C 5 B 4

TUMOR DETECTION AND TISSUE
SEGMENTATION IN MAGNETIC RESONANCE

BRAIN IMAGES USING FUZZY AND
OPTIMIZATION TECHNIQUES
A THESIS
Submitted by
VISHNUVARTHANAN.G
(Reg.No.201111205)
in partial fulfillment for the award of the degree

of
DOCTOR OF PHILOSOPHY
DEPARTMENT OF INSTRUMENTATION AND

CONTROL ENGINEERING
KALASALINGAM UNIVERSITY
(KALASALINGAM ACADEMY OF RESEARCH AND EDUCATION)
ANAND NAGAR, KRISHNANKOIL-626126
TAMILNADU, INDIA.
MAY 2015
(Kalasalingam Academy of Research and Education)
ANAND NAGAR, KRISHNANKOIL-626 126
CERTIFICATE
This is to certify that no corrections/suggestions were pointed out by

the Indian/Foreign Examiner(s) in the Thesis titled “TUMOR DETECTION
AND TISSUE SEGMENTATION IN MAGNETIC RESONANCE BRAIN
IMAGES USING FUZZY AND OPTIMIZATION TECHNIQUES”
submitted by Mr.G.Vishnuvarthanan
SUPERVISOR
Place: KLU
Date: 26.08.15
Minutes of the Ph.D. Viva-Voce Examination of Mr. G. Vishnuvarthanan (Reg. No.
20lltl205) held on 21't August, 2015 (Friday) at 02.30 PM in the Seminar Hall of
International Research Centre (1't floor of IRC), Kalasalingam University, Anand Nagar,
Krishnankoil-626 126.
The Ph.D. Viva-Voce Examination of Mr. G. Vishnuvarthanan (Reg. No. 20ll1l201) on his
Ph.D. thESiS ENtitIEd ..TUMOR DETECTION AND TISSUE SEGMENTATION IN
MAGNETIC RESONANCE BRAIN IMAGES USING FIIZZY AND OPTIMIZATION
TECHNIQUES" was conducted on 21't August, 2015 in the Seminar Hall of International
Research Centre (1't floor of IRC), Kalasalingam University, Anand Nagar, Krishnankoil-
626 126.
The following members of the Oral Examination Board were present:
l. Dr. M. Pallikonda Rajasekaran,

Professor and Head,
Department of Electronics and Communication Engineering, Supervisor
Kalasalingam University,
Krishnankoil-626 126.
Dr. G. Zayaraz,
Professor,
Department of Computer Science and Engineering, lndian Examiner
Pondicherry Engineering College,
Puducherry-605 014, India.
J. Dr. B. Kannapiran,
Professor and Head,
Department of Instrumentation and Control Engineering,
Kalasalingam University, Chairman/DRC
Anand Nagar, Kri.shnankoil - 626 126.
The candidate, Mr. G. VISHNUVARTHANAN presented the salient features of his Ph.D.
work. This was followed by questions from the board members. The queries and clarifications raised
by the Overseas and Indian Examiners were also put to the candidate. The candidate answered the
questions to thefull satisfaction of the board members.
The Indian and Foreign Examiners have accepted the thesis in the present form.
Based on the candidate's research work, his presentation and also the clarifications and
answers by the candidate to the questions raised by the examiners, the board recommends that
Mr. G.
VISHI\UVARTTTANAN be awarded the Ph.D. degree in the FACULTY oF
INSTRUMENTATION AND CONTROL ENGINEERING.
e}(v\"J
n Dr. GlZayaraz
(Indian Examiner) ",&kam"
(Chairman/DRC)
DECLARATION
I hereby declare that the thesis entitled “TUMOR DETECTION AND

TISSUE SEGMENTATION IN MAGNETIC RESONANCE BRAIN
IMAGES USING FUZZY AND OPTIMIZATION TECHNIQUES”
Submitted by me for the Degree of Doctor of Philosophy in Department of
Instrumentation and Control Engineering is the result of my original and
independent research work carried out under the guidance of Dr.M.Pallikonda
Rajasekaran, Kalasalingam University, Anand Nagar, Krishnankoil and it has
not been submitted for the award of any degree, diploma, associateship,
fellowship of any University or Institution.
Signature of the Candidate

G.Vishnuvarthanan
BONAFIDE CERTIFICATE
Certified that this thesis titled, “TUMOR DETECTION AND
TISSUE SEGMENTATION IN MAGNETIC RESONANCE BRAIN
IMAGES USING FUZZY AND OPTIMIZATION TECHNIQUES” is the
bonafide work of Mr. VISHNUVARTHANAN G, who carried out the
research under my supervision. Certified further, that to the best of my
knowledge the work reported herein does not form part of any other thesis or
dissertation on the basis of which a degree or award was conferred on an
earlier occasion on this or any other scholar.
Signature of the Supervisor

Dr.M.PALLIKONDA RAJASEKERAN
SUPERVISOR
Professor & Head
Department of ECE,
Kalasalingam University,
Krishnankoil-626126
Tamilnadu, India.
iii
ABSTRACT
Image processing plays a very important role in analyzing the images
of different standards. Image enhancement, image restoration and image
compression are the types of image processing techniques of which image
segmentation process under Image enhancement technique is utilized
extensively to analyze the medical images.
An image is composed of several groups of pixels. Segmentation is
an image processing technique, in which different regions of an image are
segregated based upon the pixel intensities available within the image.
Boundary and edge based segmentations are the two types of segmentation
procedures, of which, boundary based segmentation is used widely.
In the presence of several medical imaging techniques, Magnetic
Resonance Imaging (MRI) technique occupies a predominant position in the
area of diagnosing the organs of human body. The functionality of an MRI
scanner entirely relies upon the 80% of hydrogen atom present in our body.
The hydrogen atom of our human body is stimulated using resonant radio
frequency signals. The excited hydrogen atom is capable of emitting the

iv
absorbed radio frequency signals, upon which several signal processing
operations are done to obtain a clear anatomy of human organs and tissues.
Cancer cells play a major role in tumor formation inside the human
body. A survey made by “Times of India” has identified that nearly 3 million
people living in India suffer from cancer, out of which 1 million people are
registered with new forms of cancer. The effects created by tumor are
extremely pervasive, leading to the formation of lesions and tumors within the
human body.
Tissues in brain are the most complicated part of our body; a clear
examination and study are therefore required by a radiologist to identify the
pathologies embed within the tissue of brain. Normal Magnetic Resonance
(MR) scanner is capable of producing brain images with bounded tissues,
where unique and segregated views of the tissues are required. A
distinguished view upon the images is manually impossible and can be
subjected to operator - errors. Further, individual segmentation and
identification of tumor affected region in brain are quite unavailable in a
clinical MRI scanner. Due to the presence of afore mentioned hindrances,
automated algorithms capable of both tumor identification and tissue
segmentation processes are proposed through this research work.

v
With the support of soft computing techniques, automated and
unique segmentation upon the brain tissues, the identification of
heterogeneous type tumors can be effectively done. These functionalities
assist the radiologist in diagnosing process, extensively. Large volumes of
patient images can be easily diagnosed with minimal manual intervention and
non - invasiveness. The algorithms also favour pre and post radio therapeutic
planning procedures in treating the tumor region. This research work also
focuses upon the separation of tumor and edema regions.

vi
ACKNOWLEDGEMENT
My first and foremost thanks to the Almighty, for His endless

blessings showed upon me throughout this endeavor. I am indebted and
grateful to my supervisor Dr. M. Pallikonda Rajasekeran for his guidance,
inspiration and fruitful discussion during the course of this research work.
Without him this work would not have been possible.
I am extremely thankful to the chairman, A.K. Group of

Institutions “Kalvivallal” Thiru. T. Kalasalingam, Chancellor “Ilayavallal”
Thiru. K. Sridharan and Vice-Chancellor Dr. S. Saravana Sankar,
Kalasalingam University for providing the necessary facilities during the period
of my research. I like to deliver my gratitude and thanks to Dr. K.G. Srinivasan,
MD, RD, Consultant Radiologist and Dr. K.P. Usha Nandhini, DNB, KGS
Advanced MR & CT Scan - Madurai, Tamilnadu, India, for supporting the
research with the patient information and MR brain images. I thank

Dr. Bridgetta Jeyaseelan, Former Head, P.G and Research Department of
English, Fatima College, Madurai; who has gone through the entire thesis with
utmost patience and offered valuable suggestions for the correct usage of English
language.
I wish to express my sincere thanks to all the Teaching Staff

members of the Electronics and Communication Department who helped me
during the dissertation work. Last but not least, I like to thank my father Mr.
Govindaraj Ramasamy and wife Mrs. Anitha Vishnuvarthanan for their
colossal support throughout my academic career.
VISHNUVARTHANAN G
vii
TABLE OF CONTENTS
CHAPTER NO. TITLE PAGE NO.
ABSTRACT iii
LIST OF TABLES ix
LIST OF FIGURES xii
LIST OF SYMBOLS AND ABBREVIATIONS xxii
1. INTRODUCTION 1
1.1 IMAGE PROCESSING 1
1.2 APPLICATIONS OF IMAGE PROCESSING 2
1.2.1 Medical Image Diagnosis 4
1.3 TYPES OF IMAGE PROCESSING TECHNIQUES 4
1.3.1 Image Enhancement 5
1.3.2 Image Restoration 5
1.3.3 Image Segmentation 5
1.4 CONTRIBUTIONS OF THE THESIS 6
1.5 ORGANIZATION OF THE THESIS 8
2. LITERATURE SURVEY 11
2.1 MEDICAL IMAGE PROCESSING 11
2.2 MAGNETIC RESONANCE BRAIN
IMAGE PROCESSING 11
2.2.1 Edge Detection Methods 12
2.2.2 Thresholding Methods 12
2.2.3 Compression Based Methods 13
2.2.4 Histogram Based Methods 14
viii
2.2.5 Region – Growing Methods 15

2.2.6 Split and Merge Methods 16
2.2.7 Partial Differential Equation
Based Methods 17
2.2.8 Parametric Methods 17
2.2.9 Level Set Methods 18
2.2.10 Fast Marching Methods 19
2.2.11 Graph Partitioning Methods 19
2.2.12 Watershed Transformation Methods 19
2.2.13 Model Based Segmentation Methods 20
2.2.14 Multi – Scale Segmentation Methods 20
2.2.15 One – Dimensional Hierarchical
Signal Segmentation Methods 20
2.2.16 Semi – Automatic Segmentation 21
2.2.17 Trainable Segmentation 22
2.2.18 Hybrid Segmentation Techniques 26
2.3 MOTIVATION 42
2.4 OBJECTIVE OF THE THESIS 43
3. DATA ACQUIREMENT FROM CLINIC AND

COMPARISON PARAMETERS 45
3.1 INTRODUCTION 45
3.2 DETAILS OF PATIENTS 45
3.3 TYPES OF IMAGES ACQUIRED 46
3.3.1 T1 – Weighted Images 46
3.3.2 T2 – Weighted Images 47
3.3.3 Fluid Attenuated Inversion
Recovery (FLAIR) Image 48
ix
3.3.4 MPR Type Images 48

3.3.5 Contrast Enhancement Images 49
3.4 DESCRIPTION OF MRI SCANNER 50
3.5 IMAGE QUALITY PARAMETERS 50
4 SEGMENTATION OF MR BRAIN IMAGES

USING FUZZY 54
4.1 INTRODUCTION 54
4.1.1 Types of Fuzzy Controller 55
4.1.2 Basics of Membership Function 56
4.2 METHODOLOGIES 57
4.2.1 Fuzzy Rule Implementation 59
4.3 IMPLEMENTATION 59
4.4 RESULTS AND DISCUSSION 61
4.4.1 Performance Evaluation 73
4.5 SUMMARY OF CONTRIBUTIONS 75
5. SEGMENTATION OF MR BRAIN IMAGES

USING SELF ORGANIZING MAP AND
FUZZY K - MEANS 76
5.1 INTRODUCTION 76
5.2.1 SOM Algorithm 79
5.2.2 FKM Algorithm 83
x

USING PARTICLE SWARM OPTIMIZATION
BASED FUZZY C - MEANS 109
6.1 INTRODUCTION 109
6.2.1 PSO Algorithm 113
6.2.2 FCM Algorithm 114

USING PARTICLE SWARM OPTIMIZATION
AND FUZZY C - MEANS IMPROVED BY
REGION GROWING ALGORITHM 135
7.1 INTRODUCTION 135
7.2.1 PSO Algorithm 138
7.2.2 FCM Algorithm 139
7.2.3 Region Growing Algorithm 141
xi
8 CONCLUSION AND FUTURE WORK 173

8.2 CLINICAL ASSISTANCE 175
8.3 CONCLUSION 175
8.4 SCOPE FOR FUTURE WORK 175
APPENDIX 1 177
REFERENCES 179
LIST OF PUBLICATIONS 199
VITAE 200
xii
LIST OF TABLES
TABLE NO. TITLE PAGE NO.
4.1 Comparison parameters utilized to evaluate the

efficiency of the proposed Fuzzy Inference
73
System (FIS)
4.2 Comparison of SI, OF and EF values 74
5.1 Performance evaluation for the proposed SOM
100
based FKM Algorithm
5.2 Comparison of SI, of and EF values 105
5.3 Comparison of SI, of and EF values between FIS
107
and SOM – FKM algorithms
6.1 Evaluation of PSO – FCM using standard
128
comparison parameters
6.2 Comparison of FIS, SOM – FKM and PSO –
132
FCM using SI, of and EF values
7.1 Comparison parameters utilized to evaluate the
efficiency of the proposed PSO based FCM
merged with Region Growing Algorithm 160
7.2 Comparison of SI, OF and EF values 167
7.3 Overall comparison of SI, OF and EF values 169
xiii
LIST OF FIGURES
FIGURE NO. TITLE PAGE NO.
3.1 T1 weighted axial and coronary images 47

3.2 T2 Weighted axial image 47
3.3 Fluid Attenuation Inversion Recovery (FLAIR)
image 48
3.4 Multiplanar Reconstruction (MPR) image with
Contrast Enhancement 49
3.5 T2 Axial image sequence with Contrast
Enhancement 49
4.1 Image segmentation procedure adopted by FIS
(Fuzzy Inference System) 56
4.2 Triangular membership parameters chosen for
the input variables 57
4.3 Triangular membership parameters chosen for
the output variables 57
4.4 Segmented tumor region of patient 1 (age: 5)
affected with Primitive Neuro Ectodermal tumor
Image type: MPR (Multi Planar Reconstruction)
3D contrast enhanced images 62
affected with Primitive Neuro Ectodermal tumor 63
affected with Meningioma 64
affected with high grade Astrocytoma 65
xiv

affected with high grade astrocytoma
affected by supratentorial Primitive Neuro
Ectodermal Tumor 66
affected by supratentorial Primitive Neuro
Ectodermal Tumor
4.11 Segmented tumor region of the patient suffering
from Nerve Sheath Tumors 68
from low grade Glioma 69
from high grade Glioma 70
from Pituitary Adenomas 71
4.15 Segmented tumor region of the patients suffering
from Low Grade Glioma and Metastatic
Bronchogenic Carcinoma 72
4.16 Comparison of memory required attained by
executing the proposed Fuzzy Inference System
(FIS) 74
xv
4.17 Comparison parameters for evaluating the

proposed Fuzzy Inference System (FIS) 75
5.1 Flowchart of hybrid SOM - FKM algorithm 79
5.2 Gold Standard Images of the tumor detected
region for patient – 1 86
5.3 Gold Standard Image of tumor detected region
for the data available in Harvard Brain Web
Repository 87
5.4 Gold Standard Image of tumor detected region
for the patient suffering from Nerve Sheath
Tumors 87
5.5 Segmented results of patient - 1 (age: 5) suffering
from Primitive Neuro Ectodermal tumor obtained
from clinical database 88
from Primitive Neuro Ectodermal tumor obtained
from clinical database, which are of 3D Multi
Planar Reconstruction (MPR) type images 89
5.7 Segmented results of patient - 2 (age: 35)
suffering from Meningioma obtained from
clinical database 90
5.8 Segmented results of patient - 3 (age: 32)
suffering from high grade Astrocytoma obtained
from clinical database 91
5.9 Segmented result of patient - 3 (age: 32)
suffering from high grade Astrocytoma obtained
from clinical database, which is of 3D Multi
Planar Reconstruction (MPR) type image 92
xvi

from supratentorial Primitive Neuro Ectodermal
Tumor obtained from clinical database 93
from supratentorial Primitive Neuro Ectodermal
Tumor obtained from clinical database, which are
of 3D Multi Planar Reconstruction (MPR) type
images 94
5.12 Segmented results of the patients suffering from
Glioma and Metastatic Bronchogenic Carcinoma,
obtained from Harvard Brain Web Repository 95
5.13 Segmented results of the patients suffering
Pituitary Adenomas 96
Nerve Sheath Tumors 97
low grade Glioma 98
high grade Glioma 100
5.17 Comparison of Mean Square Error values
obtained from SOM – FKM and FCM algorithms 102
5.18 Comparison of Peak Signal to Noise Ratio
(PSNR), Tanimoto Coefficient (TC) and Dice
Overlap Index (DOI) values obtained from SOM
– FKM and FCM algorithms 103
5.19 Comparison of Memory Requirement or Memory
Utilization values obtained by processing SOM –
FKM and FCM algorithms 104
xvii
5.20 Comparison of Computational Time or Elapsed

Time values obtained from SOM – FKM and
FCM algorithms 104
5.21 Comparison of Sensitivity (or) Overlap Fraction
values with other competitive algorithms 105
5.22 Average Values of Comparison Parameters for
FIS and SOM – FKM Algorithms 106
5.23 Average Values of Memory Requirement for FIS
and SOM – FKM Algorithms 107
6.1 Flowchart explaining the sequential steps of the
proposed PSO based FCM algorithm 112
6.2 Segmented results derived using PSO – FCM for
the patient - 1 (age: 5) suffering from Primitive
Neuro Ectodermal tumor 117
the patient - 1 (age: 5) suffering from Primitive
Neuro Ectodermal tumor, which are of 3D Multi
Planar Reconstruction (MPR) type images 118
the patient - 2 (age: 35) suffering from
Meningioma 119
the patient - 3 (age: 32) suffering from high grade
Astrocytoma 120
the patient - 3 (age: 32) suffering from high grade
Astrocytoma 120
xviii

supratentorial Primitive Neuro Ectodermal
Tumor (PNET) 121
supratentorial Primitive Neuro Ectodermal
Tumor obtained from clinical database, which are
of 3D Multi Planar Reconstruction (MPR) type
images 122
the patients suffering from Glioma and
Metastatic Bronchogenic Carcinoma attained
from Harvard Brain Web Repository 123
the patient suffering Pituitary Adenomas 124
the patient suffering from Nerve Sheath Tumors 125
the patient suffering from low grade Glioma 126
the patient high grade Glioma 127
6.14 Comparison of FIS, SOM based FKM and PSO
based FCM using evaluation parameters 130
6.15 Comparison of FIS, SOM based FKM and PSO
based FCM using memory requirement values 132
xix
7.1 Flowchart explaining the Sequential steps

involved in the proposed algorithm 137
affected with Primitive Neuro Ectodermal tumor 146
affected with Primitive Neuro Ectodermal tumor.
affected with Meningioma 148
affected with high grade Astrocytoma 149
affected with high grade astrocytoma.
3D contrast enhanced image 149
7.7 Figure 7.6 Segmented tumor region of patient 4
(age: 3) affected by supratentorial Primitive
Neuro Ectodermal Tumor 151
affected with supratentorial Primitive Neuro
Ectodermal Tumor
3D images 151
from Pituitary Adenomas 152
xx

from Nerve Sheath Tumors 153
from low grade Glioma 154
from high grade Glioma 155
7.13 Segmented tumor region of the patients suffering
from Glioma and Metastatic Bronchogenic
Carcinoma 156
7.14 Magnetic Resonance Spectroscopic (MRS)
Image representing the diagnosis of patient 1
(age: 5) suffering from Primitive Neuro
Ectodermal tumor 157
7.15 Individual results from PSO based FCM and
Region Growing algorithms for comparison 158
7.16 Comparison of MSE values 162
7.17 Comparison of PSNR values 162
7.18 Comparison of Jaccard Tanimoto Coefficient
(TC) Index values 164
7.19 Comparison of Dice Overlap Index (DOI) values 164
7.20 Comparison of memory requirement values in
terms of Bytes 166
7.21 Comparison of computational time in terms of
Seconds 167
7.22 Overall comparison of sensitivity values with an
approximation offered by Rasoul Khayati et al
[115] 168
xxi
7.23 Comparison of PSO based FCM improved by

RGA with other soft computing algorithms 169
7.24 Comparison of MSE values derived from PSO
based FCM improved by RGA with other soft
computing algorithms 170
7.25 Comparison of memory requirement values 171
7.26 Comparison of soft computing techniques with
Gold standards using Similarity Index, Overlap
Fraction and Extra Fraction 171
xxii
LIST OF SYMBOLS AND ABBREVIATIONS
SYMBOLS
A( ) - Input image
B( ) - Output image
- Maximum pixels in input
( ) - Jaccard Tanimoto Coefficient Index
( ) - Dice Overlap Index
Rd - Diversified input data
( ) - Input vector
( ) - Sample vector
( ) - exponential decay
( ) - Neighbourhood function
- Time constant
- topological error
- Quantization error
⃗ ⃗⃗⃗ - weight function
- Data vector
-
-
-
m - Fuzziness co – efficient
- Squared Euclidean distance
- Cluster center or centroid value
l - first cluster
⃗⃗⃗ - Velocity of the particles
- Positive constants.
xxiii
& - Uniform weight values

⃗⃗⃗ ( ) - Best position of the particle
⃗⃗⃗⃗ ( ) - Global best position
⃗⃗⃗ ( ) - Position vector
TVAC - Time Varying Acceleration Constants
C - Number of clusters in Fuzzy C – Means
- Random value for membership assignment
- Degree of membership
- Data points or voxels
- connected region
- Euclidean distance evaluation
ABBREVIATIONS
DICOM - Digital Imaging and Communications in Medicine

MRI - Magnetic Resonance Imaging
FIS - Fuzzy Inference System
GUI - Graphical User Interface
SOM - Self-Organizing Map
FKM - Fuzzy K - Means
PSO - Particle Swarm Optimization
FCM - Fuzzy C – Means
RGA - Region Growing Algorithm
MSE - Mean Square Error
PSNR - Peak signal to Noise ratio
TC - Jaccard Tanimoto Coefficient Index
DOI - Dice Overlap Index
SI - Similarity Index
OF - Overlap Fraction
EF - Extra Fraction
xxiv
FLAIR - Fluid Attenuated Inversion Recovery

