai menggunakan
indeks Jaccard, indeks Dice, dan kedua-dua kadar palsu positif dan negatif.
Keputusan menunjukkan teknik rantau berkembang automatik memberikan prestasi
terbaik untuk pengsegmenan lesi sementara strok akut memberikan kadar indeks
Dice tertinggi iaitu 0.838. Kemudian, ciri-ciri statistik diekstrak daripada rantau
tarikan dan diinputkan kepada pengelas berasaskan peraturan yang telah direka untuk
disesuaikan dengan ciri lesi. Prestasi pengelas dinilai berdasarkan kejituan
keseluruhan, kepekaan dan kekhususan. Kejituan keseluruhan untuk pengkelasan
adalah 81.3%. Sebagai kesimpulannya, teknik klasifikasi automatan lesi otak yang
dicadangkan mempunyai potensi untuk mendiagnosis dan mengklasifikasi lesi otak
utama.
vii
TABLE OF CONTENTS
CHAPTER
1
2
TITLE
PAGE
DECLARATION
ii
DEDICATION
iii
ACKNOWLEDGEMENT
iv
ABSTRACT
v
ABSTRAK
vi
TABLE OF CONTENTS
vii
LIST OF TABLES
xi
LIST OF FIGURES
xiii
LIST OF ABBREVIATIONS
xvi
LIST OF SYMBOLS
xviii
LIST OF APPENDICES
xix
INTRODUCTION
1
1.1
Introduction
1
1.2
Problem Statement
2
1.3
Objectives
4
1.4
Scope of Work
5
1.5
Proposed System Design
6
1.6
Thesis Contributions
8
1.7
Thesis Outline
9
LITERATURE REVIEW
11
2.1
Introduction
11
2.2
The Human Brain
11
2.3
Neuroimaging Techniques
13
viii
2.3.1 Magnetic Resonance Imaging
15
2.3.2 Diffusion-Weighted Imaging
17
2.3.2.1 Pulse Sequence of DiffusionWeighted Imaging
18
2.3.2.2 Advantages of DiffusionWeighted Imaging
2.3.2.3 Types of Brain Lesions
2.4
2.5
2.7
3
22
Computer-Aided Detection and Diagnosis for
Brain Lesions
26
2.4.1 Stroke Lesion Analysis Techniques
26
2.4.2 Tumor Lesion Analysis Techniques
29
Diffusion-Weighted Imaging Analysis
Techniques
2.6
20
31
Brain Lesion Segmentation and Classification
Techniques
34
Summary
37
METHODOLOGY FOR LESION DETECTION,
SEGMENTATION AND CLASSIFICATION
41
3.1
Introduction
41
3.2
Data Collection
42
3.3
Image Pre-Processing
46
3.3.1 Image Normalization
47
3.3.2 Background Removal
47
3.3.3 Image Enhancement
49
3.4
3.3.3.1 Gamma-Law Transformation
49
3.3.3.2 Contrast Stretching
50
Segmentation Analysis Techniques
51
3.4.1 Adaptive Thresholding
53
3.4.2 Gray Level Co-Occurrence Matrix
59
3.4.2.1 Co-Occurrence Histogram
62
3.4.2.2 Minimum and Maximum
Threshold
65
ix
3.4.2.3 Region and Boundary
Information
3.4.2.4 Optimal Threshold Values
67
3.4.3 Region Splitting and Merging
69
3.4.4 Automated Region Growing
76
4.2.5 Fuzzy C-Means
80
Features Extraction
85
3.5.1 First Order Statistic Features
86
3.5.2 Statistical Features Analysis
87
3.6
Rule-Based Classifier
90
3.7
Performance Assessment Matrices
93
3.5
4
66
RESULTS AND PERFORMANCE ANALYSIS
97
4.1
Introduction
97
4.2
Pre-processing Stage
98
4.3
Segmentation Stage
101
4.3.1 Adaptive Thresholding
101
4.3.2 Gray Level Co-occurrence Matrix
103
4.3.3 Automatic Region Growing
104
4.3.4 Fuzzy C-Means
107
4.4
Performance Evaluation of Segmentation
Techniques
4.5
108
Performance Evaluation of Lesion
Classification
118
4.5.1 Classification Results Validated with
Manual Reference Segmentation
119
4.5.2 Classification Results for Automated
System
4.6
Performance Benchmarking with Other
122
125
Techniques
4.6.1 Lesion Segmentation Benchmarking
125
4.6.2 Lesion Classification Benchmarking
128
x
5
CONCLUSIONS AND FUTURE WORK
130
5.1
Conclusions
130
5.2
Contributions
133
5.3
Recommendations
133
REFERENCES
135
Appendices A-D
149-159
xi
LIST OF TABLE
TABLE NO.