MPR - Multi Planar Reconstruction
GM - Grey Matter
WM - White Matter
CSF - Cerebro Spinal Fluid
CAD - Computer Aided Diagnosis
CT - Computer Tomography
PET - Positron Emission Tomography
MDL - Minimum Description Length
NC - Noise Cluster
mBm - multifractional Brownian motion
GMIs - Gradient Magnitude Intensities (GMIs)
LIM - Local Intensity Minimum (LIM)
ROI - Region of Interest
PCNNs - Pulse-coupled neural networks (PCNNs)
HSOM - Hierarchical Self Organizing Map (HSOM)
IFCM - Improved FCM (IFCM)
ANN - Artificial Neural Network (ANN)
P – SVM - Probabilistic Support Vector Machine (P - SVM)
SVM - Support Vector Machine (SVM)
LS – SVM - Least Square – Support Vector Machine (LS - SVM)
BEM - Brain Extraction Methods
MRS - Magnetic Resonance Spectroscopy
MR - Magnetic Resonance
BFO - Bacteria Foraging Optimization
FCM - Kernel based Fuzzy C Means
GFCM - generalized FCM
GFCR - Generalized Fuzzy Clustering Regularization
EM - Expectation-Maximization algorithms
ML - Maximum Likelihood
xxv
MRF - Markov Rrandom Fields

BET - Brain Extraction Tool
MLP - Multi – Layered Perceptron
RFCM - Robust Fuzzy C - Means
ABF - Amended Bacterial Foraging optimization algorithm
ACM - Active Contour Model
LOGISMOS - B - Layered Optimal Graph Image Segmentation of
Multiple Objects and Surfaces
MRISIMUL - MR physics simulator
RF - Random Forests
DMRI - Diffusion MR Images
ODF - Orientation Distribution Function
MMTDCM - Measure of Medium Truth Degree based FCM
ABC - Artificial Bee Colony
SFCM - Spatial Fuzzy C – Means
GMM - Gaussian Mixture Model
MFCM - Modified Fuzzy C – Means
LIPC - local independent projection-based classification
BS - Breeding Swarm
S –FCM - Supressed FCM framework
R – FCM - Regularized FCM approach
CJD - Creutzfeldt–Jakob Disease
sJCD - sporadic Creutzfeldt–Jakob Disease
vJCD - variant Creutzfeldt–Jakob Disease
CGMM - Constrained Gaussian Mixture Model
PABIC - Parametric Bias Field Correction
ITDS - Information Theoretic Discriminative Segmentation
SITDS - supervised Information Theoretic Discriminative
Segmentation
GFCR - Generalized Fuzzy Clustering Regularization
xxvi
EM - expectation-maximization algorithms
ML - Maximum Likelihood
MRF - Markov Random Fields
TE - echo spin time
SL - Slice Thickness
TR - Repetition time
TE - Echo spin Time
GADO IV - Gadolinium IV
TP - True Positive
FP - False Positive
FN - False Negative
Trimf - Triangular Membership Function
MF - Membership Function
RAM - Random Access Memory
EZW - Embedded Zerotree Wavelet
CE - Contrast Enhancement
PNET - Primitive Neuro Ectodermal Tumor
3D - 3 – Dimensional
BEA - Brain Extraction Algorithm
ROC - Region of Convergence
ROI - Region of Interest
BMU - Branch Metric Unit
1
CHAPTER 1
INTRODUCTION
1.1 IMAGE PROCESSING
Image processing is a methodology which is capable of converting

an image into digital form and it performs certain operations on image, so as
to achieve an enhanced image or to extract some vital information from it. It
is similar to digital signal processing. In image processing, input is an image
(may be a video frame or a photograph in any format) and the output may be
an image or the characteristics of the input image. Image processing system
usually considers an image as a two dimensional signal, while processing. It is
one among the emerging technologies, with its branches of application
widespread into several domains of business. Image processing is a core
research area in engineering and it also acts as a thrust area in other
disciplines of computer science. Researchers are in need of image processing;
as it offers real time applications and the results derived from image
processing techniques are also made available to the hands of its user.
There are two main methods in image processing and they are:
1. Analog Image Processing
2. Digital Image Processing

2
1. Analog Image Processing
Analog image processing relates to the modification of an image

through analog or electrical signals. A common example of this process is the
television image.
The amplitude levels of the voltage vary to illustrate the brightness

of an image. Through the application of varied electrical signals, the
appearance of the displayed image is modified (or) altered. The brightness
and contrast levels in the image are altered with the aid of both the controls
available on amplitude and the reference of the video signals in a television
set.
2. Digital Image Processing
Digital image processing generally denotes the procedure of

processing a two-dimensional picture by a digital computer. A digital image is
comprised of an array of real numbers symbolized by a finite number of bits.
In digital image processing, digital computers are used to process the

input image. The image would be converted to a digital form using an
electronic device, say scanner, and then the further processing upon the image
is done. It can also be mentioned as the numerical representations of the
images to enable a sequence of operations to derive a desired result. Digital
image processing commences the processing with one image and finally, a
reformed version of the same image is obtained.
1.2 APPLICATIONS OF IMAGE PROCESSING
1. Visualization – To observe the objects that are not visible to

the human eye.
3
2. Image sharpening and restoration - To generate a better image.
3. Image retrieval – To identify and retrieve certain valuable

information for the interested image.
4. Measurement of pattern – Measures various objects of

different geometric shapes in an image.
5. Image Recognition – Distinguishes the specific objects present

in an image.
Some other applications:
Image processing has a wide range of applications and some of the

areas which require it are mentioned below:
1. Medicine
 Examination and interpretation of images obtained from

X-rays, MRI, PET and CT scans.
 Identification of blocks in the image comprising of both

arteries and veins.
 Analysis of cell structures and microbial images.
2. Agriculture
 Satellite/aerial views of land, for example to determine

how much land is being used for different purposes or to
investigate the suitability of different regions for
different crops.
 Inspection of fruit and vegetables distinguishing good

and fresh produce from old.
4
3. Industry
 Automatic inspection of items on a production line.
 Inspection of paper samples.
 Identification of leakage in gas pipe lines.
4. Security and Crime Detection
 Fingerprint and knuckle analysis.
 Sharpening or recovery of de-blurring in speed-camera

images.
 Identification of criminals
1.2.1 Medical Image Diagnosis
Generally, medical images obtained from hospitals are in DICOM

(Digital Imaging and Communications in Medicine) format. These image
formats are quite large in size and require higher memory space for storage.
For portability of these data, they are converted into JPEG, JPG, BMP, TIFF,
PNG file formats. Analysis of images in DICOM format is a tedious process
and the images are converted into any of the above said file format, and they
are used worldwide. Image processing is extended to such medical image
diagnosis so as to identify the pathologies present in our body, especially the
pathologies present in human brain that are difficult to diagnose. This thesis
focuses upon this particular issue and is organized to resolve it.
1.3 TYPES OF IMAGE PROCESSING TECHNIQUES
Image processing techniques are further subdivided into broader

subclasses such as image restoration, image segmentation and image
enhancement. Different algorithms are available for executing these
5
subclasses of image processing and some of the techniques implied are briefly
discussed in chapter 2.
1.3.1 Image Enhancement
This process is related to the improvement of an input image, so that

the final result obtained after improvement is more suitable for a specific
application. Some of the applications of Image enhancement are:
1. Sharpening or de-blurring an out focussed image.
2. Highlighting edges.
3. Improving the contrast levels or brightening an image.
4. Noise removal.
1.3.2 Image Restoration
Image restoration is a process of reversing the damage done to an

image by a known source, for example:
1. Removal of blur created by linear motion.
2. Removal of visual distortions.
3. Removing periodic intrusions.
1.3.3 Image Segmentation
This is the process of subdividing an image into essential parts or

segregating certain aspects of an image, so that deeper analysis could be
exhibited.
6
1. Circles or specific shapes in an image.
2. In an aerial photograph, identifying cars, trees, buildings or

roads.
3. In identifying the pathologies present in a medical image.
These subclasses are interrelated with each other. A single algorithm

can perform both image enhancement and image restoration. Care and
concern should be preserved while applying the above said techniques to an
image, as the algorithms should not deteriorate the quality of the image, or in
other words, the originality of the image should not be disturbed.
1.4 CONTRIBUTIONS OF THE THESIS
This thesis has made an evaluation upon Fuzzy and its several
combinations with other algorithms. The thesis addresses the problems of i)
tumor identification ii) tissue segmentation iii) discrimination of tumor from
edema portions. The methodologies proposed through this thesis are well
capable of giving an initial level sketching for radio therapeutic procedures or
for image guided surgeries.
1. To segment the tissues of the brain and to extract the tumor

region using Fuzzy (Fuzzy Inference System).
Fuzzy Inference System (FIS) is the process of formulating the

mapping from a given input image to an output image using Fuzzy Logic. FIS
is combined with the procedures of developing membership functions, Fuzzy
Logic operators and if-then rules. Mamdani-type and Sugeno-type are the two
types of Fuzzy Inference System that can be implemented in Fuzzy Logic. In
this case, framing the Fuzzy rules and implementing them for image
segmentation are done exclusively by a Mamdani - type controller. The Fuzzy
Logic provides a number of interactive tools that allow accessing many of the
7
functions through a Graphical User Interface (GUI). Here, Fuzzy Inference

System is used as an initial level technique for tumor detection and tissue
segmentation in MR brain images.
2. A novel combinatory algorithm using Self Organizing Map (SOM)

and Fuzzy K Means algorithm (FKM) for tumor identification and
tissue segmentation in magnetic resonance brain images.
FIS has no capability of effectively segmenting T1 - Weighted,

FLAIR and MPR brain images. Better segmentation results are offered only
with T2 – Weighted and contrast enhanced MR brain images. To overcome
these hindrances, SOM based FKM is proposed to have a clear examination
of other modalities of MR images. SOM, an unsupervised neural network
technique offers an initial level clustering and dimensionality reduction.
Additional clustering operation is performed by FKM. Since, K - Means is a
greedy algorithm, the shortcomings of FKM is minimized with the inclusion
of SOM strategy. Tumor region and tissue regions in T2 – Weighted, MPR
and FLAIR type MR brain images are excellently segmented using the SOM
based FKM technique. The results of SOM based FKM and FIS algorithms
are evaluated using nine standard comparison parameters. SOM based FKM
offers good segmentation results in comparison with FIS technique.
3. A complete automated algorithm for the segmentation of tissues

and identification of tumor region in MR brain images using
Particle Swarm Optimization and Fuzzy C – Means algorithm
(PSO – FCM).
SOM based FKM and FIS algorithms lack capability to produce

better segmentation results in T1 weighted brain images. This drawback is
completely eradicated with the assistance of PSO based FCM algorithm. PSO
produces optimized results, upon which clustering is performed using FCM
8
algorithm. The time consumption level of FCM algorithm is subsequently

reduced with the allocation of cluster centres done by PSO. This algorithm
has solidified better segmentation of T1 weighted MR brain images. PSO
based FCM produces better segmentation results in comparison with SOM
based FKM and FIS techniques.
4. Tumor identification in T1, T2, FLAIR and MPR type brain

images using a combination of Optimization and Fuzzy clustering
merged with a novel seed point initialization based Region
Growing algorithm.
The algorithm exhibits a combination of the concepts of

optimization (PSO) and Fuzzy – Clustering (FCM), with enhancement of
segmentation results using region growing algorithm. It is also observed that
the inclusion of PSO based FCM has geared up the performance of region
growing algorithm by making it quite insensitive to threshold initiation. Time
Varying Acceleration Constants (TVAC) is utilized in PSO, rather using
normal positive constants. All these factors have favoured efficient tumor
identification and tissue segmentation. PSO based FCM merged with region
growing algorithm produces excellent segmentation results when compared
with FIS, SOM based FKM and PSO based FCM methodologies.
1.5 ORGANIZATION OF THE THESIS
The thesis is organized into eight chapters and is as follows:
Chapter 1
A brief introductory about the image processing techniques and their

impact upon medical image analysis is described in this chapter. Applications
of image processing along with the original contributions and organization of
the thesis are discussed.
9
Chapter 2
Chapter 2 entirely comprises of ―Literature Survey‖, where a

detailed discussion on the technical advancements and review of the existing
technologies for MR brain image processing have been explicated. Motivation
and objectives of the thesis are also briefed in this chapter.
Chapter 3
Clinical data acquisition along with the patient details, briefing of

the images attained from Harvard brain web database and the exposition of
difficulties experienced in data acquirement are also done in this chapter.
Chapter 4
This chapter explains the functioning of Fuzzy Inference technique

(FIS) used for segmenting MR brain images. Tumor detection in T2 weighted
images and contrast enhanced sequences is done by FIS algorithm. Sparse
tissue segmentation is produced by FIS. This method tends to be a
preliminary understanding of the functionality of soft computing
methodologies.
Chapter 5
This chapter elucidates the operation of SOM based FKM technique.

The objective of this chapter is to present low Mean Square Error (MSE),
time consumption and high Peak signal to Noise ratio (PSNR). SOM based
FKM helps to provide better tumor identification and tissue segmentation and
it is clearly briefed in this chapter.
10
Chapter 6
This chapter describes the functioning of PSO based FCM

algorithm. The efficacy of the PSO based FCM in segmenting T1 weighted
MR brain images is illustrated in this chapter. Tissue segmentation and
extraction of tumor region from edema portions, along with tumor detection
are offered by PSO based FCM. All these processes are discussed in this
chapter.
Chapter 7
This chapter briefs the working principles of PSO based FCM

merged with RGA (Region Growing Algorithm). The later part of this chapter
explains the excellent performance of this technique and figures out an
effective comparison with the other methodologies. Effective tissue
segmentation and tumor detection of PSO based FCM with RGA and
validations using the gold standard images are incorporated in this chapter.
Chapter 8
This chapter explains both the summary of contributions attained

through this research work and the suggestions for future work.
11
CHAPTER 2
LITERATURE SURVEY
2.1 MEDICAL IMAGE PROCESSING
Current medical imaging systems are capable of providing large

amounts of images, which needs in - depth analysis. Normally, experts
perform certain evaluation procedures upon the medical images that may lead
to operational errors which require huge amount of time for making
evaluations. MR images are qualitatively and quantitatively analyzed by
experts based on their professional experience, but it is certainly limited by
the human vision system, where human eye vision is restricted to analyze 8
bits of grey level. Nowadays, MRI systems are capable of offering images of
the organs up to 65,535 gray levels. Certain vital information acquired
through an MRI scanner cannot be analyzed using a normal human eye,
which has visual constraints. This paves the way to use Computer Aided
Diagnosis (CAD), by which both high and low resolution MR images can be
deeply analyzed. CAD supports a radio surgeon to have a clear examination
of the aberrant regions. These factors have made medical image processing
inevitable, so that millions of life could be saved through earlier and accurate
diagnosis.
2.2 MAGNETIC RESONANCE BRAIN IMAGE PROCESSING
There are several image processing methodologies that can be

applied upon MR brain images and facilitate effective patient diagnosis.
12
There are several research issues underlying in the domain of MR brain image
processing and some such issues are mainly focused in this research. Tumor
identification and tissue segmentation present in multi - channeled MR brain
images, which are in different sequences and are of multi – axis structures, are
the major concern of this research.
2.2.1 Edge Detection Methods
Edge detection has its own reputation in the domain of image

processing. Region boundaries and edges are quite similar, since there is an
iterative alteration of intensity at the region boundaries. The edges recognized
by edge detection techniques are often non - continuous. To segment an
object or a region of desire from an image, one requires closed region
boundaries. The expected edges are the boundaries which lie between such
objects. So, edge based detection supports image segmentation. One such
edge detection technique was proposed by Canny [28], which utilizes optimal
smoothing filter to preserve the edges while performing image segmentation.
2.2.2 Thresholding Methods
The simplest method of image segmentation known as the

thresholding method was introduced by Shapiro et al [129]. The function of
this method is based on a clip-level (or a threshold value), which supports the
conversion of a gray-scale image to a binary image. The main objective of
this method is to choose the threshold value (or values when multiple-levels
are selected). Maximum entropy method, Otsu's method (maximum variance)
and k-means clustering are the most frequent thresholding techniques used in
industries. Threshold value acts as mid-value, and input values above or
below the nominated threshold value are finally displayed.
13
Thresholding methods are now being used in segmenting computed

tomography (CT), Magnetic Resonance Images (MRI), Ultrasonic images and
Positron Emission Tomography (PET) images. The basic concepts of
thresholding are used in the techniques proposed through this research work.
2.2.3 Compression Based Methods
Compression based methods explicate that the optimal segmentation

helps to minimize the overall possible segmentations, which further favour
the reduction of code length of the data [26, 54, 127]. The relationship
between these two ideologies is that the segmentation tries to identify patterns
present in an input image and to closely monitor the regularity present in the
image, which can be used in compression. Compression based methods
explain the process of segmentation done using texture and boundary shape
explain each segment by its texture and boundary shape. Each of the above -
said components are modelled using a probability distribution function and
their coding length is explicated as follows:
1. The boundary encoding briefs the fact that regions in input

images tend to have a smooth contour or smooth boundary.
This property is utilized by Huffman coding to encode the
difference chain code of the contours present in an input
image. Thus, smoother a boundary, shorter is the coding
length.
2. Texture is encoded by lossy compression technique, similar to

Minimum Description Length (MDL) principle. The length of
the given data model is estimated in approximation to the
number of samples times the entropy or the energy value of
the model. A multivariate normal distribution, having the
closed form expression is used to model the texture present in
14
each region. The estimated entropy offers an upper boundary

to the true entropy of the data. This is considered to be an
interesting property of the model described above. When
compared with other distributions of specified mean and
covariance, normal distribution has the largest entropic values.
Thus, the minimization factor offered by an algorithm is quite
higher when compared with the true coding length.
Segmentation of an image by compression technique is performed

by encoding the bits of an input image. The segmentation technique
producing the shortest coding length is considered to be an effective one. This
segmentation is achieved by a simple clustering methodology. The distortion
imparted by the lossy compression determines the coarseness of the
segmentation and the optimal value produced by lossy compression varies for
each image. Optimal value parameter can be reckoned using a heuristic
approach exhibited upon the contrast of textures present in an image.
2.2.4 Histogram-Based Method
Histogram-based methods produce efficient segmentation results in

comparison with other image segmentation methods, because they require
only one entry through the pixels [95, 105]. By this technique, a histogram is
calculated from all the pixels present in the image. The peaks and valleys
present in the histogram are used to allocate the clusters in the image. Colour
or intensity values can be used to figure out the peaks and valleys.
On iterative application of histogram based segmentation method,

the input image is partitioned into several smaller clusters. This procedure is
repeated with the formation of smaller and smaller clusters until no more
clusters are formed. One disadvantage of the histogram-seeking method is
that it is difficult to identify the peaks and valleys in the input image.
15
Histogram-based approaches can also be quickly modified to occur upon

multiple frames, while preserving their single pass efficacy. Histogram
procedures can be done in multiple facets when multiple frames are being
processed. The same procedure adopted in processing one frame can also be
applied to process multiple frames. This process is followed as the final
results are merged, so that the identification of peaks and valleys is made easy
and found to be clearly distinguishable. The histogram methodology can also
be applied on a single pixel, where the information results of each pixel are
used to define the most frequent colour for the pixel location. This approach
performs segmentation based on active objects and a static environment,
resulting in a different type of segmentation, quite useful in video tracking
applications.
2.2.5 Region-Growing Method
The first region-growing method introduced was the Statistical

Region Merging method [96]. Also, this methodology utilizes a set of seeds as
input along with the input image. The seeds denote each of the objects to be
segmented. The regions are grown iteratively by comparing all unallocated or
non-grouped neighbouring pixels to the regions. Both pixel‘s intensity value
and the region‘s mean are used as the differential parameters to measure the
similarity. The pixel with the smallest differential value measured in this
procedure is assigned to the respective region. This process takes place until
all pixels are assigned to a region. Seeded region growing requires seeds as
additional input. The segmentation results are entirely dependent on the
choice of seeds.
Noises present in the input image can cause the seeds to be poorly
placed. Unseeded region growing is a modified version of region growing
algorithm, which does not require seed value initialization. This method
performs the segmentation process with a single region and the corresponding
16
pixel value present in the chosen region does not influence final segmentation.
On the completion of a single iteration, the unseeded region growing
algorithm considers the neighbouring pixels in the same way as seeded region
growing. It is different from seeded region growing, as it uses the minimal
differential value which is less than a predefined threshold, based on which
the pixel is added to the particular region. If this criterion is not met, then the
pixel is considered to be different from all the current regions and a new
region is shaped out with this pixel.
Another variant of this technique, proposed by Haralick and Shapiro

in the year 1985 is based on the intensities of pixels. The mean value and
scatter of the region and the intensity level of the candidate pixel is used to
define a test statistic. If the test statistic is sufficiently small than the described
threshold, the pixel is appended to the region and the region‘s mean and
scatter are recalculated. Otherwise, the pixel is forbidden and is used to form
a new region.
A special region-growing method, called-connected segmentation is

based on pixel intensities and neighbourhood-linking paths. A degree of
connectivity (connectedness) will be reckoned based on a path that is formed
by pixels. Two pixels are to be called-connected, if there exists a path linking
two pixels and the connectedness of this path is at minimal levels.
Connectedness is a similarity relation, which favours image segmentation.
2.2.6 Split-and-Merge Method
Split-and-merge segmentation is dependent on a quad ‗V‘ tree

partition of an image. It is also called the Quadra segmentation. This method
begins at the root of the tree that represents the whole input image. If non –
uniformity is observed, then the tree structure further splits into four sections
and so on. On the other hand, if four son-squares are homogeneous, they can
17
be grouped as a single connected component. The node present in the tree

structure is a segmented node. This process continues repeatedly until no
further splits or merges could be made. When the split-and-merge
methodology is influenced by a special data structure sequence, then the
implementation of this technique is quite easy and the methodology requires
lesser computational time for execution.
2.2.7 Partial Differential Equation-Based Method
An image can be segmented by numerically solving a partial

differential equation (PDE)-based method [29]. Curve propagation is a
popular procedure in PDE based method, by which numerous applications
related to object extraction, object tracking, stereo reconstruction, etc. are
performed. The prime idea of using this technique is to find out an initial
curve towards the lowest potential of a cost function, where its description
relates the task to be processed. For most of the converse problems,
minimization of the cost function is of much importance and it proposes
certain smoothness constraints on the solution, which can be represented as
geometrical constraints on the evolving curve.
2.2.8 Parametric Method
Lagrangian techniques depend on parameterization of the contour

according to certain sampling strategies and then produce element values
according to images and internal terms [67]. Such techniques are fast enough
and efficient in producing segmentation results; however, the original "purely
parametric" formulation is generally critiqued for its limitations in choosing
the sampling strategy, topology changes, addressing problems in higher
dimensions, the internal geometric properties of the curve, etc. Recently,
efficient "discretized" formulations have been evolved to address the
limitations incurred during high efficiency maintenance. In the above said
18
cases, energy minimization is generally organized using steepest-gradient

descent techniques.
2.2.9 Level Set Method
Osher and Sethian [100] proposed a level set method to track

moving borders or interfaces spread across various imaging domains. It can
be used to effectively handle the problem of curve or surface propagation in
an image. The prime idea is to represent the developed contour using a signed
function, where its zero level relates the actual contour. Using the motion
equation of the contour, one can easily derive a similar flow for the unclear
surface, sequentially applied to the zero-level illustrating the propagation of
the contour. The level set method encloses several advantages: it produces
parameter free - direct estimation of the geometric properties of the
developing structure.
This technique is intrinsic and can change the topology. Further,

these methodologies can be used to define an optimization framework and
was stated by Zhao, Merriman and Osher in the year 1996. Therefore, one can
make conclusion that this framework is quite convenient to address various
applications of computer vision and medical image analysis. Continuous
evolution of this method has offered better segmentation results.
Atiq Islam et al [15] utilized a multi - resolution fractal model called

the multifractional Brownian motion (mBm) for the identification of tumor
region, which are of different sizes and textures. The Jaccard scores produced
by this method require improvement.
19
2.2.10 Fast Marching Method
The fast marching method has been used in image segmentation by

Forcade et al [43], as the model utilizes both positive and negative
propagation speed values for the segmentation of an image. This approach can
also be called the generalized fast marching method.
2.2.11 Graph Partitioning Method
Graph partitioning methods favour efficient image segmentation

[175]. In using these methods, the input image is represented as an undirected
weight based graph. The dissimilarity between the neighbourhood pixels is
usually assessed with the help of a pixel or a group of pixels associated with
nodes and edge weights. The image represented graph structure is later
partitioned as the same procedure followed to model "good" clusters. Each
partition of the nodes (pixels) represent the outputs derived from the
algorithm and is said to be a segmented object of the image. Some of the
frequently used algorithms which lie under this category are normalized cuts,
random walker, minimum cut, isoperimetric partitioning and minimum
spanning tree-based segmentation.
2.2.12 Watershed Transformation
The watershed transformation utilizes the gradient magnitude of an

image, which acts as a topographic surface [124]. Pixels having the highest
Gradient Magnitude Intensities (GMIs) form the watershed lines, and they
denote the region boundaries. Water positioned on any pixel, which is
enclosed by a common watershed line, flows downhill to a collective Local
Intensity Minimum (LIM). Pixels draining to a common minimum form a
catch basin representing the segmentation process.
20
2.2.13 Model Based Segmentation
The main aim of using this approach is that the structures of organs
have a monotonous form of geometry [119]. Hence, variations in the shape of
organs could be pinpointed using model based segmentation. This process
requires three steps to be implemented such as (i) registration of the training
examples, (ii) probabilistic representation of the variation of the registered
samples and (iii) statistical inference between the model and the image. State
of the art knowledge-based segmentation procedures involve active shape and
appearance models, active contours and deformable templates and end up
with level-set based methods.
Atlas-based segmentation of pathological MR brain images using a

model of lesion growth [84], processes T1 weighted coronal type images and
requires a manual expert support to segregate tumor regions and lesions.
Atlas renormalization for improved brain MR image segmentation

across scanner platforms suggested by Xiao Han and Bruce Fisch [165] uses
an intensity renormalization to facilitate neuroanatomical imaging studies.
2.2.14 Multi-Scale Segmentation
Image segmentations are computed at multiple scales in scale space

and sometimes approached from coarse to fine scales. Segmentation criteria
can be subjectively complex and may consider both global and local criteria
present in an image. A common requirement is that each region considered
for segmentation must be uniform in some sense.
2.2.15 One-Dimensional Hierarchical Signal Segmentation
Witkin's [164], delivered an inspiring work in scale space included

the ideology that a one - dimensional signal could be definitely segmented
21
into regions, with one scale parameter controlling the scale or rate of
segmentation. Hierarchical relations between segments at different scales are
denoted by the zero-crossings of the second derivatives (minima and maxima
of the first derivative or slope) of multi-scale-smoothed versions of an image
from a nesting tree. Specifically, slope extreme at coarse scales can be traced
back to corresponding features at fine scales. Normally three segmented
regions are produced as a result of intervention of slope maximum and slope
minimum. When both these two values annihilate at a larger scale, the three
segmented regions merge together to form a single segmented region, which
defines the functioning of hierarchy of segmentation.
2.2.16 Semi-Automatic Segmentation
In this type of segmentation, the user outlines the Region of Interest

(ROI) with the mouse and the algorithms are applied so that the path best fits
the edge of the image is implemented. This is how the segmentation
procedure is accomplished using semi – automatic algorithm and one such
algorithm is used in this work and it is called the seed based region growing
algorithm. Initial level ROI is selected using a mouse click and that region or
voxel act as the seed point based on which the segmentation is performed.
Tumor-Cut: segmentation of brain tumors on contrast enhanced MR

images for radio surgery applications [8] is capable of segmenting only
contrast enhanced T1 weighted images. The usage of the tumor cut algorithm
is limited towards T1 contrast enhanced images and not upon other modalities
of brain images. Tumor detection is done using the basic principle of Region
of Interest (ROI) and requires manual assistance for segmentation procedures.
22
2.2.17 Trainable Segmentation
Most segmentation methods rely upon the colour information of

pixels present in an image. While performing image segmentation, humans
use a wider knowledge than colour information. But, implementing this
knowledge would require ample computation time and huge knowledge based
database, which are presently not available. There are trainable segmentation
methods, which can model some of this knowledge and can produce better
results when compared to conventional knowledge based methods.
Neural Network segmentation relies on processing small areas of an

image using an artificial neural network or a set of neural networks. After
such initial level processing, the decision-making mechanism mark the areas
of an image accordingly to the recognition offered by the neural network. A
type of neural network designed specifically for this purpose is the Kohonen
map and one such unsupervised neural network technique is used in this
work.
Pulse-coupled neural networks (PCNNs) are neural models

characterized based on functioning of a cat‘s visual cortex and further
developed for high-performance biomimetic image processing. Eckhorn
(1989) introduced a neural model to resemble the mechanism of a cat‘s visual
cortex. The Eckhorn model was extensively used for studying and analysing
the visual cortex of small mammals and was soon recognized as having
significant application in the domain of image processing. The Eckhorn
(1994) model was extended as an image processing algorithm by Johnson,
who named the algorithm as Pulse-Coupled Neural Network.
Over the past ten years, PCNNs have been extensively utilized for a
diversified image processing applications, such as image segmentation,
feature generation, face extraction, motion detection, region growing, noise
23
reduction and so on. PCNN is a two-dimensional neural network technique in

which each neuron in the network corresponds to one pixel in an input image
and this correspondence is established by receiving corresponding pixel‘s
colour information (e.g. intensity). Thus colour information acts as an
external stimulus. Based on the reception of local stimuli from the
neighbouring neurons, interconnection is established between each neuron
and neighbouring neurons. An internal activation system is responsible for
holding up the external and local stimuli. The activation system accumulates
the stimuli until it exceeds a dynamic threshold, resulting in a pulse output.
Through recursive computation, PCNN neurons produce sequential series of
pulse outputs. The sequential series of pulse outputs contain valid information
of input images and this information can be used in various image processing
applications, such as image segmentation and feature generation.
Compared with the conventional image processing means, PCNNs

have several significant advantages, which include immunity against noise,
individuality of geometric variations in input patterns, capability of linking
minor intensity variations in input patterns, etc.
Two fully-unsupervised methods for MR brain image segmentation

using SOM-based strategies are recommended by Oritz et al [99]; segment T1
weighted images are based upon SOM (Self Organizing Map) technique. The
usage of the segmentation algorithm has been confined to process T1
weighted images. Further, the sensitivity value produced by the algorithm
requires improvement.
Logeswari and Karnan [78] reported brain tumor detection using a

segmentation process based on Self Organizing Maps (SOM). The
segmentation process is curbed to segment T2 weighted image sequences.
24
Logeswari and Karnan [79] offered a brain tumor detection using

segmentation process based on Hierarchical Self Organizing Map (HSOM)
technique, where T2 weighted images were segmented in an average time
period of 29.9708 seconds, which requires further minimization.
Yan Li and Zheru et al [167] performed MR brain image

segmentation using SOM, which requires reduction in MSE values.
Shen et al [130] projected a novel FCM algorithm, named Improved

FCM (IFCM). Two important parameters that address neighborhood pixel
attraction are suggested through this work. The first parameter is concerned
with the feature difference between neighbouring pixels present in the image and
the second parameter is linked towards the calculation of relative location of the
neighbouring pixels. This methodology performs segmentation based upon the
neighbouring pixel‘s intensities and their specific locations. Choosing optimum
parameters still remains a tedious procedure as it paves the way for inaccuracies
in image segmentation and stands to be a major hindrance in utilizing IFCM for
real time applications. Moreover, IFCM requires an Artificial Neural Network
(ANN) for computing these two parameters. Designing ANN and assigning the
control parameters using ANN is a complicated process and results in higher
time consumption for execution. Inappropriate selection of control parameters
leads to improper image partitioning and the effort taken to offer effective image
segmentation proves to be futile. Statistical classifiers use preset learning rules to
group the voxels present in the input image. It accomplishes clustering
techniques which classify voxels in an unsupervised mode, by which similar
voxels are grouped into the same class.
A similarity protocol is to be extensively studied to analyze the

grouping of voxels into a class. As said by Riad et al [118], classifiers are
capable of generating different classes in which group of voxels with the same
properties are held together.
25
Generalization property exhibited by the support vector classifiers

suggested by Vapnik [153] has favoured its utilization in the domain of image
segmentation. Parra et al [103] also explained that the functioning of these
classifiers, which are governed by statistical learning theory.
Sahoo et al [120], Yang and Laaksonen [169], Guler et al [47], Ong

et al [97], Alirezaie et al [5], Sun [138] and Fan et al [40] have developed
segmentation techniques using ANN classifiers.
Segmentation algorithms put forth by Guler et al [47], Ong et al

[97], Alirezaie et al [5], Sun [138] and Jiang et al [59] have utilized
unsupervised neural network technique namely, self-organizing maps (SOM).
Elliott et al [33] have introduced an algorithm which identifies

multiple sclerosis lesions. The algorithm utilizes Bayesian classifier to
identify the brain tissues and random-forest for lesion identification. The false
detection rate produced by the algorithm requires significant reduction.
Segmentation of T1 – Weighted images alone is done using an

automated algorithm, which combines stationary wavelets and Self
Organizing Map (SOM) was developed by Ayse Demirhan et al [17].
Ulagammai et al [149] used wavelet based neural network technique,

initially trained by BFO algorithm for load forecasting.
Anthony Bianchi et al [10] estimated the advancements of mild

traumatic brain injury (mTBI) from Magnetic Resonance Imaging (MRI).
This method utilizes the texture features available in an MR image with
further assistance offered by a Probabilistic Support Vector Machine (P -
SVM), which enhances the differentiations in unimodal MR images. The
overall time duration required for executing the algorithm is found to be less
26
than five minutes. This recommended algorithm suffers from computational

complexity.
Vijayakumar and Ashish Chaturvedi [15] have utilized Gaussian

filter for image pre – processing, fluid vector flow for feature extraction and
brain tumor extraction, and Support Vector Machine (SVM) for final
classification of abnormalities. This method is capable of segmenting T1 and
T2 – Weighted MR brain images and not other MR sequences. Further, the
computational time involved in segmenting the images is also high.
Kasiri et al [68] have utilized Least Square – Support Vector

Machine (LS - SVM) to segment the Grey Matter (GM) and White Matter
(WM) regions. The training samples for LS – SVM is obtained from the
registered brain atlas domain. The suggested methodology is in need of
manual intervention to complete the segmentation process. Moreover, training
and testing procedures of the SVM classifier has led to computational
complexity.
Applications of Image Segmentation:
The identification and segmentation of the brain tumor will be much

helpful in the following domains:
1. Surgery planning.
2. Virtual surgery simulation.
3. Intra-surgery navigation.
2.2.18 Hybrid Segmentation Techniques
Some of the recent advancements in image segmentation are the

development of hybrid algorithms, which combine the working principles of
27
two or more techniques and are capable to deliver different segmentation

results. It has been observed that different combinations of soft computing
techniques have been tried under this category, which would be of great
beneficiary to clinical oncology.
A neural network based method for brain abnormality detection in

MR images using Gabor wavelets [7] segments T1 and T2 weighted MR brain
images. 0.025 is the average MSE value obtained through segmenting the
input MR brain images, which requires minimization.
An improved MR brain image segmentation to detect cerebrospinal

fluid level using anisotropic diffused Fuzzy C Means [139], segments T2
weighted images alone.
Automatic brain extraction methods for T1 magnetic resonance

images using region labeling and morphological operations [133] perform
segmentation only upon T1 weighted images using BEM (Brain Extraction
Methods). Tissues of brain are segregated from cerebrospinal fluid and skull
regions, automatically.
Automatic segmentation of Meningioma from non-contrasted brain

MRI integrating fuzzy clustering and region growing [141], segments non
contrasted T1 and T2 weighted images of patients suffering from
meningioma. Usage of the segmentation algorithm is narrowed down to
segment homogeneous tumor type (meningioma alone).
Fully automatic segmentation of multiple sclerosis lesions in

MRFLAIR brain images using adaptive mixtures method and markov random
field model [115] is limited towards processing FLAIR type images.
Similarity Index, Overlap Fraction and Extra Fraction values produced by the
above said segmentation algorithm require further improvement.
28
Identification and localization of prostate cancer with combined use

of T2-Weighted, Diffusion Weighted MRI and Proton MR Spectroscopy,
correlation with histopathology [18] segments T2 weighted and MRS
(Magnetic Resonance Spectroscopy) images. Sensitivity, specificity and
Signal to Noise Ratio values produced by the above said algorithm need to be
enhanced.
Image mosaicking for evaluation of MR brain tissue abnormalities

segmentation study [125], segments the tumor region only in T1 weighted
images.
The study of evaluating the impact of different factors on voxel-

based classification methods of ADNI structural MR brain images [114]
performs segmentation upon T1 weighted images. Latency acts to be a major
setback in theses voxel based classification methods.
Segmentation of multiple sclerosis lesions in brain MR images using

Spatially Constrained Possibilistic Fuzzy C-means Classification [50]
validates segmentation of T1 and T2 weighted images using FCM clustering
technique. Possibilistic FCM produces better segmentation accuracy, which is
confined to small and tiny lesions present in the input MR brain images.
Spatial information based image clustering with a swarm approach

[101], segments MR images using Artificial Bee Colony algorithm combined
with FCM clustering technique, where time consumption has to be carefully
observed.
A fully automated algorithm under modified FCM framework for

improved brain MR image segmentation [65] performs segmentation in T1
and T2 weighted axial type images using modified FCM clustering technique.
Increasing and decreasing the number of iterations for modified FCM
29
framework affect the overall computational time required for processing the
input MR brain images.
Improving fuzzy algorithms for automatic image segmentation [135]

describes an automated fuzzy algorithm for MR brain image segmentation.
Time consumption proves to be a major setback for the proposed automated
fuzzy algorithm. Also, this method requires improvement of segmentation
accuracy.
Segmentation of MR brain images for tumor extraction using fuzzy

[45] is capable of segmenting T1 weighted axial type images. Values of
PSNR and MSE produced by this method can further be improved. It has
shortcomings in terms of computational time, PSNR and MSE.
Nandha Gopal [94] referred an automatic detection of brain tumor,

capable of producing an accuracy of 95.16% in segmenting MR (Magnetic
Resonance) brain images, and it requires further improvement.
El-Hachemi Guerrout et al [38] used Markov Random Field Theory

for medical image segmentation. The algorithm performs medical image
segmentation on a Cluster of PCs using Markov Random Field Theory.
Segmentation process is done only upon T1 and contrast enhanced T1
weighted images.
Ben George and Karnan [19] attempted to produce segmentation

results of MR brain images in a computational time of 31.21 seconds using
Bacteria Foraging Optimization (BFO) technique. Computational time of the
algorithm plays a major role in the process of diagnosing. Lesser
computational time supports the radiologist to segment large volumes of
clinical data.
30
Sultan Aljahdali et al [136] suggested an automated fuzzy rule based

algorithms for reliable image segmentation. T1 weighted images were
segmented using both fuzzy K means and Kernel based Fuzzy C Means
(FCM) algorithms. Time consumption required by the algorithm plays a
major constraint in segmenting the images. Also, the algorithm produces a
segmentation accuracy ranging between 59.454 and 80.46, which can further
be incremented.
Vasuda and Satheesh [154] introduced an improved FCM algorithm

for MR (Magnetic Resonance) brain image segmentation. Reduction of
convergence rate of the proposed algorithm remains a major setback.
Artem Mikheev et al [14] have recommended a segmentation

algorithm which works only in segmenting high resolute T1 weighted images
using techniques such as thresholding, surface detection, connectivity and a
novel constrained growing operator.
Vida Harati et al [155] have utilized Fuzzy Connectedness (FC) for

segmenting and identifying tumor region only in T1 weighted images. Yu and
Yang [174] recommended a generalized FCM (GFCM) structure to
consolidate the variations found in FCM and then observed its optimality test
using parameter selection. GFCM does not involve the optimality test
associated with both clusters and membership functions. Rather, GFCM
involved only optimality test with the cluster prototype.
Yu and Yang [173] recorded a substitution method for GFCM and

named it as Generalized Fuzzy Clustering Regularization (GFCR). GFCR is
associated with the optimality test done using membership functions.
31
Certain statistical classifiers rely on algorithms such as expectation-

maximization algorithms (EM) recommended by Kapur et al [64], Tsai et al
[147], Xie et al [166], Maximum Likelihood (ML) estimation referred by
Smith [132] or Markov random fields recommended by Wells et al [160] and
Mohamed et al [89].
Liyuan et al [77] recorded that K-means and Fuzzy k-means are

widely used in the field of image segmentation as they avoid abrupt
transferences during the classification process.
DICOM based MR brain image segmentation offered by a hybrid

approach of fuzzy techniques proposed by Uma Maheswari and Radhamani
[151] shows better segmentation result than the conventional Otsu‘s
segmentation process.
Brain tumor detection of MR images using color-converted hybrid

PSO and K-means Clustering segmentation proposed by Rajalakshmi and
Lakshmi Prabha [113] is a color-based MR brain image segmentation method
that uses hybrid PSO (Particle Swarm Optimization) and K-Means clustering
technique to track the tumor region in MR brain images.
Diagnosis of brain tumor through MRI using image processing

clustering algorithms such as Fuzzy C - Means along with intelligent
optimization techniques recommended by Nandha Gopal and Karnan [93]
uses Fuzzy C - Means with Particle Swarm Optimization algorithm to
segment the tumor part present in MR brain image and produces an accuracy
rate of 92.43%, which can further be increased.
A high speed parallel Fuzzy C - Means algorithm for segmenting

brain tumor proposed by Murugavalli and Rajamani [91] is capable of
32
segmenting MR brain images at higher speed rate and requires average time
duration of six seconds for processing 256 × 256 input brain images.
MR image segmentation using a power transformation approach

referred by Juin-Der Lee et al [61] segments T1 weighted images alone.
A complete automatic brain extraction algorithm (BET – Brain

Extraction Tool) for axial T2 - weighted magnetic resonance images
developed by Somasundaram and Kalaiselvi [134] produces segmentation
results only upon T2 weighted images.
Tumor growth estimation proposed by Islem Rekik et al [58]

performs tumor segmentation using seed point initialization strategy and its
usage is confined to process MR brain images affected with low grade
gliomas.
Hanning Chen et al [48, 49] have proposed adaptive and cooperative

type bacteria foraging optimization algorithms. Both these algorithms have
the limitation of utilizing excessive run length time for exploring better food
source.
Vijayakumar et al [157] utilized Multi – Layered Perceptron (MLP),

Self Organizing Map (SOM) and Radial Basis Function (RBF) for segmenting
contrast enhanced and diffusion weighted images. Sensitivity value produced
by these algorithm ranges between 92.47 and 94.37, which could further be
elevated.
Lim et al [74] proposed a method which segments color images and

smoothes the gray-level histogram using Gaussian convolution technique.
This methodology requires high computational time for processing the input
color images.
33
Du-Ming Tsai [42] introduced a fast threshold selection procedure

for unimodal and multimodal histograms. This methodology comprises
Gaussian kernel smoothing and local maximum curvature techniques to detect
peaks and valleys in a multimodal histogram and utilized points of
discontinuity for a unimodal histogram. This technique requires large time
consumption for processing the input images.
Lio et al [75] provided a recursive algorithm, which greatly reduces

the computational complexity in assessing multilevel threshold values. A
look-up table is used to determine the threshold values and this technique
performs far better than the conventional Otsu and Kapur methods. The
critical shortcoming faced by this technique is that the number of thresholds
―m‖ increases; the processing time too increases.
The most popular clustering approach used to segment medical

images is the Fuzzy C - Means (FCM) clustering and expectation–
maximization (EM) algorithms. Applications of the EM algorithm to MR
brain image segmentation are reported in [38] and [45]. A general
disadvantage of EM algorithms is that the normal intensity distribution is
developed to favor the segmentation of noiseless brain images and not the
images filled with noise. FCM algorithm and its various forms have also been
applied in several other research activities.
Introduction of FCM clustering by Bezdek [22] focuses in

minimizing an objective function perpended to the fuzzy membership and a
set of cluster centroids. In presence of noise signals, FCM increases the
membership function, resulting in higher computational complexity and
inefficiency in image segmentation. To overcome these hindrances, Tolias
and Panas [144] made further processing upon the membership function so as
to smooth the effect of noise. Dave [36] recommended a technique called
Noise Cluster (NC), which could perform clustering on noisy input images.
34
Removal of noisy pixels is done by NC and it could not be used in the domain
of image segmentation, as there is a loss of vital information during the noise
removal process.
Pham [107] rendered a new approach of FCM and named it as

Robust Fuzzy C - Means (RFCM) algorithm. Ahmed et al [2] developed
certain revised versions of FCM by introducing modifications in the objective
functions. All the above said methodologies suffer from unavoidable
computational complexities, even though modifications of most equations
along with the modification of the objective function are accomplished.
Shan et al [126] developed the automated brain segmentation of T1

weighted 3D MR brain images using histograms. A fully automated 3D brain
image segmentation using elastic template matching approach trailed by a
stochastic heuristic approach is presented by Marroquin et al [82]. It requires
higher computational time for processing the input images.
Zhen Ma et al [176] projected a review on the current segmentation

algorithms used for medical images. The algorithms are categorized according
to the usage of methodologies based on thresholds, deformable models and
clustering techniques.
Some of the recently developed evolutionary algorithms introduced

by Passino [104] and Sathya et al [123] are Bacterial Foraging (BFO)
algorithm and Amended Bacterial Foraging optimization algorithm (ABF).
The Peak Signal to Noise Ratio (PSNR) values produced by ABF, suggested
by Sathya et al [123], needs improvement. The usage of this algorithm is
narrowed down to segment T2 – Weighted MR brain images.
BFO algorithm has been extended to the areas of harmonic

estimation, as suggested by Mishra [87].
35
BFO algorithm has been used in the reduction of transmission loss as

recommended by Tripathy et al [146].
Load compensation by active power filter has been performed using

bio inspired BFO algorithm developed by Mishra and Bhende [86].
Tripathy and Mishra optimized real power and voltage stability

using Bacteria Foraging Optimization algorithm [145].
Acharya et al [1] suggested a BFO based independent component

analysis and Datta et al [35] put forth an improved BFO algorithm to have an
efficient and quicker optimization of antenna array.
Albert Huang et al [3] have formulated a deformable model for

performing an automatic segmentation of 3D MR brain images. This method
requires improvement in the domains of White Matter (WM) segmentation
and tumor identification.
Ayelet Akselrod-Ballin et al [16] have developed a 3D MR brain

image segmentation algorithm using Atlas guidance and SVM classifier. The
Jaccard Index scores produced by this methodology require amplification.
An Auto-Context based MR brain image segmentation proposed by