TITLE
PAGE
2.1
Major parts of the brain
12
2.2
Comparison of reliabilities in neuroimaging modalities
21
2.3
Lesion in DWI and ADC image
24
2.4
Description of brain lesions, types, symptoms and
pathological findings
2.5
25
Summary of methods for DWI segmentation and
classification
39
3.1
Summary of dataset for segmentation
43
3.2
Summary of training and validation dataset for features
extraction and classification
44
3.3
Gray level description from DWI raw data
46
3.4
Types of intensity label according to hyperintensity or
hypointensity lesions
46
3.5
Average optimal threshold values for hyperintense lesions
58
3.6
Range of brain regions from GLCM
65
3.7
Intensity of lesions based on histogram
71
3.8
Threshold value of selected homogeneity criteria
72
3.9
Statistical features for region level 2
73
3.10
Statistical features for region level 3
74
3.11
Feature extraction comparison with medical diagnosis
85
3.12
Performance evaluation for classification of Class A
96
4.1
Segmentation results using adaptive thresholding
102
4.2
Performance evaluation of each lesion
103
4.3
Segmentation results using GLCM
104
4.4
Segmentation results using region splitting and merging
105
4.5
Segmentation results using automatic region growing
106
xii
4.6
Segmentation results using FCM
108
4.7
Classifier input with 30 % training data
118
4.8
Classifier input with 50 % training data
118
4.9
Classifier input with 70 % training data
119
4.10
Classifier input with 90 % training data
119
4.11
Confusion matrix for 30 % training data
120
4.12
Confusion matrix for 50 % training data
120
4.13
Confusion matrix for 70 % training data
120
4.14
Confusion matrix for 90 % training data
120
4.15
Overall classification accuracy with different training data
validated with manual reference segmentation
4.16
Confusion matrix for 50 % training data for fully
automatic classification system
4.17
121
123
Evaluation results for proposed fully automatic
region growing
125
4.18
Performance comparison of stroke segmentation
127
4.19
Classification results using rule-based classifier
128
4.20
Comparison with previous researchers
129
xiii
LIST OF FIGURES
FIGURE NO.
TITLE
PAGE
1.1
Proposed system design
6
2.1
Major parts of the human brain
12
2.2
Electromagnetic spectrum in medical imaging
14
2.3
Schematic diagram of an MRI machine
16
2.4
CT and conventional MRI of patient with tumor
17
2.5
DWI pulse sequence
18
2.6
Acute stroke
20
2.7
Average sensitivity by different modalities for
neuroimaging of acute stroke
23
3.1
Flow chart of DWI image analysis
43
3.2
DWI images of the brain and lesions indicated by
neuroradiologists
45
3.3
Intensity description from image dataset
45
3.4
DWI of solid tumor
49
3.5
Image response for gamma-law transformation
50
3.6
Input and output response for contrast stretching
51
3.7
Flowchart of the proposed adaptive thresholding
technique
54
3.8
32 × 32 pixels per macro-block region
55
3.9
Histogram distribution of each macro-block region
55
3.10
Normal and abnormal histogram of macro-block regions
56
3.11
Superimposed histogram
57
3.12
Gradient function to determine the optimal threshold
57
3.13
Brain lesion segmentation using adaptive thresholding
59
3.14
Eight nearest-neighbor pixel cells to pixel
60
xiv
3.15
Segmentation process using GLCM
61
3.16
GLCM for DWI
62
3.17
Plot of co-occurrence histogram
64
3.18
Image cluster based on COH
65
3.19
GLCM cross-section and gradient function
66
3.20
Gray level region and boundary information
66
3.21
Example of GLCM segmentation
68
3.22
Flow chart of the splitting and merging algorithm
69
3.23
Example of the split and merge segmentation process
70
3.24
Region splitting at first stage shows split image of
region 3 and 4
3.25
Histogram of region 3 and 4, the arrow shows the
histogram of hyperintense lesion
3.26
71
Region splitting at second level with lesion apprearance
in region 4-1 and 4-3
3.27
71
73
Region splitting at third level with lesion apprerance in
region 4-1-4
74
3.28
Hyperintense lesions and their segmentation results
75
3.29
Segmentation results for hypointense lesion and normal
76
3.30
Flowchart of the proposed method
77
3.31
Region splitting and merging
79
3.32
Histogram from region slitting and merging
79
3.33
Hyperintense lesions and their segmentation results for
acute stroke
3.34
80
Hypointense lesions and their segmentation results for
chronic stroke
80
3.35
Flow chart of FCM segmentation
83
3.36
Flowchart of FCM clustering in this study