Zhuowen Tu and Xiang Bai [177] performs segmentation in 3D MR brain
structures and it suffers from computational complexity. The algorithm
developed is a supervised model and it requires initial level sincere training
modules (necessity of abundant ground truth values for training), which
ultimately results in higher time consumption.
Arnaldo Mayer and Hayit Greenspan [12] presented an Adaptive

Mean-Shift algorithm for MR brain image segmentation. The methodology is
in need of large computational time to perform the segmentation process. The
36
average time duration of processing the four modalities of MR brain images is

30 minutes, which has to be considerably reduced.
Pham and Klaus Berger [148] propounded an automatic detection of

changes observed in White Matter (WM) region of elder people. This
methodology relies upon fuzzy clustering, possibilistic clustering,
geostatistics and knowledge combination techniques. The total number of
under detections observed in WM region needs to be minimized.
Mariano Rivera et al [80] proffered a novel Markov Random Field

(MRF) controlled by entropy values for image segmentation. The
segmentation results obtained using this methodology make a clear cut
illustration that the improvement of Jaccard Tanimoto Coefficient values and
reduction in time consumption are strongly required.
Dimitrios Zarpalas et al [37] put forth Hippocampus region

segmentation using Active Contour Model (ACM), enhanced by Gradient-
Based reliability map technique. The shortcoming observed in this technique
is the huge time consumption involved during non – rigid registration process.
Moumen El-Melegy and Hashim Mokhtar [90] developed a

prior-information-guided FCM (PIGFCM) for tumor detection in MR brain
images. To identify the tumor region subjected to bias field or intensity
inhomogeneity, PIGFCM requires fine tuning. Moreover, this algorithm has
the capability to identify tumor region present only in T1 weighted images.
Ipek Oguz and Milan Sonka [57] developed an algorithm named as

LOGISMOS - B (Layered Optimal Graph Image Segmentation of Multiple
Objects and Surfaces) for segmenting the cortical surfaces present in MRI
human brain. For segmenting White Matter (WM) and Grey Matter (GM)
regions, LOGISMOS – B requires 2.55 hours per subject.
37
Christos et al [32] presented a new concept in MR image analysis,

called the MR physics simulator (MRISIMUL) of the Bloch equations. This
method helps the radio surgeon to have an initial level ideology in identifying
the anomalies present in the patient, as MRISIMUL provides MRI
simulations. The overall computational time for modelling out an MR image
using MRISIMUL is found to be 2.2 min. Here, time consumption is a major
shortcoming.
Yaqub et al [170] have rendered an MR image segmentation and 3D

volumetric fetal femoral ultrasound image segmentation using RF (Random
Forests) technique. Frequent tuning of parameters and high training time are
the major setbacks of this methodology.
Ertan Cetingul et al [39] utilized Orientation Distribution Function

(ODF) to model diffusion and perform segmentation upon Diffusion MR
Images (DMRI). ODF combined with clustering approach is also used to
segment DMRI. The proposed technique is vulnerable to computational
complexity, as frequent parametric tuning for clustering is needed.
Zhou et al [95] suggested an improved Fuzzy C – Means (FCM) or

MMTDFCM (Measure of Medium Truth Degree based FCM), which entirely
relies on Measure of Medium Truth Degree (MMTD) for MR brain image
segmentation. The Peak Signal to Noise Ratio (PSNR) values produced by
this method has to be eventually increased.
Salima et al [101] have recommended a Spatial Fuzzy C – Means

(SFCM) optimized by Artificial Bee Colony (ABC) algorithm. This ABC
based SFCM is capable of effectively segmenting T1 weighted images and
not the other modalities of MR brain images.
38
Karim Kalti and Mohamed Mahjoub [66] have utilized Gaussian

Mixture Model (GMM) and Modified Fuzzy C – Means (MFCM) framework
for MR brain image segmentation. The entire run time of the algorithm is
about 72.28 seconds, which requires minimization.
Shantha Kumar and Ganesh Kumar [128] have recommended a pre-

processing, anisotropic diffusion, feature extraction and classification
procedures for tumor identification in MR brain images. For classification,
SVM is used and this methodology requires an average computational time of
36 seconds for processing the input images. The time duration for processing
the input images is quite low and the algorithm is executed on a desktop fitted
with advanced i3 processor with 2GB RAM.
Prabhjot Juneja et al [109] contributed a Validation Index (VI) to

assure the accuracy of the automated segmentation algorithms, which suits
radiotherapy procedures. A new control parameter derived from the opinion
offered by multiple experts is used to validate or evaluate the automated
segmentation algorithms. The major drawback of this technique is the limited
usage in assessing clinical images. Large volumes of clinical data could not
be assessed with the above said methodology.
Huang et al [83] developed a brain tumor segmentation using local

independent projection-based classification (LIPC) technique to group the
voxels present in the input image. The patch type characteristic is insufficient
to distinguish the brain tumor segmentation process, as it is quite complex to
analyse and represent the tissue regions and lesions in MR brain images.
Mohamad et al [88] have utilized Improved FCM algorithm for

segmenting MR brain images, where Breeding Swarm (BS) optimization
values are assigned as the input parameters for FCM algorithm. The usage of
this algorithm is narrowed down to segment T1 weighted images.
39
Sarah Parisot et al [122] developed a methodology to simultaneously

address both image segmentation and registration using a combined pairwise
discrete Markov Random Field (MRF) model on a sparse grid overlaid upon
the input image or the image domain. The main drawback of this algorithm is
that the classification process exhibited by Gentle Adaboost algorithm
developed by Friedman et al [44] requires continuous tuning, which requires
higher time duration to carry out the segmentation process.
Resmi et al [117] suggested an automatic segmentation algorithm to

segment the tumor region, WM and GM present in T1 –, T2 – Weighted and
FLAIR type MR brain images. The usage of the proposed algorithm is
confined to segment low grade astrocytoma and high grade gliomas and not
the other forms of aberrant regions.
Hassan Khotanlou et al [69] suggested fuzzy based classification

approach to detect tumor region in MR brain images. This method also
includes symmetry analysis for tumor identification and spatially constrained
deformable models for tumor surface evolution. The limitation of this method
is that the symmetry analysis fails to evaluate the tumor region present in the
mid sagittal plane. Moreover, the methodology requires 3.5 min for producing
the segmentation results.
Siyal and Lin Yu [131] recommended a Modified approach to FCM

algorithm to derive out MR brain image segmentation. Computational
complexity has upper hand in the case of modified – FCM framework, which
requires primary addressal.
Kannan et al [62] developed a FCM algorithm whose functions

depend upon the Kernel based induced distance calculation procedure. This
technique has the ability to segment only T1 and T2 weighted images. The
overall time consumption required by the suggested methodology for
undergoing four iterations could be slightly reduced.
40
Hung et al [162] suggested a Supressed FCM framework (S - FCM)

for MR brain image segmentation and this technique requires ample time
duration to process the images and to produce segmentation results. It is
found that S – FCM requires 5min in average for processing the images.
Hou et al [55] presented a Regularized FCM approach (R – FCM)

for brain tissue segmentation. The methodology suffers from computational
complexity.
Kannan et al [63] suggested an automatic Robust FCM algorithm for

segmenting medical images. The methodology segments T1 and T2 weighted
MR brain images, where the calculated segmentation accuracy requires
improvement.
George et al [81] developed a novel segmentation algorithm using

image registration techniques to remove false positives during diagnosis. This
process is accomplished using segmented anatomical data present in atlases.
This methodology is mainly rendered for the examination of basal ganglia
(including the thalamus region) for an accurate detection of human spongiform
encephalopathy in MR brain image sequences such as, T1 –, T2 – Weighted and
FLAIR – T2 Weighted. Creutzfeldt–Jakob Disease (CJD), sporadic
Creutzfeldt–Jakob Disease (sCJD) and variant Creutzfeldt–Jakob Disease
(vCJD) were examined with the support of this novel algorithm. It is observed
that the false positive values occur due to the presence of motion artefacts in the
input images. Thus the algorithm is vulnerable to motion artefacts.
Crum et al [163] introduced generalized overlap measures for

evaluating and validating the segmentation results of MR brain images. Lack
of gold standards has paved way to develop such an evaluation tool for
analysing medical image processing. Nonrigid registration problems have not
been properly addressed by the above said technique.
41
Berengere et al [20] have developed twenty brain digital phantoms

for validating the MR brain databases. The main shortcoming of this
methodology is that the mixing skull, tissue, Dura layer and fluid region occur
during the development of brain phantoms.
Hayit et al [52] suggested a Constrained Gaussian Mixture Model

(CGMM) for automatically segmenting noisy, low-contrast MR brain images.
Computational complexity plays a major role in segmenting the images using
CGMM.
Xia et al [168] recommended a knowledge driven algorithm to

segment the Caudate Nucleus present in MR brain images. Spatial overlap
metric values produced by the segmentation algorithm require improvement.
Yongxin et al [171] developed a novel automatic MR brain image

segmentation algorithm by combining atlas registration, Fuzzy Connectedness
(FC) and Parametric Bias Field Correction (PABIC) strategies. This algorithm
is inapplicable to the images filled with severe intensity overlapping.
Lisa Jonasson et al [76] presented a technology to view Diffusion

MRI (DMRI) in a 5D spatial representation, so that to have a distinguished
view upon the white matter tracts. Segmentation is afforded by thresholding
method. Segmentation methods presented in this method identify White
Matter (WM) regions and are unable to measure the connectivity between
gray matter regions as similarly done in tractographic procedures.
Youyong et al [172] have utilized discriminative clustering and

feature selection to perform MR brain image segmentation. Information
Theoretic Discriminative Segmentation (ITDS) and supervised Information
Theoretic Discriminative Segmentation (SITDS) are the two techniques
recommended by Youyong et al, which require fine tuning to produce better
segmentation accuracy.
42
Thomas and Gordon [142] suggested an Open – Box spectral

clustering in application towards medical image analysis. Through spectral
clustering, 3D medical image analysis could be effectively done. The major
hindrance faced in converting a spectral clustering procedure to an open-box
tool procedure is the visualization of high – dimensional spectral feature and
interlinking it to the input 3D data.
2.3 MOTIVATION
Automated detection of tumor region and segregation of tumor

region from dense edema portions has not been instantiated. The imaging
systems utilized in hospitals do not favor the following conditions:
1. Normal tissue regions cannot be segmented by a conventional

image processing tool available in a scan center.
2. Tumor region bounded between the tissues of the brain cannot

be efficiently segregated.
3. Differentiated view between the tumor infiltration levels and

the edema covered portions are quite impossible.
4. Larger data volumes require maximum time duration and

manpower for a better diagnosis.
These conditions prove to be a major challenge for a radio surgeon,

in diagnosing patients suffering from different pathologies, lesions and tumor
nature.
If an automated view is made available, a radio surgeon can

efficiently plan towards pre and post radio therapeutic procedures. Further,
patient images can be efficiently analyzed within minimum time duration
while reducing the valuable time and energy of a radio surgeon involved in
the diagnosing process.
43
Accurate detection of the tumor region with its exact topology,

normally unknown by a radiologist is also made possible with the application
of the proposed soft computing techniques. The soft computing techniques
proposed through this research are capable of performing segmentation upon
both the tissues and identification of the heterogeneous tumor types present in
T1 – w, T2 – W, FLAIR and MPR type MR brain image sequences. Exact
segregation between tumor region and dense edema portion is made available
using afore - mentioned techniques. The research activities so far carried out
have addressed two to three types of tumor identification, tissue segmentation,
multiple lesions, and identification of blocks in circle of Willis and
segmentation of two to three modalities of MR brain images. The proposed
techniques through this research have the ability to process multimodal MR
brain images with the identification of heterogeneous tumor regions.
2.4 OBJECTIVES OF THE THESIS
The core idea of this thesis is to summarize the selected

methodologies for medical image analysis. It introduces several combinations
of Fuzzy and other methodologies to facilitate tumor diagnosis. The
objectives of the thesis are listed below:
1. To segment the tissues of the brain and to extract the tumor

region using Fuzzy (Fuzzy Inference System).
2. To create a novel combinatory algorithm using Self

Organizing Map (SOM) and Fuzzy K Means (FKM)
algorithm for tumor identification and tissue segmentation in
magnetic resonance brain images.
3. To produce a complete automated algorithm for the

segmentation of tissues and identification of tumor region in
T1, T2, MPR (Multiplanar Reconstruction) and FLAIR (Fluid
44
Attenuated Inversion Recovery) type brain images using

Particle Swarm Optimization and Fuzzy C – Means algorithm
(PSO based FCM).
4. To introduce tumor identification in T1, T2, FLAIR and MPR

(Multi Planar Reconstruction) brain images using a
combination of Optimization and Fuzzy clustering improved
by novel seed based Region Growing algorithm.
45
CHAPTER 3
DATA ACQUIREMENT FROM CLINIC
3.1 INTRODUCTION
Some of the sequences that are acquired through a MRI scanner are
T1, T2, FLAIR, and MRS. MRI sequence is a systematic combination of
radio frequency and gradient pulses designed to obtain the data to form the
image. The data to create an MR image are attained through a sequence of
steps. 1. In the presence of a slice select gradient, RF pulse excites the
magnetization for tissues. 2. The other two vital elements of the sequence are
phase encoding and frequency encoding (or) read out procedures, which are
necessary to localize the protons in space (spatial domain), that is the other
two dimensions. 3. After the collection of data, the process is iterated for a
sequence of phase encoding steps. The images of dimensions 256 × 256, 480
× 375, 1105 × 650 and 763 × 664 are used in this research.
3.2 DETAILS OF PATIENTS
Nearly 38 images acquired from ten different patients of different

age groups have been used in this work. Also, images from Harvard Brain
Web repository have been used. The details of the patients are listed below.
1. Clinical images of patient of age five suffering from PNET

(Primitive Neuro Ectodermal tumor) are used.
46
2. Clinical images of patient of age thirty five suffering from

meningioma are used.
3. Clinical images of patient of age thirty two suffering from

high grade Astrocytoma are used.
4. Clinical images of patient of age three suffering from

supratentorial Primitive Neuro Ectodermal Tumor are used.
5. Images of the patients suffering from Glioma and Metastatic

Bronchogenic Carcinoma, taken from Harvard Brain Web
Repository are used.
6. Images of the patients suffering from Pituitary Adenomas,

Nerve Sheath tumor, low grade Glioma and high grade
Glioma, taken from Harvard Brain Web Repository are used.
3.3 TYPES OF IMAGES ACQUIRED
3.3.1 T1 Weighted Image
A T1 weighted image is the one where contrast entirely depends on

the differences in the pulse times obtained from the tissues (say fat and
water). The echo spin time (TE) governs the amount of T1 decay that is
allowed to appear before the signal is received. To achieve T1 weighting, the
TE value must be long enough to let both fat and water time to decay. If the
TE value is too small, both fat and water would have no time to decay and
therefore their differences in T1 times are not illustrated in the image.
T1 – weighted images used in this research work have Slice Thickness (SL)
of 4mm, Repetition time (TR) and Echo spin Time (TE) values of 10ms and
550ms.
47
Figure 3.1 T1 weighted axial and coronary images
3.3.2 T2 Weighted Image
To show T2 –Weighted and T2 contrast enhanced images, specific

values of TR and TE are chosen for a given pulse sequence. The choice of
appropriate TR and TE values weight an image, so that one contrast
mechanism dominates over the other two. A T2 weighted image is one where
the contrast depends predominantly on the differences in the T2 times
between tissues e.g. fat and water. The Repetition Time (TR) governs the
expression of each vector and can recover before it is excited by the next RF
pulse. In order to obtain T2 weighted images, the TR value must be smaller so
that both fat and water have sufficient time to return to B0. Here, B0 is the
contrast level between fat and water. T2 – weighted images used in this
research have Slice Thickness (SL) 5mm, Repetition time (TR) and Echo spin
Time (TE) values of 117ms and 4000ms.
Figure 3.2 T2 Weighted axial image

48
3.3.3 Fluid Attenuated Inversion Recovery (FLAIR) image
FLAIR is another type of inversion recovery sequence. In FLAIR,

the electric signal from fluid, say cerebrospinal fluid (CSF), is nullified by
selecting a TI value corresponding to the time of recovery of CSF from 180 º
inversions to the transverse plane. The signal from CSF is blocked and
FLAIR type sequence is utilized to suppress the high CSF signal in T2 and
proton density weighted images, so that the pathologies nearby CSF region
are clearly visualized.
FLAIR type image sequences used in this research have Slice

Thickness (SL) of 5mm, Repetition time (TR) and Echo spin Time (TE)
values of 10ms and 550ms. In addition, an Inversion Time (TI) of 2500ms is
used to attain FLAIR images.
Figure 3.3 Fluid Attenuation Inversion Recovery (FLAIR) image
3.3.4 Multi Planar Reconstruction (MPR) Image
Multiplanar Reconstruction (MPR) image helps us to visualize the

anatomy of human brain in all possible axes. MPR is nothing but a post
processing reformation of a 3D data set into 2D slices of random thickness at
any angle view. For a better resolution, thin slice embodied with isotropic
voxel and overlapping slices with good signals are advantageous. An axial
MPR image is shown in Figure 3.4. MPR images of Slice Thickness (SL):
49
0.7mm, Repetition time (TR) and Echo spin Time (TE) values of 4.52ms and
1160ms are used in this research.
Figure 3.4 Multiplanar Reconstruction (MPR) image with Contrast

Enhancement
3.3.5 Contrast Enhancement
Contrast enhancement is induced for the above - said image

sequences to have a clear view of the pathologies present in the brain.
Contrast Enhancing agent GADO IV (Gadolinium IV) is used extensively in
the previously mentioned MRI sequences. Normally, contrast agents favour a
clear cut examination of the pathologies. Some of the contrast enhanced
image sequences used in this research are given below for better
understanding.
Figure 3.5 T2 Axial image sequence with Contrast Enhancement

50
Collecting live clinical data samples have consumed much time in

this research process. It took nearly two years to frame out the data to be used
in the soft computing techniques. Moreover, data accumulation of
heterogeneous tumor types is a tedious process and it too has been done to
verify the robustness of the proposed algorithms. Finally, the results produced
by the algorithms have been cross validated by an experienced radiologist.
3.4 DESCRIPTION OF THE MRI SCANNER
The images have been obtained using a Siemens – Aera 1.5 T

Magnetom Avanto-Tim technologies (World‘s advanced MRI scanner)
equipped MRI scanner of 1.5 T magnetic field intensity. The total number of
channels required to capture the image is around 48. So, the images obtained
using these scanner can also be called multi – channelled MR brain images.
3.5 IMAGE QUALITY PARAMETERS
MSE (Mean Square Error)
Mean Square Error defines the process of squaring the differentiated

quantities [143, 161]. It is expressed as the average of the squares of the errors
obtained by subtracting the input and the output values. MSE is the
cumulative squared error value between the input image A( ) and the
segmented image B( ).
( ) ∑ ∑ , ( ) ( )- (3.1)
‗m‘ & ‗n‘ denote the number of rows and columns present in an input image.
i = m, to perform incremental operation for ‗rows‘ in a ‗for‘ loop statement.
j = n, to perform incremental operation for ‗columns‘ in a ‗for‘ loop

statement.
51
MSE value of the resulting segmented image should be as low as

possible. Reduced MSE values offer minimized error occurrence in the
segmented MR brain images.
PSNR (Peak Signal to Noise Ratio)
Peak Signal to Noise Ratio refers to the immunity of an image to

noise signals [46, 143]. If the PSNR value is high, then the impact of noise
interference signal upon the MR brain image is quite low. The highest pixel
value of the input MR brain image is represented as (Say 255). PSNR
is expressed with the help of MSE values. The algorithm producing PSNR
values ranging between 40 to 100 dB is considered to be less sensitive to
noise signals. Segmented results obtained with high PSNR values can be used
in image compression procedures for image storage and image transmission
through the internet facilities. PSNR depicts resistance to the noise signals
and also briefs the quality of the segmented image. PSNR is usually expressed
as:
. / . / ( )
√
(3.2)
Memory Requirement
The memory space required by the combinational algorithm to

perform the segmentation process is described in terms of memory
requirement. The memory required for an image segmentation process should
be minimal. Lower memory requirement proves the efficacy of the
segmentation algorithm and is expressed in bytes.
52
Computational Time
The time required for the accomplishment of segmentation process is

denoted in terms of seconds. Computational time plays a major role in clinical
patient diagnosis.
Jaccard (Tanimoto Co - efficient) Index
Jaccard index is denoted as the ratio of the common voxels present

in the input image (A) and the segmented output image (B) to the union
function or the collection of voxels present in the input image (A) and the
segmented output image (B) [8, 17]. In other words, it is the ratio between the
intersection and the union functions of the voxel values present in the input
image (A) and the segmented resulting image (B). Explanation of TC value is
given in Equation (3.3).
( )
( ) (3.3)
( )
Dice Overlap Index (DOI)
DOI value is expressed with the aid of Jaccard index J (A, B). DOI
defines the overlapping function of the input image (A) and the segmented
output image (B) [8, 17]. Calculation of DOI is mentioned in Equation (3.4).
( )
( ) (3.4)
( )
Similarity Index (SI)
Similarity Index describes the similar or identical values between the

input image (A) and the segmented output image (B) [92, 115]. It relates to the
similarity found between the input image and the final segmented image
53
which comprises detected tumor region along with the identification of tissue
regions and is shown in Equation (3.5). The values used to calculate SI are:
1. True Positive (TP): Exact classification of tissue region and

identification of tumor region.
2. False Positive (FP): Misidentification or misclassification of

normal tissue region as tumor region.
3. False Negative (FN): Tumor region undetected.
(3.5)
Overlap Fraction (OF) or Sensitivity
Overlap Fraction (OF) or sensitivity value refers to the proper

segmentation or classification of the input image [92, 115]. Moreover, OF
defines the success rate in exact identification of tumor region and other tissue
regions. It is stated in Equation (3.6).
(3.6)
Extra Fraction (EF)
Extra Fraction (EF) mentions the number of voxels falsely detected

as tumor region. In addition, the misclassification rate of tissue regions is too
taken into account [115]. The algorithm which is capable of producing lower
EF value offers better segmentation results. EF is described in Equation (3.7).
(3.7)
54
CHAPTER 4
SEGMENTATION OF MR BRAIN IMAGES USING FUZZY
4.1 INTRODUCTION
Image segmentation procedure adopted to analyze medical images is

quite a challenging task. It refers to the process of partitioning a digital image
into multiple regions. The goal of segmentation is to change the
representation of an image into something that is more meaningful and easier
to analyze [23]. Image segmentation is used to locate objects and boundaries
in images. The result of image segmentation occurs as a set of regions that
collectively covers the entire image.
Over the years, various image segmentation techniques have been

proposed in order to analyze the images in an effective way. Brain tumor can
be identified visually, but it is difficult to determine the location and volume
of the tumor bounded by brain tissues [159]. Using Fuzzy Inference System
technique, the tumor present in the brain can be uniquely identified. It is
found that the tumor region in brain image has been perfectly segmented with
minimum rate of dissimilarity and error. The fuzzy rules are tested and
verified through simulation for the segmentation of MR brain images. Several
authors have suggested various methodologies and algorithms for image
segmentation. Ozkan et al [102] suggested neural network based image
segmentation for multimodal medical images. Chaplot et al [30] focused on
the classification of MR brain images using wavelet as input supported by
55
neural network. Fletcher et al [41] discussed the automatic techniques used to

segment non - enhancing brain tumors in MR Images.
Neural network technique based on brain abnormality detection in

MR Images using GABOR wavelets was discussed by Lashkari et al [72].
Murgavalli et al [91] suggested a high speed parallel fuzzy c means algorithm
for brain tumor segmentation. Gayatri Mirajkar and Barbadekar [85] have
proposed a multiresolution approach using undecimated wavelet transform
which allows the low-low (LL), low-high (LH), high-low (HL), and high high
(HH) sub-bands to remain at full size. Vijay Kumar [156] has recommended
computer aided diagnosis system to detect malignant texture in biological
study with the help of several techniques. The tumor mass detection and
cluster micro classification is used as the processing method for cancer
prediction. A Neuro fuzzy approach is used for the recognition of the
extracted region in a given MR image. The implementation is observed on
various types of MR images with different types of cancer regions. Tamilarasi
and Palanisamy [140] have proposed an algorithm for the compression of
medical images using fuzzy C – means based Contourlet transform. Raja
Kumar et al [112] recommended a methodology to enhance the contrast level
and segment MR images using fuzzy. The main aim of this work is to identify
the tumor part present in the brain with an efficient computer supported fuzzy
algorithm designed by using image processing methodologies. Since, fuzzy is
the basic of all soft computing methodologies; this research work is initiated
with FIS technique. Our proposed algorithm gives higher PSNR values in
comparison with the existing algorithms.
4.1.1 Types of Fuzzy Controller
Fuzzy inference is the process of formulating the mapping from a

given input to an output using fuzzy logic [21]. The process of fuzzy
inference involves developing membership functions, fuzzy logic operators
56
and if-then rules. Mamdani-type and Sugeno-type are the two types of Fuzzy
Inference Systems that can be implemented in Fuzzy Logic. In this case,
framing the fuzzy rules and implementing them for image segmentation are
done by Mamdani type controller exclusively. The Fuzzy Logic provides a
number of interactive tools that allow accessing many of the functions
through a graphical user interface (GUI). Here we use Fuzzy Inference
System for image segmentation on MR brain images. The hierarchy of image
segmentation process performed using FIS technique is shown in Figure 4.1.
MR brain image given as

input
Framing fuzzy rules using If-Then

functions
MR brain image segmented and tumor

part uniquely identified using FUZZY
INFERENCE SYSTEM
Performance Evaluation
Figure 4.1 Image segmentation procedure adopted by FIS (Fuzzy

Inference System)
4.1.2 Basics of Membership Function
The proposed Fuzzy Inference System (FIS) uses Triangular

Membership functions for setting out the threshold parameters for tumor
extraction and tissue segmentation. Amid the presence of several membership
parameters such as trapezoidal, sigmoidal, Gaussian, etc., triangular
membership function is predominantly used and it helps in making an exact
excavation of tumor region bound between the normal brain tissues.
57
Triangular membership functions assigned for input and output variables to

segment the MR brain images are illustrated in the figures below.
Figure 4.2 Triangular membership parameters chosen for the input

variables
Figure 4.3 Triangular membership parameters chosen for the output

variables
4.2 METHODOLOGIES
The methodology has been designed with membership functions and

other descriptions using fuzzy logic. The sequential steps involved in
designing the methodology using fuzzy logic are given below:
58
Step 1 : The input value and the output value of the fuzzy system are
assigned first and the range for the values for both input and output
is chosen between 0 and 256.
Step 2 : The input and the output values of Fuzzy Inference System are
mentioned with name such as input1, input2 & input3 and output1,
output2 & output3 for mf1, mf2 and mf3 respectively.
Step 3 : The type of mf1, mf2 and mf3 value is assigned with trimf
(triangular membership functions).
Step 4 : The membership parameter values are chosen for the input values
and output values and it is briefed in Figure 4.2 and 4.3.
Step 5 : The antecedent and the consequent part of the fuzzy logic along with
the weight and connection part used in framing the fuzzy rules are
assigned.
Step 6 : Based on fuzzy rules the clustering of the image is done and it is
executed with three clusters namely clust1, clust2 and clust3. The
values of the cluster are finally stored in the variable values AA1,
AA2 and AA3 respectively.
Step 7 : Using a FOR loop command, iterative clustering is done over the
image and stored in AA1, AA2 and AA3.
Step 8 : Finally, the segmented resulting images are evaluated using

comparison parameters.
59
4.2.1 Fuzzy Rule Implementation
An input MR brain image for segmentation is given and processed

using the image processing functions. Clinical images and Harvard Brain
Web Repository images of different dimensions, as previously mentioned in
Chapter 3, are processed by FIS algorithm. Further with the processed image,
fuzzy rules are framed and executed [34]. The fuzzy rules are framed as,
FuzzySys.rule (1).antecedent= [1];
This is particularly developed for the antecedent part. For e.g., the
basic fuzzy rule can be stated as if x is 1 and y is 2 then z is true. In this basic
fuzzy rule ‗x is 1‘ is said to be the antecedent and ‗y is 2‘ is the consequent.
The values 1 & 2 are said to be assigned fuzzy values. The coding mentioned
above represents fuzzy rule assigned for the antecedent part [51].
FuzzySys.rule (1).consequent= [1 0 0];
This coding is developed for the consequent part with the fuzzy rule
procedure [78, 56].
4.3 IMPLEMENTATION
Coding is developed to ensure the rate of similarity between the

segmented images by using Peak Signal to Noise Ratio, Mean Square Error
values, Jaccard Index, Dice Overlap Index, Similarity Index, Overlap Fraction
60
and Extra Fraction. The values prove that the MR Images segmented have a
diminished rate of similarity between them. The amount of segmentation of
the MR image samples ranges between 80 – 85% with a misclassification of
about 10 – 15% for 12 input MR brain slice image samples.
The sequential steps followed to derive the segmentation results are

briefed below:
Step 1 : The values for i & j are assigned with rows and columns of the
image, resolved into matrix form. The matrix converted image is
instantiated as AA.
Step 2 : Swapping function is done towards AA1 and assigned to the value
named as ‗aa‘.
Step 3 : The initial value of ‗aa‘ for the image AA1 is assigned to be greater
than 20. (This iteration process takes place till the value of ‗aa‘
overcomes 20 till 255 pixel values of the image are analyzed). AA1
denotes Grey Matter (GM) segmentation.
image, resolved into matrix form (for denoting AA1).
named as ‗aa‘.
overcomes 90 till 255 pixel values of the image are analyzed). AA2
defines White Matter (WM) segmentation.
61
image, resolved into matrix form (for denoting AA2).
named as ‗aa‘.
overcomes 150 till 255 pixel of the image is completely analyzed).
AA3 represents tumor detection.
4.4 RESULTS AND DISCUSSION
This work mainly focuses on the identification of brain tumor in an

efficient manner with lesser rate of similarity by using Fuzzy Inference
System. A surgeon can have a clear knowledge about the tumor and its
location in the brain. The value of Mean Square Error and Peak Signal to
Noise Ratio obtained proves that there is minimum dissimilarity between the
segmented images [152, 108]. Moreover, to process the 38 input samples, an
Intel(R) core(TM) 2 Duo CPU with 2.27 GHZ processor and a memory RAM
of 2GB provide a computational time of 39.28 seconds in average. The PSNR
value of existing methods, such as Modified Embedded Zerotree Wavelet
(Modified - EZW) [140] and conventional fuzzy [112] based segmentation
methodology ranges 30.36840 to 32.08. Our proposed methodology gives
PSNR values ranging 26.3131 – 33.4221. Higher the value of PSNR, lesser
will be the interference of noise signals.
62
Figure 4.4 Segmented tumor region of patient 1 (age: 5) affected with

Primitive Neuro Ectodermal tumor
Figure 4.4 illustrates the segmented results of the patient suffering

from Primitive Neuro Ectodermal Tumor (PNET). T2 FLAIR Axial with
Contract Enhancement (CE), T2 Coronary and T1 Coronary with Contract
Enhancement (CE) are shown in Figure 4.4. Minimal amount GM and WM
segmentations can be viewed. Tumor region present in T2 FLAIR Axial with
CE is detected. The tumor regions in other two images are left undetected and
they can be categorized under False Negative (FN).
63
White Matter Grey Matter Tumor

Input Image
Segmentation Segmentation Identification

Image type : MPR (Multi Planar Reconstruction) 3D contrast enhanced

images
Figure 4.5 explains the segmentation results of the patient suffering

from PNET. The images are of Multiplanar Reconstruction (MPR) type. The
tumor region present in the second image of Figure 4.5 can be viewed. Partial
segmentation of GM and WM regions is done by the proposed FIS technique.
64

Input Image
Meningioma
Figure 4.6 denotes the segmentation results of the patient affected

with Meningioma. Improper identification of tumor region is delivered by FIS
technique. The tumor region has been detected only in the third image of
Figure 4.6. Perfectionism in GM and WM segmentation could not be
observed too. The images of Figure 4.6 are Flair Axial, T1 Sagittal CE and T1
Coronary CE.
65
Input Image White Matter Grey Matter Tumor

high grade Astrocytoma
Figure 4.7 indicates the segmentation results of the patient suffering

from high grade Astrocytoma. Astrocytoma is caused by the astrocytes
present in the brain cells. Tumor region present in all the three images of
Figure 4.7 have been successfully identified. The images present in Figure 4.7
are T2 Axial, T1 Sagittal CE and T1 Coronary CE types. WM segmentation
of T2 Axial image has been done. Minimal segmentation of GM and WM
regions in other two MR sequences has been exhibited.
66

Input Image
high grade astrocytoma
Image type: MPR (Multi Planar Reconstruction) 3D contrast enhanced image
Figure 4.8 specifies the segmentation results of the patient suffering

high grade astrocytoma and the images are of 3D MPR type. Tumor region has
been clearly detected. Least segmentation of GM and WM regions is performed.
Input Image White Matter Grey Matter Tumor

Figure 4.9 Segmented tumor region of patient 4 (age: 3) affected by

supratentorial Primitive Neuro Ectodermal Tumor
67
Figure 4.9 shows the segmentation results of the patient suffering

supratentorial PNET. Partial identification of tumor region and proper
segmentation of WM region is made available. The types of MR brain image
sequences of Figure 4.9 are T2 Axial, T1 Sagittal CE and T1 Coronal types.
Exact identification of GM region in T2 Axial is done and a clear level WM
segmentation is produced by the proposed FIS technique.

Input Image

Image type: MPR (Multi Planar Reconstruction) 3D images
Figure 4.10 shows the segmentation results of the patient suffering

supratentorial PNET. The images of Figure 4.10 are 3D MPR type. A clear
identification of tumor region is not available. Exact WM segmentation of
both the images is done. Improper GM segmentation is obtained.
68

Input Image
Figure 4.11 Segmented tumor region of the patient suffering from Nerve
Sheath Tumors
Figure 4.11 expresses the segmentation results of the patient affected

with Nerve Sheath Tumors. Agreeable segmentation of GM region of all four
T1 Axial CE images is done. Tumor identification of all the images is
69
performed, with an exception of third image. WM segmentation offered FIS

technique is not up to the expectation.

Input Image
Figure 4.12 Segmented tumor region of the patient suffering from low
grade Glioma
Figure 4.12 explains the segmentation results of the patient affected

with low grade Glioma. Tumor detection of all the three images in Figure
4.12 is done. Proper segmentation of GM region is also exhibited. Only the
segmentation levels of WM region are improper. The images of Figure 4.12
are of T1 Axial with CE.
70

Input Image
Figure 4.13 Segmented tumor region of the patient suffering from high
grade Glioma
Figure 4.13 describes the segmentation results of the patient affected

with high grade Glioma. Better segmentation of GM region is done.
Disintegration has occurred in the detected tumor region of the first three
71
images present in Figure 4.13. Tumor region of last two images of Figure 4.13
is left undetected. Efficient WM segmentation is also unavailable. These
factors denote the shortcomings of the proposed FIS algorithm.

Input Image
Figure 4.14 Segmented tumor region of the patient suffering from

Pituitary Adenomas
72
Figure 4.14 describes the segmentation results of the patient affected

with Pituitary Adenomas. Normally, Pituitary Adenoma occurs in frontal lobe
of the brain. A clear identification of the tumor region is unattainable. Good
GM segmentation of all the images in Figure 4.14 is available. Inefficacy of
FIS technique in terms of WM segmentation is visualized.

Input Image
Figure 4.15 Segmented tumor region of the patients suffering from Low
Grade Glioma and Metastatic Bronchogenic Carcinoma
Figure 4.15 shows the segmentation results of the patient affected

with Glioma and Metastatic Bronchogenic Carcinoma. Good identification of
tumor region of all the three images is done. WM identification in the first
73
image of Figure 4.15 is shown. WM and GM segmentations of all the three

images are not up to the mark. The first image (T2 – W Axial with Contrast
Enhancement) of Figure 4.15 shows the tumor region of the patient affected
with low grade Glioma. The next two images, likely to be T2 - W Axial and
T1 – W Axial with contrast enhancement depict the patient suffering from
metastatic bronchogenic carcinoma.
4.4.1 Performance Evaluation
Table 4.1 Comparison parameters utilized to evaluate the efficiency of

the proposed Fuzzy Inference System (FIS)
PSNR in E. TIME MEMORY REQ

Images MSE TC DOI
dB in seconds in bytes
T2 FLAIR Axial with CE 23.4067 34.4374 0.3645 0.5343 7.842212 5.0023e+013
T2 Coronary 21.0517 34.8979 0.3645 0.5343 5.447456 4.8855e+013
T1 Coronary with CE 38.4376 32.2832 0.2728 0.4286 5.142342 4.9474e+013
MPR with CE 39.0473 32.2149 0.2885 0.4478 5.246051 4.9474e+013
MPR with CE 35.0568 32.6831 0.2381 0.3846 5.408834 4.9474e+013
Flair Axial 25.7366 34.0253 0.3645 0.5343 5.382239 4.8855e+013
T1 Sagittal CE 24.8983 34.1691 0.3757 0.5462 5.163085 4.8855e+013
T1 Coronary CE 25.4741 34.0698 0.3873 0.5583 5.512368 4.8855e+013
T2 Axial 24.2471 34.2842 0.4117 0.5833 5.566776 4.8855e+013
T1 Sagittal CE 42.1056 31.8874 0.2728 0.4286 5.234851 4.8924e+013
T1 Coronary CE 40.4889 32.0574 0.2653 0.4194 5.494187 4.8855e+013
MPR with CE 39.2535 32.1920 0.2581 0.4103 5.433121 4.8855e+013
T2 Axial 36.0098 32.5666 0.2653 0.4194 5.423707 4.8855e+013
T1 Sagittal CE 34.7324 32.7235 0.2445 0.3929 5.179654 4.8855e+013
T1 Coronal 37.0336 32.4448 0.2581 0.4103 5.083428 4.9474e+013
T1 MPR 32.6779 32.9883 0.2805 0.4381 4.848224 4.8924e+013
T1 MPR 30.3269 33.3125 0.3144 0.4784 5.280535 4.8924e+013
T1 Axial with CE 27.4550 33.7446 0.3144 0.4784 5.220793 4.8924e+013
T1 Axial with CE 41.4439 31.9562 0.4263 0.5977 59.826547 5.1323e+014
T1 Axial with CE 54.8033 30.7427 0.4263 0.5977 58.009256 5.0083e+014
T1 Axial with CE 56.2454 30.6299 0.4263 0.5977 57.991498 4.8841e+014
T1 Axial with CE 45.2065 31.5788 0.4263 0.5977 58.439681 5.1549e+014
T1 Axial with CE 19.7784 35.1689 0.4232 0.5988 56.411125 4.9104e+014
T1 Axial with CE 22.8846 34.5354 0.4886 0.6564 83.504557 6.8949e+014
T1 Axial with CE 70.1977 29.6676 0.4274 0.5988 81.961691 6.6313e+014
T1 Axial with CE 66.6764 29.8911 0.4274 0.5988 81.381565 7.1144e+014
T1 Axial with CE 43.7310 31.7229 0.4312 0.6025 83.756634 6.7631e+014
T1 Axial with CE 65.0852 29.9960 0.4312 0.6025 80.653480 7.1144e+014
74
T1 Axial with CE 55.3804 30.6972 0.4312 0.6025 83.529461 6.6753e+014

T1 Axial with CE 54.4160 30.7735 0.4312 0.6025 86.838945 7.0266e+014
T1 Axial with CE 46.0636 31.4972 0.4312 0.6025 84.796626 6.7631e+014
T1 Axial with CE 60.0001 30.3493 0.4274 0.5988 84.043186 7.1144e+014
T1 Axial with CE 56.2588 30.6289 0.4274 0.5988 82.821115 6.7631e+014
T1 Axial with CE 37.4805 32.3928 0.4468 0.6176 86.644710 7.0266e+014
T1 Axial with CE 46.7102 31.4367 0.4312 0.6025 84.736285 6.7631e+014
T2 Axial with CE 27.4550 33.7446 0.3144 0.4784 85.309352 4.5625e+013
T2 Axial with CE 84.3289 28.8710 0.2258 0.3684 5.374565 5.0023e+013
T1 Axial with CE 71.6604 29.5780 0.2653 0.4194 6.336158 4.8924e+013
MSE – Mean Square Error, PSNR – Peak Signal to Noise Ratio, TC – Tanimoto Coefficient, DOI –
Dice Overlap Index
Table 4.1 represents the comparison parameters used for evaluating

the efficiency of the proposed FIS technique.
Table 4.2 Comparison of SI, OF and EF values
Algorithm Similarity Index (SI) Overlap Fraction (OF) Extra Fraction (EF)
Proposed FIS Algorithm 0.7119 0.6364 0.1515
Table 4.2 briefs the segmentation accuracy of the proposed

methodology in terms of Similarity Index (SI), Overlap Fraction (OF) and
Extra Fraction (EF). For a good segmentation algorithm, SI and OF should be
near to ‗1‘ and EF should be close to ‗0‘.
Memory Requirement
8.00E+14
7.00E+14
6.00E+14
5.00E+14
Values in bytes
4.00E+14
3.00E+14 MEMORY REQ
2.00E+14
1.00E+14
0.00E+00
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37
Input Images
Figure 4.16 Comparison of memory required attained by executing the

proposed Fuzzy Inference System (FIS)
75
Results from FIS algorithm

100
90
80
70
Comparison values
60 MSE
50 PSNR
40 TC
30 DOI
20 E. TIME
10
0
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37
Input images
Figure 4.17 Comparison parameters for evaluating the proposed Fuzzy

Inference System (FIS)
FIS requires a computational time of 39.28 seconds and Memory

Requirement of 3.11E+14 bytes. MSE of 42.19, PSNR of 32.18 dB, Jaccard
Index of 36.07% and DOI of 52.54% are average comparison parameter
values produced by the recommended FIS algorithm.
4.6 SUMMARY OF CONTRIBUTIONS
The membership parameters are assigned manually and frequent

tuning of threshold value is not required. FIS is used as an initial level
methodology to segment MR brain images. Acceptable results were obtained
with the initialized parameters, with no further variations. FIS is only capable
of segmenting T2 – W and T2 – W contrast Enhanced MR brain images. The
proposed methodology is incapable of segmenting other MR brain image
modalities. FIS could be used in offline clinical diagnosis to a minimal scope.
76
CHAPTER 5
SEGMENTATION OF MR BRAIN IMAGES USING SELF

ORGANIZING MAP AND FUZZY K - MEANS
5.1 INTRODUCTION
Malignant and benign types of tumor infiltrated in human brain are

diagnosed with the help of an MRI scanner. With the slice images obtained
using an MRI scanner, certain image processing techniques are improvised to
have a clear anatomy of brain tissues. One such image processing technique is
hybrid Self Organizing Map (SOM) with Fuzzy K Means (FKM) algorithm,
which offers successful identification of tumor and good segmentation of
tissue regions present in an MR brain image. The algorithm proposed through
this work has better data handling capacities and it also performs efficient
processing upon the input Magnetic Resonance (MR) brain images.
Automatic detection of tumor region in MR (Magnetic Resonance) brain
images has a high impact in helping the radio surgeons to assess the size of
the tumor present inside the tissues of brain; it also supports in identifying the
exact topographical location of tumor region. The proposed hybrid SOM –
FKM algorithm assists the radio surgeon by providing an automated tissue
segmentation and tumor identification, thus enhancing radio therapeutic
procedures.
Conventional analysis of MR brain images preferred by a medical

expert is quite a complicated and time consuming task. The accuracy of
diagnosis entirely vests upon the experience possessed by the expert. Manual
77
procedures involving tumor identification and tissue segmentation are prone

to error and inaccuracy, where noises or interferences produced by the
magnetic field and radio frequency signals occupy a predominant position.
So, it is inevitable to have an automatic segmentation technique to offer
considerable results.
Some of the MR (Magnetic Resonance) brain image sequences are

classified as T1, T2, FLAIR (Fluid Attenuated Inversion Recovery) and MRS
(Magnetic Resonance Spectroscopy) images. T1 weighted images depict
white matter as white in color (inner tissue region) and grey matter (outer
tissue region) as grey in color. Whereas, it is white color for grey matter and
grey color for white matter in case of T2 weighted images. FLAIR images
support the radiologist in viewing the tissues of the brain in a better
perspective by suppressing the fluid content (Water and Cerebrospinal fluid)
in brain. T1, T2 and FLAIR image sequences are segmented by the proposed
SOM – FKM algorithm.
Generalization property exhibited by the support vector classifiers