84
3.37
FCM segmentation
84
3.38
Features extraction process
88
3.39
Distributions of median and mean
89
3.40
Distributions of mode and standard deviation
89
3.41
Distributions of compactness and mean of boundary
89
3.42
code for a rule-based classifier
90
xv
3.43
Rule-Based classification process
91
3.44
Code for the classification process
92
3.45
Segmentation assessment indices
94
4.1
Flowchart of the analysis
98
4.2
DWI of acute stroke and histograms for pre-processing
stage
100
4.3
Jaccard index of the segmentation results
110
4.4
Average Jaccard index for each technique
110
4.5
FPR of the segmentation results
111
4.6
Average FPR for each technique
112
4. 7
FNR of the segmentation results
113
4.8
Dice index for the segmentation results
114
4.9
Average Dice index
114
4.10
Average simulation time
115
4.11
Comparison of the segmentation techniques in
terms of similarity indices
4.12
Comparison of the segmentation techniques in
terms of error rates
4.13
117
Sensitivity of the lesion classification validated with
manual reference segmentation
4.14
117
121
Sensitivity of the lesion classification
for fully automatic classification system
122
4.15
Sensitivity of the lesion classification for 50 % training data 123
4.16
Specificity for 50 % training data
124
4.17
Overall classification accuracy for 50 % training data
124
4.18
Automatic region growing performances for 28 acute
stroke cases
126
xvi
LIST OF ABBREVIATIONS
ADC
-
Apparent Diffusion Coefficient
ANFIS
-
Adaptive Network Based Fuzzy Inference System
AO
-
Area Overlap
AS
-
Acute Stroke
BG
-
Background Image
CAD
-
Computer Aided Detection and Diagnosis
CS
-
Chronic Stroke
CSF
-
Cerebral Spinal Fluid
CT
-
Computed Tomography
DICOM
-
Digital Imaging and Communications in Medicine
DWI
-
Diffusion-Weighted Imaging
FCM
-
Fuzzy C-Means
FDA
-
Food and Drug Administration
FNR
-
Fast Negative Rate
FPR
-
Fast Positive Rate
GLCM
-
Gray Level Co-Occurrence Matrix
IBSR
-
Internet Brain Segmentation Repository
kNN
-
k-Nearest Neighbors
k-SOM
-
Kohonen Self-Organizing Map
MA
-
Misclassified Error
MAPE
-
Mean Absolute Percentage Error
MLP
-
Multi-Layer Perceptron
MRI
-
Magnetic Resonance Imaging
MS
-
Mean-Shift
xvii
NC
-
Necrosis
PD
-
Proton Density
PET
-
Positron Emitted Tomography
RBF
-
Radial Basis Function
ROI
-
Region of Interest
S
-
Sensitivity
SNR
-
Signal to Noise Ratio
Sp
-
Specificity
SPM
-
Statistical Parametric Mapping
ST
-
Solid Tumor
SVM
-
Support Vector Machine
TE
-
Time Echo
TN
-
True Negative
TP
-
True Positive
TR
-
Time Repetition
WHO
-
World Health Organization
WM/GM
-
White Matter/Gray Matter
xviii
LIST OF SYMBOLS
B0
-
Magnetic Field Strength
b
-
Diffusion Gradient
D
-
Diffusion Coefficient
E
-
Energy
eV
-
Electron Volts
Hz
-
Hertz
T
-
Tesla
r err
-
Absolute Error Ratio
γ
-
Gamma
µ
-
Mean
σ
-
Standard Deviation
xix
LIST OF APPENDICES
APPENDIX
TITLE
PAGE
A
Enhancement Performance Analysis
150
B
Segmentation Performance Evaluation
151
C
List of Publications
158
D
List of Awards
160
1
CHAPTER 1
INTRODUCTION
1.1
Introduction
Diffusion-weighted magnetic resonance imaging (DW-MRI or DWI) is
increasingly having an important role in the diagnosis of many brain diseases. This
medical imaging technique provides higher pathologic or lesion contrast based on
diffusion of water molecules in brain tissues, compared to conventional MRI. This
unique information of diffusion properties plays a key role in evaluating multiple
neurologic diseases, especially for stroke detection (Mukherji et al., 2002,
Holdsworth and Bammer, 2008). It also gives additional information for cerebral
diseases such as stages in neoplasm (cancer, tumor, and necrosis), infections, and
others.
DWI is considered as the most sensitive technique in detecting acute
infarction and is useful in giving details of the component of brain lesions (Mukherji
et al., 2002). In DWI, image intensity and contrast only depend on the strength of
diffusivity of tissue. Tissue with altered diffusion rates may appear with either
hyperintense or hypointense on a pixel basis, which is absent in healthy tissue. Such
information forms vital image characteristics that may lead to the classification of
several brain-related diseases.
Automatic brain lesion detection and classification is necessary to implement
successful therapy and treatment planning. Computer-aided detection and diagnosis
system (CAD) has become a major research subject in diagnostic radiology to assist