[153] has favored its utilization in the domain of image segmentation.
Functioning of these classifiers is governed by statistical learning theory
[103].
Some other segmentation techniques are developed using ANN

classifiers [5, 40, 47, 97, 120, 138, 169] such as self-organizing maps (SOM)
[5, 47, 59, 97, 138]. In presence of several image segmentation procedurals,
no one technique is available to segment heterogeneous tumor region present
in different sequences of MR brain image. With this objective, a novel
technique combining the working principles of SOM and FKM algorithms is
proposed through this work. This technique almost satisfies the needs of a
radio surgeon and is up to the mark.
78
The main aim of our methodology is to offer an automated hybrid

algorithm to identify the tumor region present in the brain images. An
automated algorithm greatly assists the radio surgeon in the process of non –
invasive diagnosis of the patients affected with tumor. Further, image
sequences of a MRI scanner with all the three axes namely, coronal, sagittal
and axial are segmented with the support of the proposed algorithm. Levels of
tumor infiltration are clearly analyzed with the help of the proposed
methodology. A radio surgeon can efficiently plan for pre and post radio
therapeutic procedures using the hybrid SOM – FKM algorithm.
Comparison parameters validate the performance of the algorithm

and the results are compared with Fuzzy C Means (FCM) algorithm. Since,
FCM and FKM are the two standard competitive clustering algorithms, we
choose FCM as an algorithm to compare with FKM combined with SOM
strategy.
5.2 METHODOLOGIES
A simple flowchart explaining the sequential steps of the proposed

method is given in Figure 5 for easy understanding. Initial level pre-
processing is done to remove the skull region from the given input images.
The images have been obtained using a Siemens – Aera, Magnetom Avanto -
Tim technologies equipped MRI scanner of 1.5T magnetic field intensity. The
total number of channels required to capture the image is around 48. The
total time duration for acquiring the images is found to be less than 10 min.
The slice thickness of the MR brain images obtained from four patients
suffering from different types of lesions is found to be 1mm×1mm×1mm.
Removal of skull region is inhibited using Brain Extraction Tool (BET).
Segmentation of tissue region and identification of tumor region are the main
concerns and the above - said factor has compelled the researcher to use BET
[25] and ROI (Region of Interest) based brain masking for initial level image
79
processing. Images from Harvard Brain Repository have also been used to
validate the proficiency of the proposed algorithm.
Input
MR brain image
as
Image pre-
processing
Skull stripped Membership function

image updation using Cj
Dimensionality reduction and Prototype grouping

prototype preparation using done using FKM
SOM
Centroid allocation (Cj) in

SOM prototype: Objective function
or Maximum
iteration reached?
Distance No
Evaluation (𝒅𝒊𝒋 )
Yes
Segmented tissue
regions
Tumor region
Identified
Figure 5.1 Flowchart of hybrid SOM - FKM algorithm
5.2.1 SOM Algorithm
The first step of processing a SOM algorithm commences with

initialization, where voxel intensities are initially assessed and modelled with
the assistance of SOM prototypes. This process is executed using histogram
80
analysis, which acts as a replacement for mixtures of probability density

functions. Histogram based model helps to analyze the peaks and valleys of
the input image and further supports in sustaining unique informative values
of voxel, which can be greatly used in segmentation process [70, 71, 99]. The
SOM algorithm is explained as follows:
Let X Rd, where Rd represents the diversified input data and X is

the input vector. The winning unit is calculated in each iterative step as,
( ) *‖ ( ) ( )‖+ (5.1)
x X, where ( ) is defined as the input vector at time instant ‗t‘ and ( ) is

the sample vector related to the unit ‗i‘. The unit quite closer to the input
vector ( ) is denoted as winning unit and the associated prototype or the
sample vector is sequentially updated. The adaptive learning process in SOM
is completed by updating the prototypes of the units in the neighbourhood of
the winning unit. This procedure is described as,
( ) ( ) ( ) ( )( ( ) ( )) (5.2)
Here, ( ) is the exponential decay learning factor and ( ) is the

neighbourhood function related to the unit ‗i‘. Both learning factor and the
neighbourhood function decrease with respect to time and the adaptability of
prototypes becomes slower as the neighborhood of the unit ‗i‘ contains few
units.
( (‖ ‖ ⁄ ( ) ))
( ) (5.3)
Equation 5.3 explains the neighbourhood function, where denotes

the position on the output space and ‖ ‖ is the distance between the
winning unit and the unit ‗i‘ on the output space. The neighborhood function
81
is explicated using a Gaussian function, which consistently shrinks on

completion of every iteration step and is briefed in Equation 5.4. The winning
unit found during this process is called the Best Matching Unit (BMU).
Consecutively, ( ) controls the reduction of the Gaussian neighbourhood in
each iteration step using the time constant .
( ⁄ )
( ) (5.4)
The prototype present in SOM acts as the centroid of a Voronoi

polyhedron. SOM is capable of projecting the prototypes either in two
dimensional or three dimensional spaces. This representation occurs based on
the dimension of the output layer. Initial level grouping of the prototypes is
done using SOM. Prototypes which are quite similar in nature are placed
closer to the output space. This placement of prototypes is usually performed
based on Euclidean distance calculation. The location of the prototypes on the
output space tends to be a valuable source of information and can be utilized
to cluster the SOM [99, 121]. Based on the eigenvalues and eigenvectors of
the training data, linear initialization of the SOM prototypes is performed.
The prototypes are sorted in the form of hexagonal lattices. This process is
performed based on the width of the lattice and its proportionality towards the
standard deviation of the first principal component [4, 99, 71]. This indicates
the arrangement of the prototypes, as the first dimension of the prototypes is
aligned proportionally to the first principal component and the second
dimension of the prototypes is aligned proportionally to the second principal
component. On complete estimation of the optimal number of map units, the
unsupervised neural network is nearly trained. The above - mentioned
calculations are executed during the initialization of SOM map. On
completion of the map initialization process, feature vectors obtained from the
normalized dimensionality confined feature space developed for zero mean
and unity variance is used to train the unsupervised neural network. This
82
procedure can also be called data whitening. Normalization of the feature

vectors used for training the map supports in reducing the influence of one
dimensional value, which occurs due to the extracted features of variable
property. As an outcome of the adopted training procedurals, a group of
prototypes assembled on the SOM layer supports in modeling the features of
the volume image histogram.
A unique classification process is performed based on assigning the

closest prototype for each voxel. Exact depiction of the data is accomplished
with the aid of the prototypes obtained during the training process, which
indicates the quality of the SOM technique. The effectiveness of the trained
map is ascertained using the following conditions (i) generalization of the
input data using the prototypes and (ii) distance evaluation between the
similar prototypes present in output space or in other words it is the
denotation of topology preservation.
Quantization error ( ) and the topological error ( ) are the two

important quality assessment parameters for the trained map. Quantization
error ( ) is related to the measurement of average distance between the
individual data vector and the Best Matching Unit (BMU). Topological error
( ) defines the measurement of data vectors proportionality, for which the
first and second Best Matching Unit (BMUS) are non – adjacent.
Quantization error and the topological error are described in Equations (5.5)
and (5.6).
∑ (⃗ ) (5.5)
∑ ‖⃗ ⃗⃗⃗⃗⃗⃗ ‖ (5.6)
83
In Equation (5.5), denotes the total data vectors; ( ⃗ ) = 1 if non –

adjacency is preserved by first and second BMU for ( ⃗ ), or else ( ⃗ ) = 0. In
Equation (5.6), ⃗⃗⃗⃗⃗⃗ represents weight function related to the BMU of data
vector ⃗ , where ⃗ is the ith data vector present in the input space and is
capable of defining the quantization error. Hence, lower quantization error
( ) and topological error ( ) affirm better data representation and
topological preservation and therefore provide better clustering result.
Betterment of SOM algorithm is coined by lower quantization error and
topological error [7, 13, 99].
Here, SOM itself is an unsupervised neural network based clustering

technique, which can be directly used in segmentation process [79, 99].
5.2.2 FKM Algorithm
Fuzzy K means algorithm, also called FKM, acts a stiff competitor

to FCM algorithm. As mentioned by Chih-Tang Chang et al [31], FKM
partitions the given input image into ‗k‘ clusters (partitioning the voxels
present in the image).
∑ ∑ (5.7)
represents the number of data points or voxels present in the input

MR brain image. refers to the number of clusters formed by FKM. i & j
explain the rows and columns of the input image. defines the membership
functions of FKM with fuzziness coefficient m (fuzziness co – efficient is
obtained from the overlapping of clusters). is the squared Euclidean
distance between pixel and cluster representative (Cluster center or
centroid value). FKM algorithm has quicker convergence rate and requires
less computational time for processing an image. It has limitations in handling
84
the data of wider range. Taking favour of quicker convergence rate, FKM is
used in the proposed methodology and the problem of data range handling is
overcome using SOM.
5.3 IMPLEMENTATION
The combination of SOM and FKM is the main novelty proposed in

this research work. SOM is used for initial clustering and it offers
dimensionality reduction [99]. The image given as input to the algorithm is of
varied dimensions, which is reduced with the aid of SOM algorithm and
displayed as a two dimensional map. The size of the map is 8 × 8.
Prototype , defined in [70] acts as the resultant value obtained using SOM.
The classified prototypes obtained using SOM is given as the inputs to FKM
algorithm. Thus the problem aroused due to FKM in handling wide and large
data range is minimized with the inclusion of SOM. Next, FKM utilizes the
output from SOM and performs the clustering operation as specified in [31].
Prototype is the input fed to FKM algorithm and the following process
takes place sequentially:
∑ ∑ (5.8)
Here, describes the total number of iterations. represents the

membership function and its calculation is described in Equation (5.9).
represents the Euclidean distance calculation. (Number of clusters is 4).
,( ) ∑ ( ) - (5.9)
Here, ‖ ‖ ‗m‘ denotes the fuzziness co – efficient

obtained due to the overlapping of clusters. ‗m‘ is entered as 2.0 as it gives
out optimal segmentation results and is derived from Karan Sikka et al [65].
85
‘k’ denotes the number of clusters into which the prototype is to be

partitioned. l = 1 represents the first cluster. ‗ ‘ is the location of the data
th
at the position and it belongs to the prototype . This specific task has
been motivated by Ortiz et al [9, 98].
• High membership function is assigned to the data point which

lies very near to the cluster centre.
• In Equation (5.10), Cj denotes the cluster center. Cj can also be

called cluster representative.
∑
∑
(5.10)
• Equation (5.11) explicates that the sum function of

membership values of each data point and should be equal to
one.
∑ = 1, for i = 1 to N (5.11)
Function , explains the final segmentation result attained from

FKM algorithm. As a whole, SOM has greatly supported FKM algorithm by
offering an initial level segmentation.
The efficiency of the proposed methodology is explained using

comparison parameters and is briefed in results and discussions.
Gold standard or the manually segmented images [65] help in

assessing the proficiency of the proposed methodology. These images are
obtained by the manual intervention offered by an expert radiologist, who has
gained deep knowledge in diagnosis as an outcome of prolonged experience.
86
Sample gold standard images are shown in Figure 5.2, 5.3 and 5.4. These
images serve as an assessment tool for analyzing Figure 5.5, 5.12 (A) and
5.14 (A). Due to the availability of gold standard images for both clinical and
Harvard datasets, a cross comparison between SOM based FKM and FCM
algorithm is performed.
Gold Standard Images of patient - 1 (age: 5) suffering

from Primitive Neuro Ectodermal tumor
Input image Gold Standard
Figure 5.2 Gold Standard Images of the tumor detected region for patient – 1
87
Gold Standard Image for Harvard database

Figure 5.3 Gold Standard Image of tumor detected region for the data
available in Harvard Brain Web Repository
Gold Standard Image for Harvard database

Figure 5.4 Gold Standard Image of tumor detected region for the
patient suffering from Nerve Sheath Tumors
Figure 5.5 & 5.6 are obtained from the patient of age 5, suffering
from PNET (Primitive Neuro Ectodermal tumor). Perfect tumor identification
along with tissue segmentation is performed. Ipsilateral ventricular system is
extremely compressed in the case of patient 1. T2 FLAIR Axial with Contrast
Enhancement - Figure 5.5 (A), T2 Coronary - Figure 5.5 (B) and T1 Coronary
with Contrast Enhancement – Figure 5.5 (C) are obtained during the clinical
88
diagnosis of PNET. Minimal intensity levels of Grey Matter (GM) and White
Matter (WM) structures are observed in input images of Figure 5.5 (B) and
5.5 (C), where successive segmentation of both GM and WM is presented by
the proposed SOM based FKM algorithm, which can be referred to the result
part. Further, a clear cut demarcation between the tumor region and the edema
portion is visualized in Figure 5.5 (A), which proves the segmentation
efficiency and robustness of the proposed algorithm.
SOM – FKM
Input Image Skull Stripped image
Resulting Image
5.5
(A)
5.5
(B)
5.5
(C)
Figure 5.5 Segmented results of patient - 1 (age: 5) suffering from

Primitive Neuro Ectodermal tumor obtained from clinical
database
89
SOM – FKM
Resulting Image
5.6(A)
5.6 (B)

Primitive Neuro Ectodermal tumor obtained from clinical
database, which are of 3D Multi Planar Reconstruction
(MPR) type images
Figure 5.6 is a 3D Multi Planar Reconstruction (MPR) Contrast

Enhanced image of patient 1. Contrast enhancement was favoured by
Gadolinium or Gado IV; it is a common contrasting agent which supports
efficient diagnosis of the lesions and pathologies. MPR images help in
analyzing the anatomy of brain. A clear structure of GM and WM is
visualized in the result part of Figure 5.6 (A). Even though severe intensity
variation is observed in the input image of 5.6 (B), the proposed algorithm has
produced a clear segmentation of GM and WM with successful detection of
tumor region and can be seen in the result part of Figure 5.6 (B). The efficacy
of the proposed algorithm is literally proved in Figure 5.6.
Figure 5.7 is the complete imagery attained from the patient of age
35 suffering from meningioma. Calcified meningioma with dural tail is
reported in the right parietal convexity of the brain. In Figure 5.7 (A),
90
distinguished view upon the grooves (Gyri and Sulci) is made available and
can be observed in the result part. Moreover, no clear examination of GM and
WM can be done with the input images of Figure 5.7 (B) and Figure 5.7 (C).
This difficulty is overcome with the help of the suggested hybrid SOM based
FKM algorithm and could be verified in the result part. These challenging
cases of tumor identification and tissue segmentation have been resolved with
the aid of the recommended hybrid algorithm. Flair Axial – Figure 5.7 (A), T1
Sagittal - Figure 5.7 (B) and T1 Coronary with Contrast Enhancement -
Figure 5.7 (C) was used in clinical practices to diagnose the tumor region
present in the meninges of brain.
SOM – FKM
Resulting Image
5.7
(A)
5.7
(B)
5.7
(C)

Meningioma obtained from clinical database
91
After the injection of Gadolinium (or) Gado IV contrast agent and on

a specific attention paid towards the right ganglio capsular region and
thalamus, a heterogeneous enhancement was identified. Presence of intense
edema has led to the effacement of ipsilateral ventricular system observed
during clinical diagnosis. The high grade Astrocytoma tumor region bounded
between the dense edema regions is exactly detected and it is clearly
visualized in Figure 5.8.
SOM – FKM
Resulting Image
5.8
(A)
5.8 (B)
5.8 (C)

high grade Astrocytoma obtained from clinical database
92
T2 Axial – Figure 5.8 (A), T1 Sagittal – Figure 5.8 (B) and T1

Coronary with Contrast Enhancement – Figure 5.8 (C) was used to diagnose
the tumor region. The above figure shows the effectiveness of the proposed
methodology in resolving challenging clinical datasets, where computer aided
diagnosis is extensively preferred by the radio surgeon. Demarcation between
tumor region and edema portions is made available in the result of Figure 5.8
(A); improper segmentation of tissue regions is also observable. This reduces
the count of true negative value. Perfect tumor identification and tissue
segmentation can be seen in the result part of 5.8 (B) and 5.8 (C). A 3D MPR
with Contrast Enhancement was also used to identify the impact of high grade
Astrocytoma upon the patient. The proposed hybrid algorithm succeeded in
segmenting the tissue regions and has delivered perfection in tumor
identification and the above said details are clearly stated in Figure 5.9. In
Figure 5.9, it is also observed that the continuous intensity variations present
in the input image does not have any effect upon the working of SOM based
FKM technique.
SOM – FKM Resulting

Image
Figure 5.9 Segmented result of patient - 3 (age: 32) suffering from high
grade Astrocytoma obtained from clinical database, which is
of 3D Multi Planar Reconstruction (MPR) type image
93
Figure 5.10 & Figure 5.11 were completely acquired from Patient 4
affected with supratentorial PNET and subjected to radio therapeutic
procedures. Partial identification of tumor region is observed in 5.10 (A).
SOM – FKM
Resulting Image
5.10
(A)
5.10
(B)
5.10
(C)

supratentorial Primitive Neuro Ectodermal Tumor obtained
from clinical database
Exact identification of tumor and segmentation of tissue regions are

characterized by the proposed algorithm and they are verified in the result part
of Figure 5.10 (B) and Figure 5.10 (C). On injection of Gado IV contrast
agent, a mixed enhancement was observed in the areas of right parieto-
94
occipital lobe structures, cortical and subcortical regions. Effacement of Sulci

and right lateral ventricle were visualized due to the presence of cerebral
edema in larger volumes. Tumor region impinged in the dense edema portions
was successfully identified and can be verified in Figure 5.10 (B) and Figure
5.10 (C). A prominent partition between the tumor and the edema filled
regions done by the proposed hybrid algorithm is illustrated in 5.10 (C). T2
Axial – Figure 5.10 (A), T1 Sagittal with Contrast Enhancement – Figure 5.10
(B) and T1 Coronal – Figure 5.10 (C) are image sequences used to diagnose
the pathologies.
SOM – FKM
Resulting Image
5.11
(A)
5.11
(B)

supratentorial Primitive Neuro Ectodermal Tumor obtained
from clinical database, which are of 3D Multi Planar
Reconstruction (MPR) type images
A 3D MPR sequence was used to examine the impact of tumor

region and edema portion upon Gyri and Sulci structures. Demarcation
between tumor and edema regions was exactly done with the help of SOM
95
based FKM algorithm and it is illustrated in Figure 5.11 (A) and Figure 5.11
(B). Fine tumor identification made by the proposed methodology is shown in
Figure 5.11 (B).
Figure 5.12 is obtained from the patients affected with low grade
Glioma and metastatic bronchogenic carcinoma. T2 Axial with Contrast
Enhancement - Figure 5.12 (A), is used to diagnose the malignant tumor
region present in the left occipital lobe of the brain affected by low grade
Glioma. T2 Axial - Figure 5.12 (B) and T1 Axial with Contrast Enhancement
- Figure 5.12 (C) are the images used to diagnose the patient suffering from
Metastatic Bronchogenic Carcinoma.
Skull Stripped SOM – FKM

Input Image
image Resulting Image
5.12(A)
5.12 (B)
5.12 (C)
Figure 5.12 Segmented results of the patients suffering from Glioma and
Metastatic Bronchogenic Carcinoma, obtained from
Harvard Brain Web Repository
96
The suggested SOM based FKM algorithm effectively segments the

tumor region and is proven in the result part of Figure 5.12 (A), (B) and (C).
Distinguished view upon the tumor region, which is surrounded by dense
edema region is made available by the proposed algorithm and is shown in
Figure 5.12 (A) and (B). Excellent tissue segmentation is also offered by the
proposed methodology and can be visualized in Figure 5.12.
Figure 5.13, 5.14, 5.15 & 5.16 are Contrast Enhanced images and
were attained from Harvard brain web database. These images were used to
affirm the performance levels of the proposed SOM based FKM algorithm.
Skull Stripped SOM – FKM

Input Image
image Resulting Image
5.13(A)
5.13 (B)
5.13 (C)
5.13(D)
Figure 5.13 Segmented results of the patients suffering Pituitary

Adenomas
97
Figure 5.13 was obtained from the patient affected with Pituitary
Adenomas. T1 axial images with Contrast Enhancement offered by Gado – IV
were used to diagnose the patient.
The shortcoming of the proposed methodology in misidentifying the

tumor region is clearly visible in Figure 5.13(A), 5.13 (B), 5.13 (C) and 5.13
(D). These images also show the perfection of the proposed methodology in
segmenting the tissues present in the input MR brain images. Tumor region in
the above said images could not be detected and can be considered under the
count of false negative.
Figure 5.14 is acquired from the patient affected with nerve sheath
tumors. The inefficiency of proposed algorithm related to misidentification of
tumor region is shown in Figure 5.14 (B). Tissue regions present in the input
images have been exactly segmented and can be verified in the result part of
Figure 5.14.
SOM – FKM
Resulting Image
5.14
(A)
5.14
(B)
5.14
(C)
Figure 5.14 Segmented results of the patients suffering from Nerve

Sheath Tumors
98
Figure 5.15 states the imagery of the patient suffering from low
grade Glioma. Although the images suffer from slight noise distortions caused
by the radio frequency signals produced by a MRI scanner, the proposed
algorithm identifies the tumor region and segments the tissue region accurately.
The proficiency of the proposed algorithm can be confirmed by viewing the
result part of Figure 5.15 (A), 5.15 (B), 5.15 (C), 5.15 (D) and 5.15 (E).
SOM – FKM
Resulting Image
5.15
(A)
5.15
(B)
5.15
(C)
5.15
(D)
5.15
(E)
Figure 5.15 Segmented results of the patients suffering from low grade
Glioma
99
Variable dimensions of the tumor region present in the patient

suffering from high grade Glioma are perfectly identified with the support of
the proposed SOM based FKM algorithm. In spite of the presence of intensity
variations in the tissues of brain very near to the tumor region, the proposed
algorithm successfully identifies and segments the tumor region and tissue
regions respectively. The helpfulness of the proposed algorithm can be
understood by looking into the segmentation results of Figure 5.16.
SOM – FKM
Resulting Image
5.16
(A)
5.16
(B)
5.16
(C)
5.16
(D)
100
5.16
(E)
5.16
(F)
Figure 5.16 Segmented results of the patients suffering from high grade
Glioma
Table 5.1 Performance evaluation for the proposed SOM based FKM
Algorithm
Computational Memory
Images MSE PSNR TC DOI
time Requirement
T2 Coronary 3.4964 42.6946 0.3142 0.4781 2.166391 3.1064e+013
MPR with CE 2.8268 43.6178 0.2268 0.3698 5.128464 3.6144e+013
MPR with CE 2.6551 43.8900 0.2052 0.3406 2.452931 3.1393e+013
Flair Axial 2.2346 44.6388 0.2115 0.3491 3.729248 2.4774e+013
T1 Sagittal CE 0.9899 48.1751 0.3335 0.5002 1.967794 3.7894e+013
T1 Coronary CE 2.2226 44.6622 0.2859 0.4447 2.125825 2.6166e+013
T2 Axial 2.4823 44.1823 0.2130 0.3512 3.066892 4.3516e+013
T1 Sagittal CE 1.8648 45.4244 0.3198 0.4846 2.403510 3.2310e+013
T1 Coronary CE 3.4029 42.8123 0.4152 0.5867 2.402933 2.8402e+013
MPR with CE 2.5523 44.0615 0.2228 0.3644 3.685131 3.7398e+013
T2 Axial 1.9620 5.2038 0.2182 0.3582 2.131511 4.1478e+013
101
T1 Sagittal CE 1.9004 45.3423 0.2957 0.4564 1.924268 3.4124e+013

T1 Coronal 1.4354 46.5611 0.3359 0.5029 1.944054 4.0958e+013
T1 MPR 1.7301 45.7502 0.2770 0.4338 1.465142 3.6330e+013
T1 MPR 5.7986 40.4976 0.1982 0.3308 4.866730 3.1151e+013
T1 Axial with CE 0.9830 48.2055 0.3759 0.5464 2.279997 2.3616e+013
T1 Axial with CE 1.8620 45.4309 0.3564 0.5255 3.245901 3.1409e+013
T1 Axial with CE 1.1874 47.3847 0.3178 0.4823 2.400932 3.2659e+013
T1 Axial with CE 2.1070 44.8941 0.3093 0.4724 1.572801 3.0744e+013
T1 Axial with CE 2.1285 44.8501 0.3341 0.5009 2.965126 3.3951e+013
T1 Axial with CE 0.8359 48.9094 0.3981 0.5695 2.779671 3.3601e+013
T1 Axial with CE 3.3151 42.9259 0.3894 0.5606 6.453621 3.6221e+013
T1 Axial with CE 1.6880 45.8572 0.3939 0.5652 1.684540 3.3825e+013
T1 Axial with CE 2.3832 44.3591 0.4273 0.5988 2.651933 3.7745e+013
T1 Axial with CE 2.3011 44.5114 0.3521 0.5208 1.708257 3.2618e+013
T1 Axial with CE 2.8465 43.5877 0.3478 0.5161 3.968782 3.6706e+013
T1 Axial with CE 1.9070 45.3274 0.4024 0.5739 2.839513 3.6674e+013
T1 Axial with CE 0.4625 1.4796 0.3851 0.5560 1.619065 3.6615e+013
T1 Axial with CE 2.7641 43.7153 0.3674 0.5373 4.361223 3.4649e+013
T1 Axial with CE 0.9855 48.1940 0.3981 0.5695 2.177429 3.3149e+013
T1 Axial with CE 0.8573 48.7995 0.4504 0.6211 1.715861 2.6559e+013
T1 Axial with CE 0.9316 8.4385 0.4189 0.5904 1.806791 2.8988e+013
T1 Axial with CE 1.4246 46.5939 0.4189 0.5904 3.309900 3.1353e+013
T2 Axial with CE 2.3401 44.4384 0.1968 0.3289 2.597943 3.6327e+013
T2 Axial with CE 2.9245 43.4703 0.1807 0.3060 2.393811 2.6458e+013
T1 Axial with CE 2.5364 44.0887 0.1760 0.2993 1.686739 2.0459e+013
Dice Overlap Index
Table 5.1 defines the calculated comparison parameter values used

for validating the proposed SOM based FKM algorithm.
102
Comparison of MSE values

1.6
1.4
1.2
MSE values
1
0.8
0.6 MSE
0.4
0.2
0
SOM - FKM FCM
Algorithms
Figure 5.17 Comparison of Mean Square Error values obtained from

SOM – FKM and FCM algorithms
While speaking about the size of the input images, the clinical
datasets are of dimensions 480 × 375 for axial plane, 1105 × 650 for Coronal
plane and 763 × 664 for sagittal plane. The axial planed images obtained from
Harvard brain repository is found to be of dimension 256 × 256. Either an
image compression algorithm or a segmentation algorithm should produce
PSNR value ranging between 40 to 60 dB or even more. In the case of
proposed segmentation methodology, 47.20 is the average PSNR value and is
lesser than the PSNR value produced by conventional FCM algorithm. Since,
PSNR and MSE values are inversely proportional, FCM algorithm has
exceeded SOM based FKM in these terms. But, SOM based FKM has
significance in producing better visualization of tumor and tissue regions.
MSE and PSNR values have minimal impact upon the efficiency of the
proposed methodology and they are the factors of least consideration. Jaccard
Tanimoto Coeffecient Index (TC), usually expressed as percentage (%),
gives out information regarding the similarity of voxels present in the input
image and the segmented output image. The proposed SOM based FKM
algorithm excels FCM by providing 30.29% as TC value, whereas, FCM
produces a TC value of 21.11 %. Dice Overlap Index (DOI), which is quite
103
related to TC gives out the significance of uniformity of the voxels present in

input and output images. The suggested SOM – FKM algorithm is capable of
producing an average DOI value of 45.4 % and FCM algorithm produces
34.85% as average DOI value. SOM based FKM has the ability to produce
better DOI values in comparison with FCM algorithm. All the above - said
comparisons are clearly illustrated in Figure 5.18.
Comparison of PSNR, TC and DOI values

70
60
50
Values
40
30 SOM - FKM
20 FCM
10
0
PSNR TC in % DOI in %
Comparison parameters
Figure 5.18 Comparison of Peak Signal to Noise Ratio (PSNR),

Tanimoto Coefficient (TC) and Dice Overlap Index (DOI)
values obtained from SOM – FKM and FCM algorithms
To say about memory requirement, both the soft computing

methodologies were run on a 32 Bit operating system with Pentium 4
processor (2.27 GHz) and 2 GB RAM. As said in section 5.3, lower the
memory space higher is the segmentation efficiency of the algorithm. The
proposed algorithm has outperformed FCM in terms of memory requirement
as it provides an average value of 3.11E+14 bytes. The inclusion of SOM,
which acts as an initial level clustering, has impacted FKM algorithm to
consume lesser memory space for its operation. FCM algorithm requires
4.48E+14 bytes for processing 38 input images. This is the advantage of the
recommended SOM based FKM algorithm and it is pictured in Figure 5.19.
104
Comparison of Memory Requirement

5.00E+14
4.50E+14
4.00E+14
3.50E+14
3.00E+14
Bytes
2.50E+14
2.00E+14 Memory Requirement
1.50E+14
1.00E+14
5.00E+13
0.00E+00
SOM - FKM FCM
Algorithms
Figure 5.19 Comparison of Memory Requirement or Memory

Utilization values obtained by processing SOM – FKM and
FCM algorithms
The average computational time for processing 38 input MR brain

images by the recommended SOM based FKM algorithm is found to be less
than 16.7 seconds. On the other hand, FCM algorithm requires average time
duration of 24.77 seconds for processing the input images. These explanations
are clearly mentioned in Figure 5.20.
Comparison of computational time

30
25
Time in seconds
20
15
10 Elapsed Time
0
SOM - FKM FCM
Algorithms
Figure 5.20 Comparison of Computational Time or Elapsed Time values

obtained from SOM – FKM and FCM algorithms
105
Table 5.2 defines the efficacy of the proposed SOM – FKM

algorithm using SI, OF, EF and Sensitivity values. The proposed technique
outperforms the FCM algorithm by providing better SI, OF and EF values. EF
value is sizably reduced using SOM based FKM algorithm. For a good image
segmentation algorithm, SI and OF value should be close to 1 and EF value
should be very nearer to 0 [115]. Afore mentioned values listed in table 5.2
make a confirmation that the proposed hybrid SOM – FKM algorithm is
superior to FCM algorithm.
Algorithm Similarity Criteria Overlap Fraction Extra Fraction

FCM 0.7273 0.7500 0.3125
SOM - FKM 0.9189 0.8718 0.0256
Sensitivity values
1
0.9
0.8
0.7
0.6
Values
0.5
Sensitivity
0.4
0.3
0.2
0.1
0
Leemput et al (2001)
Boudraa et al (2000)
Proposed SOM - FKM

Hassan Khotanlou et al
Zijdenbos et al (2002)
Johnston et al (1996)
Rasoul Khayat et al (2008)
(2011)
Methodologies
Figure 5.21 Comparison of Sensitivity (or) Overlap Fraction values with

other competitive algorithms
106
Figure 5.21 has been completely derived from the comparison of

sensitivity values produced by the proposed algorithm and other algorithms
suggested by Khayati et al [115], Johnston et al [60], Boudraa et al [24],
Leemput et al [73], Zijdenbos et al [178] and Hassan Khotanlou et al [50].
The proposed SOM based FKM algorithm is far superior to other rival
segmentation techniques as it produces 0.8718 as sensitivity.
Comparison of FIS and SOM - FKM

60
50
40
values
30
SOM - FKM
20
FIS
10
0
MSE PSNR in TC DOI E. TIME in
dB seconds
Figure 5.22 Average Values of Comparison Parameters for FIS and

SOM – FKM Algorithms
From Figure 5.22, it is clearly identifiable that SOM – FKM

algorithm excels FIS algorithm in terms of MSE, PSNR and time duration.
Minimum deformation of segmented results (MSE) with PSNR value above
40 decibels in average time duration of 16.7 seconds is produced by
SOM - FKM. Though reduction of TC and DOI values occur in comparison
with FIS, the quality of segmentation and tumor identification is not
compromised in the case of SOM – FKM algorithm. SOM based FKM
produces 30.29% and 45.4% as the average TC and DOI. Furthermore, Figure
5.23 stands as a testimonial to exclaim the efficiency of SOM – FKM
algorithm in terms of memory requirement.
107
Memory requirement in bytes

3.35E+14
3.30E+14
3.25E+14
3.20E+14
Bytes
3.15E+14 MEMORY REQ in

bytes
3.10E+14
3.05E+14
3.00E+14
FIS SOM - FKM
Algorithms
Figure 5.23 Average Values of Memory Requirement for FIS and

SOM – FKM Algorithms
Table 5.3 Comparison of SI, OF and EF values between FIS and

SOM – FKM algorithms
Similarity Index Overlap Fraction Extra Fraction

Algorithm
(SI) (OF) (EF)
FIS 0.7119 0.6364 0.1515
SOM - FKM 0.9189 0.8718 0.0256
Table 5.3 gives out a clear cut idea about the best operability and
efficiency of the SOM based FKM algorithm in comparison with FIS
algorithm. The proposed SOM – FKM technique offers better SI, OF and EF
values than the FIS technique. The derived values were derived with the
assistance of an expert radio surgeon (manual segmented images or gold
standard images).
108
A novel hybrid SOM based FKM algorithm has been proposed

through in this chapter and the recommended hybrid SOM – FKM algorithm
excels FIS algorithm in the process of segmenting MR brain images. The
effectiveness of the proposed algorithm is demonstrated using Tables 5.1, 5.2
& Figure 5.21. Varied clinical datasets along with standard database images
have been used to prove the efficacy of the proposed algorithm. A novel
automated technique which requires minimal manual intervention to segment
the tissues of brain, especially GM, WM and CSF regions, along with the
identification of different types of tumor at different locations has been
proposed. The ultimate goal of this research work is to assist the clinician or
radio surgeon in diagnosing the patients within reduced time duration and is
greatly achieved through the combination of SOM and FKM techniques.
Improving the values of DOI, TC and PSNR has to be done. Time duration
for processing the images and MSE values can also be further reduced.
Moreover, the shortcomings experienced in FIS technique have been rectified
to a considerable level using SOM – FKM technique. The proposed algorithm
can be extensively used in pre and post radio surgical applications.
109
CHAPTER 6
SEGMENTATION OF MR BRAIN IMAGES USING PARTICLE

SWARM OPTIMIZATION BASED FUZZY C - MEANS
6.1 INTRODUCTION
Image processing plays a very important role in analyzing the

images of different standards. Image enhancement, Image restoration and
Image compression are the types of image processing techniques of which
image segmentation process under Image enhancement technique is utilized
to analyse the MR (Magnetic Resonance) brain images. Complex tissue
region in the image can be uniquely identified and any pathology or tumor -
infiltrated region in the tissues of the brain can also be brought under
observation [11].
Tissues in brain are the most complicated parts of our body, and a
clear examination and study are therefore required by a radiologist to identify
the pathologies. Normal Magnetic Resonance (MR) scanner is capable of
producing brain images with bounded tissues, where unique and segregated
views of the tissues are required. A distinguished view upon the images is
manually impossible and can be subjected to operator - errors.
Magnetic Resonance imaging technique has a great importance in

the domain of diagnosing the organs of human body. Hydrogen atom, the
basic constituent of our human body is stimulated using resonant radio
frequency signals. The excited hydrogen atom is capable of emitting the
absorbed radio frequency signals, upon which several signal processing
110
operations are done to obtain a clear anatomy of human organs and tissues.
Brain tissues are bounded together and are complex to analyse. Moreover,
edema and tumor - infiltrated region in the tissues are difficult both to identify
and to analyse.
With the assistance of a soft computing technique, an automated

unique segmentation upon the brain tissues along with the identification of the
tumor region can be effectively done. These functionalities assist the
radiologist extensively. Several soft computing techniques have been
proposed and one such technique which is being proposed is PSO based FCM
algorithm.
The proposed hybrid PSO - FCM algorithm segments the tissues and
identifies the edema and tumor affected regions in the brain and assists the
radio surgeon in computer aided surgeries [27]. The existence of one such
technique is unavailable and with a motive to improve the convergence rate
and efficiency of FCM algorithm, PSO based FCM is introduced.
T1 -, T2 – weighted, MPR and FLAIR type brain images are

obtained with the help of an MRI scanner. T1 weighted images help in
observing and analysing the anatomy of the brain and T2 weighted images
assist in identifying the pathologies of brain. FLAIR type images support in
viewing the tissues of brain by suppressing the fluid contents. Clinical
imagery data of four patients and images from Harvard Brainweb database are
utilized to observe the efficacy of the algorithm in segmentation process.
Tumor - cut: segmentation of brain tumors on contrast enhanced MR

images for radio surgery applications proposed by Andac et al [8] is capable
of segmenting only contrast enhanced T1 weighted images. The usage of the
tumor cut algorithm is limited towards T1 contrast enhanced images and not
upon other modalities of brain images. Tumor detection is done using the
111
basic principle of Region of Interest (ROI) and requires manual assistance for
segmentation procedures.
An improved MRI brain image segmentation to detect cerebrospinal

fluid level using Anisotropic Diffused Fuzzy C - Means proposed by Tamije
Selvy et al [139], segments T2 weighted images. Inclusion of an optimization
technique to clustering process would enhance segmentation accuracy is the
conclusive statement made through this research.
Spatial information based image clustering with a swarm approach

suggested by Ouadfel Salima et al [101], segments MR images using
Artificial Bee Colony (ABC) algorithm combined with FCM clustering
technique, where time consumption has to be reduced. The usage of the
segmentation algorithm has been confined to process T1 weighted images.
Segmentation of MR brain images for tumor extraction using fuzzy

illustrated by Govindaraj Vishnuvarthanan and Murugan Pallikonda
Rajasekeran [45] is capable of segmenting T1 weighted axial type images.
Values of PSNR and MSE can further be improved. The proposed PSO based
FCM technique is capable of segmenting T1, T2, MPR and FLAIR type
images in all the three axes of MR brain imaging.
6.2 METHODOLOGIES
Methodology
The following flowchart explains the sequential procedure involved

in the proposed PSO – FCM algorithm.
112
Input MR Brain
image
Pre-processing steps
(Image resizing, skull
stripping and RGB to
grayscale conversion)
Finding the best position

of the pixels using PSO
Clustering using FCM

based on the optimized
results from PSO
Segmented
output image
Figure 6.1 Flowchart explaining the sequential steps of the proposed

PSO based FCM algorithm
Step 1 : MR brain images of different pixel size are fed as input to the
proposed methodology.
Step 2 : MR brain images with different dimensions and sizes are used as
input, say 256 × 256, 480 × 375, 1105 × 650 and 763 × 664. To
standardize the segmentation procedure, pre – processing steps
involving image resizing (Conversion to 512 × 512 pixel size),
RGB (Red Green Blue) to grayscale conversion, skull stripping and
patient detail removal are done. Adjustment of image intensity or
color map values is performed to handle medical information losses
aroused due to image resizing.
113
Step 3 : To identify the best or the optimum cluster value using PSO
algorithm.
Step 4 : The best known position of the pixel along with the clusters found
using PSO, acts as a leader or the centroid value for FCM
algorithm.
Step 5 : FCM algorithm is associated with grouping of pixels in a cluster

with the centroid value. The clusters are grouped around the
centroid value found using PSO algorithm.
Step 6 : The process of grouping the pixels of each cluster with the centroid
value takes place for nearly K number of iterations,
Step 7 : Output image with the tissues of brain segmented and the tumor
region identified is obtained from the proposed algorithm.
6.2.1 PSO Algorithm
Particle Swarm Optimization (PSO) performs a searching operation

upon a population of particles which is entirely associated with the pixels of
the MR brain images. The particles together constituting the population are
randomly generated. The individual particle represents a vivid solution and it
is denoted by a position vector. The particles are free to move in a problem
space. The position of the particle in the search space is described as the
position best ( best) of the particle and further, the position of the particle
with its best known position in comparison with the neighbouring particles is
described as global best ( best). The velocity of the particle moving through
the search space is defined using a velocity vector. The updation of gbest,
best and velocity vector of the particle is done in every individual step [42].
Obtaining the best positions of the particle through PSO (Particle Swarm
114
Optimization) reduces the procedure of finding the total number of clusters

and centroid of the pixels using FCM (Fuzzy C – Mean) algorithm.
⃗⃗⃗ ( ) ⃗⃗⃗ ( ) (⃗⃗⃗ ( ) ⃗⃗⃗ ( )) .⃗⃗⃗⃗ ( ) ⃗⃗⃗ ( )/ (6.1)
⃗⃗⃗ - Velocity of the particles which lies within (Range expressed in

terms of Pixels)
- Positive constants.
& - Uniform weight values ranging between 0 and 1.
⃗⃗⃗ ( ) - Best position of the particle associated with fitness function ( best).
⃗⃗⃗⃗ ( ) - Global best position of the particle ( best).
⃗⃗⃗ ( ) - Position vector of the particle.
- Time instant.
⃗⃗⃗ ( ) = ⃗⃗⃗ ( ) ⃗⃗⃗ ( ) (6.2)
The above equation represents the updated velocity ⃗⃗⃗ ( ) [116]

and the corresponding position vector updation of the particle as ⃗⃗⃗ ( ).
Normally, based upon the calculation of fitness function, intensity value of
the image and time duration, PSO based image segmentation can be
implemented [110].
6.2.2 FCM Algorithm
Fuzzy C - means (FCM) algorithm is a clustering technique

developed by Dunn, improved by Bezdek and further improved by Matteo
Matteucci; in this technique the pixels (data) of the Magnetic Resonance
(MR) brain images are classified and grouped under ‗n‘ number of clusters
specified by PSO algorithm. ⃗⃗⃗ ( ), the updated position vector of the
115
particle (voxel or pixel) derived with the help of PSO algorithm acts as the
centroid value and based upon this value, the number of clusters for FCM
algorithm are also assigned using PSO. Using the position vector, centre
pixel of each cluster is identified and assigned with a high membership grade
value through FCM. The neighbouring pixels of least mean distance from the
centroid pixel are assigned with low membership grade value and are grown
around the centroid value, hierarchically. The membership grade and the
cluster centres are iteratively updated to reduce the objective function of
grouping the pixels.
∑ ∑ ‖ ‖ (6.3)
N - Number of data points given as an input to the algorithm

(Total number of pixels)
K - Number of iterations to be performed.
C - Number of clusters into which the pixels are to be grouped.

‗C‘ is calculated using the position vector obtained from PSO
algorithm.
- Centre vector for the clusters (Centroid value of the pixels).
- Data points (pixels).
- Degree of membership for the ith data point of in cluster j.
‖ ‖ - Measuring the similarity or mean distance of the neighbouring

pixels present in the data point to the centre vector (centroid
value) of cluster j.
116
Degree of membership is given as,
(6.4)
‖ ‖
∑ ( )
‖ ‖
m - Fuzziness co – efficient obtained from the overlapping of clusters and it is

defined as 1.5 ≤ m ≤ 2.5 for optimum segmentation results [65].
Centre vector (Centroid voxel) for the cluster is defined as,
∑
∑
(6.5)
During the initial processing of the algorithm, .
- A random value initialized as, (Usually expressed as 0, 0.1,

0.2, 0.4….1), such that ∑ .
6.3 IMPLEMENTATION
The value, ⃗⃗⃗ ( ) the updated position vector of the particle

(Voxel or pixel) derived with the help of PSO algorithm acts as the centroid
value and based upon this value, the number of clusters for FCM algorithm is
also assigned using PSO. Using the position vector, center pixel of each
cluster is identified and assigned with a high membership grade value through
FCM.
∑ ∑ ‖ ‖ (6.6)
N - Number of data points given as an input to the algorithm

(Total number of pixels).
K - Number of iterations to be performed.
C - ⃗⃗⃗ ( ) - Number of clusters into which the pixels are to be grouped.

‗C‘ is calculated using the position vector obtained from
PSO algorithm. (Number of clusters is 4).
117
Thus the output from PSO algorithm is fed as an input to the FCM
algorithm to obtain better segmentation results and minimize the time
consumption to process the input MR brain images.
In Figure 6.2(A), edema covered tumor region has been perfectly

identified. The same is the case of Figure 6.2(B) and 6.2(C). GM and WM
region cannot be viewed in the input images of Figure 6.2. The segmented
results show a clear segmentation of WM and GM region. Segmentation of
GM and WM region in Figure 6.2(B) is a challenging one and it has been
perfectly addressed with the aid of PSO based FCM algorithm.
Input Image Skull Stripped Image Segmented Result
6.2
(A)
6.2
(B)
6.2
(C)
Figure 6.2 Segmented results derived using PSO – FCM for the patient -
1 (age: 5) suffering from Primitive Neuro Ectodermal tumor
118
Exact boundary line of the tumor region is shown in Figure 6.3(A)

and 6.3(B). Also, WM and GM region have been effectively segmented with
the support of PSO based FCM technique.
6.3 (A)
6.3 (B)
Figure 6.3 Segmented results derived using PSO – FCM for the patient -
1 (age: 5) suffering from Primitive Neuro Ectodermal
tumor, which are of 3D Multi Planar Reconstruction (MPR)
type images
Partial identification of tumor region is illustrated in Figure 6.4(A),

whereas, exact identification of tumor region can be seen in Figure 6.4(B) and
6.4(C). Moreover, the hard task of differentiating WM and GM regions has
been entirely accomplished by the proposed PSO – FCM algorithm and it can
be seen in the segmented results part of Figure 6.4.
119
6.4 (A)
6.4 (B)
6.4 (C)
Figure 6.4 Segmented results derived using PSO – FCM for the patient
- 2 (age: 35) suffering from Meningioma
High grade Astrocytoma situated within the dense edema region is

excavated using the proposed methodology and it can be seen in the
segmented result part of Figure 6.5. Good segmentation of WM and GM
regions is also performed. This shows the perfection of the PSO based FCM
algorithm.
120
6.5
(A)
6.5
(B)
6.5
(C)
- 3 (age: 32) suffering from high grade Astrocytoma
In spite of the presence of severe intensity variations, the proposed

PSO based FCM algorithm has the ability to identify aberrant region present
in the input image of Figure 6.6. Perfect GM and WM segmentations can also
be seen in the segmented results of Figure 6.6.
- 3 (age: 32) suffering from high grade Astrocytoma
121
Figure 6.7 has been completely acquired from the patient who
underwent clinical diagnosis after radio therapeutic procedures. Clear
identification of Supratentorial PNET and perfect segmentation of GM and
WM regions can be observed in the segmented result part of Figure 6.7.
6.7 (A)
6.7 (B)
6.7 (C)
- 4 (age: 3) suffering from supratentorial Primitive Neuro
Ectodermal Tumor (PNET)
PSO based FCM algorithm renders meticulous identification

of supratentorial - PNET region and effective segmentation of GM and
WM regions. It is proved in the segmented results of Figure 6.8 (A) and
Figure 6.8 (B).
122
6.8 (A)
6.8 (B)
- 4 (age: 3) suffering from supratentorial Primitive Neuro
Ectodermal Tumor obtained from clinical database, which
are of 3D Multi Planar Reconstruction (MPR) type images
The tumor region surrounded by dense edema portion of Figure

6.9(B) has been clearly detected. The tumor region and edema portion have
been effectively discriminated in both the cases of Figure 6.9(A) and 6.9(B).
Partial identification of tumor region is seen in Figure 6.9(C). Augmentation
in WM and GM segmentation can be observed in the segmented results of
Figure 6.9.
123
6.9 (A)
6.9 (B)
6.9 (C)
Figure 6.9 Segmented results derived using PSO – FCM for the
patients suffering from Glioma and Metastatic
Bronchogenic Carcinoma attained from Harvard Brain
Web Repository
Proper identification of Pituitary Adenomas is exhibited in Figure

6.10 (B) and 6.10 (D). But in the case of Figure 6.10 (A) and 6.10 (C),
subsequent leakage of tumor region into the segmented WM region is
observed. This makes a count for False Negative (FN). Partial segmentation
of WM and GM regions in almost all the segmented results of Figure 6.10 is
visualized. This particular shortcoming of the proposed PSO based FCM
requires improvement.
124
6.10
(A)
6.10
(B)
6.10
(C)
6.10
(D)
suffering Pituitary Adenomas
125
Nerve sheath tumors present in the input images of Figure 6.11 have
been exactly detected by the recommended PSO based FCM algorithm.
Optimal segmentation of WM and GM regions have also been done and it can
be verified in the segmented results of Figure 6.11.
6.11
(A)
6.11
(B)
6.11
(C)
suffering from Nerve Sheath Tumors
Effective tumor identification and better demarcation between WM

and GM regions for the input images of Figure 6.12(A) are made available
with the support offered by PSO based FCM algorithm. This can be verified
in the segmented results of Figure 6.12.
126
6.12
(A)
6.12
(B)
6.12
(C)
6.12
(D)
6.12
(E)
suffering from low grade Glioma
127
Figure 6.13 has been acquired from the patient suffering from high
grade Glioma. In clinical practice, it is difficult to examine the GM and WM
structures in Figures 6.13 (D), 6.13 (E) and 6.13 (F). But, with the assistance
of PSO based FCM, clear distinguishing between WM and GM is done. The
tumor region present in all the input images of Figure 6.13 has almost been
perfectly identified. This proves out the efficacy of the suggested PSO based
FCM algorithm.
6.13
(A)
6.13
(B)
6.13
(C)
6.13
(D)
128
6.13
(E)
6.13
(F)
high grade Glioma
Table 6.1 Evaluation of PSO – FCM using standard comparison

parameters
PSNR in E. TIME in MEMORY REQ in

Images MSE TC DOI
dB seconds bytes
T2 Coronary 0.0269 63.8389 0.4310 0.4310 28.3148 2.1920e+014
MPR with CE 0.0261 63.9570 0.2000 0.3333 28.1567 2.2011e+014
MPR with CE 0.0267 63.8668 0.2433 0.3913 28.0425 2.1800e+014
Flair Axial 0.0231 64.4944 0.2003 0.3338 25.4172 2.1789e+014
T1 Sagittal CE 0.0225 64.6174 0.2009 0.3346 28.0030 2.1789e+014
T1 Coronary CE 0.0208 64.9435 0.2006 0.3342 29.2214 2.2174e+014
T2 Axial 0.0201 65.0886 0.2006 0.3342 21.1131 2.2011e+014
T1 Sagittal CE 0.0232 64.4711 0.2003 0.3338 25.2651 2.0410e+014
T1 Coronary CE 0.0221 64.6918 0.2000 0.3333 29.1302 2.0410e+014
MPR with CE 0.0219 64.7276 0.2000 0.3333 27.7571 2.0245e+014

129
T2 Axial 0.0214 64.8174 0.2529 0.4037 27.0830 2.2078e+014
T1 Sagittal CE 0.0213 64.8369 0.2000 0.3333 28.9669 2.2062e+014
T1 Coronal 0.0190 65.3510 0.2000 0.3333 25.9118 2.2397e+014
T1 MPR 0.0192 65.2981 0.2000 0.3333 28.7703 2.0243e+014
T1 MPR 0.0182 65.5190 0.2000 0.3333 13.2510 2.0243e+014
T1 Axial with CE 0.0177 65.6492 0.2003 0.3338 28.9584 2.1016e+014
T1 Axial with CE 0.0077 69.2911 0.2019 0.3360 22.1458 1.8899e+014
T1 Axial with CE 0.0121 67.2913 0.2041 0.3390 27.1773 1.9165e+014
T1 Axial with CE 0.0115 67.5159 0.2038 0.3386 16.4155 1.8006e+014
T1 Axial with CE 0.0082 68.9848 0.2100 0.3471 20.3074 1.8393e+014
T1 Axial with CE 0.0233 64.4629 0.2104 0.3476 34.7525 1.8779e+014
T1 Axial with CE 0.0064 70.0579 0.2296 0.3734 30.3059 1.7785e+014
T1 Axial with CE 0.0115 67.5314 0.2009 0.3346 29.4348 1.7785e+014
T1 Axial with CE 0.0117 67.4391 0.2013 0.3351 30.1128 1.7398e+014
T1 Axial with CE 0.0110 67.7096 0.2064 0.3421 24.8852 1.7398e+014
T1 Axial with CE 0.0126 67.1316 0.2032 0.3377 29.7671 1.9165e+014
T1 Axial with CE 0.0107 67.8421 0.2032 0.3377 14.8882 1.7013e+014
T1 Axial with CE 0.0126 67.1139 0.2025 0.3368 20.7122 1.8559e+014
T1 Axial with CE 0.0107 67.8269 0.2067 0.3426 30.3216 1.8006e+014
T1 Axial with CE 0.0125 67.1599 0.2013 0.3351 28.8237 1.8831e+014
T1 Axial with CE 0.0119 67.3854 0.2019 0.3360 26.7234 1.8006e+014
T1 Axial with CE 0.0073 69.5141 0.2165 0.3559 26.3957 1.8779e+014
T1 Axial with CE 0.0082 69.0150 0.2064 0.3422 29.8243 1.8945e+014
T2 Axial with CE 0.0177 65.6552 0.2003 0.3338 22.9219 1.7279e+014
T2 Axial with CE 0.0280 63.6533 0.2288 0.3724 27.9590 1.7666e+014
T1 Axial with CE 0.0252 64.1147 0.2672 0.4217 30.1500 1.8439e+014

Dice Overlap Index
Table 6.1 gives an ideology about the efficiency of the proposed

PSO based FCM algorithm using standard evaluation parameters.
130
Comparison of FIS, SOM - FKM and PSO -

FCM
70
60
50 FIS
40
Values
SOM - FKM
30 PSO - FCM
20
10
0
MSE PSNR in dB TC DOI E. TIME in
seconds
Figure 6.14 Comparison of FIS, SOM based FKM and PSO based FCM
using evaluation parameters
Petronella Anbeek et al [106] processed 256 × 256 dimensional size

images in a minimum time duration of 10 minutes, using Pentium 4 processor
(1.7 GHz) and 512 Mb RAM, whereas, the proposed algorithm processes 512
× 512 dimensional images using Pentium 4 processor (2.27 GHz) and 2 GB
RAM with an average computational time less than 26.36 seconds. The same
was implemented on a Pentium 4 processor (1.7 GHz) and 512 Mb RAM as
suggested by Anbeek et al [106], where an average computational time of 3
min was required to process the proposed methodology. Reduced
computational time enhances the radio surgeon to analyze larger data
volumes. The time duration required by the proposed PSO based FCM is
higher than the SOM – FKM technique, since it utilizes both optimization
(PSO) and fuzzy – clustering (FCM) techniques.
The average MSE value offered by PSO based FCM is 0.017507895,

whereas, 1.370931579 and 42.19068684 are the average MSE values
produced by FIS and SOM – FKM techniques, respectively.
131
The average PSNR value generated by PSO – FCM is 66.06. SOM

based FKM and FIS algorithms produce 47.20 and 32.18 as the average
PSNR values. Average TC value produced by PSO based FCM is 21.41,
which is a little bit lower than the TC values produced by SOM – FKM and
FIS methodologies. As, TC and DOI values are directly proportional, 34.31 is
the average DOI value formed by PSO based FCM. This value is quite lower
when compared with the DOI values produced by FIS and SOM – FKM
techniques.
The average MSE and PSNR values produced by Govindaraj

Vishnuvarthanan and Murugan Pallikonda Rajasekeran [45] are 41.61 and
32.62. 0.017 and 66.06 are the average MSE and PSNR values produced by
the proposed PSO based FCM algorithm.
Meritxell et al [84] produced 37.19 as the mean MSE value. 0.025 is

the average MSE value obtained by Amir Ehsan Lashkar [7] through
segmenting the input MR brain images. Lower MSE value indicates minimum
deformation in the segmentation of brain tissues and tumor region. The
proposed algorithm satisfies this criterion by offering a mean MSE value of
0.0175.
Figure 6.14 and Figure 6.15 make a clear inscription that PSO based
FCM offers lower MSE and memory requirement values, with far better
PSNR values in comparison with FIS and SOM – FKM algorithms. Though
there is a slight reduction in TC and DOI values, PSO based FCM has the
ability to offer better tissue segmentation and tumor identification when
compared with FIS and SOM – FKM algorithms and it is evidentially
visualized in Figures 6.2, 6.3, 6.4, 6.5, 6.6, 6.7, 6.8, 6.9, 6.10, 6.11, 6.12 and
6.13.
132
MEMORY REQ in bytes

3.50E+14
3.00E+14
2.50E+14
2.00E+14
Bytes
1.50E+14
MEMORY
1.00E+14 REQ in bytes
5.00E+13
0.00E+00
FIS SOM - FKM PSO - FCM
Algorithms
Figure 6.15 Comparison of FIS, SOM based FKM and PSO based FCM
using memory requirement values
As per Figure 6.15, the average memory requirement value required

by the proposed PSO based FCM algorithm is found to be 1.97E+14, which is
much lower than the average memory consumption required by FIS and SOM
– FKM algorithms. In average, FIS requires 3.31E+14 bytes and SOM – FKM
requires 3.11E+14 bytes for processing thirty eight MR brain images.
Table 6.2 Comparison of FIS, SOM – FKM and PSO – FCM using SI,
oF and EF values

Algorithm
(SI) (OF) (EF)
FIS 0.7119 0.6364 0.1515
SOM -
0.9189 0.8718 0.0256
FKM
PSO -
0.9333 0.9211 0.0263
FCM
133
An algorithm producing similarity index value above 0.7 is

considered to be efficient [115]. As per table 6.2, the proposed PSO based
FCM algorithm outruns the rival techniques by producing a SI value of
0.9333. FIS and SOM based FKM produce 0.7119 and 0.9189 as Similarity
Index values.
Overlap Fraction of an algorithm should be near to 1 and Extra

Fraction should be close to 0 [115]. Lower EF values are derived with the aid
of the proposed methodology. PSO based FCM algorithm produces an
average EF value of 0.0263 for segmenting all the three brain tissues and
tumor region. When compared with the EF value produced by SOM based
FKM, slight increment in EF value is generated by PSO – FCM. But, PSO
based FCM favours optimal segmentation results in comparison with SOM –
FKM technique.
Approximated SI values produced by Rasoul Khayat et al [115] and

Petronella Anbeek et al [106] are 0.884733 and 0.759325. The proposed
algorithm performs remarkably by producing an approximated SI value of
0.9333.
Hassan Khotanlou and Mahlagha Afrasiabi [50] derived 0.82 as the

all patient SI value, whereas, the proposed methodology produces 0.9333 as
the all patient SI value.
Normally, FCM requires 100 to 200 iterations for processing the

input MR brain images. With the addition of PSO, the iteration level has been
considerably reduced to 70 to 100 iterations.
134
Traditional techniques described in related works are capable of

segmenting either T1 or T2 weighted and FLAIR type images. But, it is
clearly identified that the proposed PSO based FCM algorithm is capable of
segmenting all T1, T2, MPR and FLAIR type brain images with all sequential
axis such as coronal, sagittal and axial. Also, images with low radio frequency
signals are segmented using PSO – FCM, which proves the robustness of the
algorithm. Results and discussions provide evidence that the proposed
algorithm out performs the FIS and SOM based FKM algorithms. Future
work lies with improving Jaccard Index and DOI values, with reduction in
memory requirement for processing the PSO based FCM algorithm. PSO
plays a prominent role in offering cluster values to the FCM algorithm, based
on which the segmentation process is carried out. Thus a new approach to
segment MR brain images using PSO based FCM algorithm is proposed
through this research work. Further, the convergence rate of FCM algorithm
is reduced. Concurrently, segmentation accuracy of FCM algorithm is
prominently improved with the inclusion of PSO technique.
135
CHAPTER 7
SEGMENTATION OF MR BRAIN IMAGES USING PARTICLE

SWARM OPTIMIZATION AND FUZZY C - MEANS
IMPROVED BY REGION GROWING ALGORITHM
7.1 INTRODUCTION
MRI scanning is the most extensively used, since its functions use
non - ionized electromagnetic radiations and strong magnetic fields. Based on
the principle of Nuclear Magnetic Resonance (NMR), MRI scanner provides
a better augmentation in analyzing heart, lungs, pelvis, abdomen and soft
tissues in brain and offers a good resolute image by which we can have a clear
and distinguished view upon the tissues of brain. On the obtained MR brain
images, segmentation is done by the combination of PSO (Particle Swarm
Optimization), FCM (Fuzzy C – Means) and Region Growing Algorithm
(RGA) for unique tumor identification. Image segmentation is a reliable
method in image processing where pixels of image are segregated based upon
the boundaries. The pixels within the specified boundary limits (nominal
threshold) are grouped together. Image segmentation is a methodology among
image processing techniques and it is preferably opted for tumor detection
and identification.
Image segmentation helps in identifying the tumor part present in the

brain image [178]. The proposed Region Growing methodology helps in
effectively identifying the tumor part and eases the complication in
136
identifying the tumor infiltrated region done manually by a radio surgeon who
has acquaintance with radio surgery applications.
The proposed methodology segments Axial, Coronal and Sagittal

sequences of T1 - w, T2– w, FLAIR and 3D MPR type MR brain images. In
T1 weighted images, tissues comprising water and fluid appear to be dark and
the tissues containing fat appear white in color. It is vice – versa in the case of
T2 weighted images, where tissues filled with water and fluids appear to be
bright and tissues with fat appear to be dark in color. FLAIR images help in
reducing the vision of fluid content (Say Cerebro Spinal Fluid and Edema
portion) in the brain. The proposed methodology has the ability to segment T1
- w, T2 – w, FLAIR, Multiplanar Reconstruction (MPR) and Contrast
Enhanced (CE) images, where exact identification of tumor region is in need.
7.2 METHODOLOGIES
Novelty of the algorithm is exhibited by combining the concepts of

optimization (PSO) and Fuzzy – Clustering (FCM), with the enhancement of
segmentation results using PSO based FCM merged with Region growing
algorithm. It is also observed that the inclusion of PSO based FCM has geared
up the Region Growing algorithm by making it quite insensitive to threshold
initiation. Time Varying Acceleration Constants (TVAC) is utilized in PSO
(in this case PSO can be called as Modified PSO [110]), rather using normal
positive constants. Based on a wide literature survey, it is ostensible that the
combination of all these three algorithms is non - existing.
The algorithms so far developed are capable of segmenting T1 – w,

T2 – w, MPR and FLAIR type MR brain images individually. The proposed
PSO based FCM merged with Region Growing algorithm is capable of
segmenting all four types of images, with the inclusion of segmenting contrast
enhanced and MPR based brain images.
137
Start
Input MR Brain images
Optimization using PSO
Centroid allocation and cluster

identification by PSO
Clustering done by FCM based on

PSO results
Seed point initialization by

Region Growing Algorithm
Region growing based on

initialized seed points with
zero thresholding
Tumor region detected
Euclidean Distance Intensity

Calculation Correlation
Tumor Region Uniquely

Identified
End
Figure 7.1 Flowchart explaining the Sequential steps involved in the

proposed algorithm
138
7.2.1 PSO Algorithm
The PSO is capable of exhibiting a searching operation upon a

population of particles or a group of particles (Swarm) which is related to the
voxels present in the input MR brain images. The particles named as the
population are randomly generated. Each and every individual particle
represents a vivid solution and is characterized by a position vector. The
particles possess perfect mobility in a problem space, so that an optimized
result could be concluded during the search operation. The best known
position of the particle in the search space is described as best of the particle.
The best position of the particle, whilst comparing the position of the
neighbouring particles is expressed as global best ( best). The velocity of the
particle moving around the search space is expressed using a velocity vector.
Finding out the best position of the particles using PSO reduces the
complication in deriving the total number of clusters and centroid of the
voxels used in FCM algorithm [42, 137].
⃗⃗⃗ ( ) ⃗⃗⃗ ( ) (⃗⃗⃗ ( ) ⃗⃗⃗ ( )) .⃗⃗⃗⃗ ( ) ⃗⃗⃗ ( )/ (5.1)
⃗⃗⃗ - Velocity of the particles which lies within ⃗ (Range

expressed in terms of voxels: -5 to +5).
- PSO - TVAC (Particle Swarm Optimization based Time

Varying Acceleration Constant).
& - Uniform weight values ranging between 0 and 1.
⃗⃗⃗ ( ) - Best position of the particle linked with fitness function

( best).
⃗⃗⃗⃗ ( ) - Global best position of the particle ( best).
⃗⃗⃗ ( ) - Position vector of the particle.
- Time instant.
139
⃗⃗⃗ ( ) = ⃗⃗⃗ ( ) ⃗⃗⃗ ( ) (5.2
Equation (5.2) represents the updated velocity ⃗⃗⃗ ( ) and the

corresponding position vector updation of the particle as ⃗⃗⃗ ( ) [116, 137].
Normally, based upon the calculation of fitness function, intensity value of
the image and time duration, PSO based image segmentation can be
implemented [110]. Acceleration coefficients were assigned as 1.494
as described in [110, 137], whereas, the proposed PSO algorithm includes
TVAC (Time Varying Acceleration Constants) to enhance the optimization
procedure and better centroid value allocation for FCM algorithm. According
to TVAC criteria, decremented from 2.5 to 0.5 and is incremented
from 0.5 to 2.5 [137].
7.2.2 FCM Algorithm
Fuzzy C - means (FCM) algorithm is a basic clustering technique

introduced by Dunn, flavoured up by Bezdek and further promoted by Matteo
Matteucci. The voxels of the input MR brain image are classified and grouped
under ‗n‘ number of clusters formulated by PSO algorithm. ⃗⃗⃗ ( ), the
updated position vector of the particle, reckoned with the aid of PSO
algorithm, is the centroid value of FCM clustering procedure. Consequently,
the number of clusters for FCM algorithm is also assigned using PSO (with
reference from centroid value). Based on the position vector value, a center
voxel for each cluster is identified and assigned with a high membership
grade value through Fuzzy. The distance between the neighboring voxels and
the centroid voxel is evaluated and the voxels which possess least mean
distance are assigned with low membership grade value and are hierarchically
grown around the centroid value. To diminish the objective function value
involved in grouping of the voxels, the membership grade and the cluster
centres are iteratively updated.
140
∑ ∑ ‖ ‖ (5.3)
N - Number of data points given as input to the algorithm (Total

number of voxels or voxels)
K - Number of iterations involved.
C - Number of clusters into which the voxels are to be grouped.

⃗⃗⃗ ( ) of Equation (5.2) plays a major role in calculating ‗C‘.
- Centre vector for the clusters.
- Data points or voxels.
- Degree of membership for the ith data point of in cluster j.
‖ ‖ - Similarity measure or mean distance of the neighbouring

voxels present in the data point to the centre vector
(centroid value) of cluster j.
Degree of membership is denoted as,
(5.4)
‖ ‖
∑ ( )
‖ ‖
M - Fuzziness co – efficient derived from the overlapping of

clusters is represented as 1.5 ≤ m ≤ 2.5 for optimum
segmentation results [65].
Centre vector for the cluster is expressed as,
∑
∑
(5.5)
141
During the initial processing of the algorithm, .
- A random value initialized as, (Usually expressed as 0, 0.1,

0.2, 0.4….1), such that ∑ .
The output from PSO is given to FCM. This procedure reduces the
complexity of finding out the number of clusters and centroid value for FCM
[46]. Finally, both these two algorithms have a very good impact upon the
segmentation results offered by Region Growing algorithm.
7.2.3 Region Growing Algorithm
Let R be assumed as the entire image region. Segmentation involves

partitioning the R into n sub –regions, say, R1, R2, R3…Rn. So that,
1. ⋃ Indicates the complete description of the voxel in

a region.
2. is a connected region, i = 1, 2, 3 …. n. It denotes

connectivity between the points in a region.
3. R i ∩ Rj = for all i and j, where ‗i‘ represents the rows and ‗j‘
represents the columns of an image, i j. is the null set
operation. It defines that there should not be any intersecting
values between the regions.
4. P (Ri) = TRUE for i = 1, 2, 3 …. n. P (Ri) is the logical

prediction of the points in Ri.
5. P (Ri∩ Rj) = FALSE for i j. Conditions 4 & 5 represent the

properties to be exhibited by the voxels of the segmented
region in an image.
142
These steps indicate the conventional region growing algorithm and

its segmentation procedures [111]. The proposed PSO – FCM based Region
Growing algorithm handles the problem of growing regions and stopping
instances experienced by the conventional technique by offering a mean
distance calculation.
For voxels P and Q with coordinates (x, y) and (s, t) [111], the
Euclidean distance is calculated as
( ) (( ) ( ) ) (5.6)
The proposed method infers the identification of tumor region

bounded between the tissues in 3D MR brain images [111]. So, the above said
Euclidean distance is reformatted as
( ) (( ) ( ) ( )) (5.7)
Here, x, y, z and s, t, u described in Equation (5.7) are the spatial

domain values of the given input MR brain image with a Field of View (FOV)
matrix value as 256 × 256. The distances between the points are assessed by a
single seed point selection, say P. It includes the spatial coordinate values x, y
and z. The other spatial coordinates of Q (say s, t and u) are estimated based
on the minimal distance and intensity correlation criterion. Tumor detection
using average values of local Maxima is exhibited by Aminah Abdul Malek et
al [6], where region growing is performed based only on local maxima and
not upon local minima values. As a result of choosing local maxima, 0.94 is
the accuracy value proclaimed by Aminah Abdul Malek et al [6], which has
been considerably incremented using the proposed methodology.
143
Here, voxel P is the initialized seed point. Q is the neighbouring

voxel which has a similar intensity level when compared with voxel P. Based
upon the Euclidean distance calculation (De) and the intensity observance, Q
is grouped with P. Accompanying this procedure, the adjacent and the nearby
voxels which satisfy the above said criterion is grown around the initial seed
point P. With the assistance of the manually drawn largest diameter line by
the user, the seed points are initialized and the tumor region in contrast
enhanced T1 – w images is contoured using the tumor cut algorithm
proposed by Andac Hamamci et al [8]. Using the coordinates specified by the
line, seed points are formulated in both background and foreground regions to
perform segmentation operation in both higher and lower intensity levels of
the tumor region. The cellular automata algorithm performs the contouring
operation using this initialization, whereas, the proposed method requires
initial seed point selection and performs the region growing process using the
following conditions:
1. One time single seed point selection. (Selection of voxel - P)
2. Neighbouring voxels are automatically grouped with the

initial seed point based on Euclidean distance calculation (De).
3. The entire tumor region is detected depending upon the

correlation between the intensity levels of the neighbouring
voxels and the initially assigned seed point P.
This defines the novelty of the proposed region growing algorithm

which performs detection of tumor region using growing of region based on
distance calculation and intensity matching. The proposed methodology is
found to be less sensitive to seed point initialization. With the inclusion of the
distance calculation metric De, the time consumption for processing the entire
algorithm is considerably reduced.
144
7.3 IMPLEMENTATION
Step 1 : MR brain images are produced with the assistance of world‘s

advanced imaging systems provided by Siemens - Aera 1.5 Tesla X
4 G, DOT & TIM – 48 Channel & Siemens - 1.5 Tesla Magnetom
Avanto-Tim technologies.. The entire acquisition time for obtaining
the images is found to be less than 10 min. T1 weighted images
from a database can also be utilized to indicate the proficiency of
an algorithm, but the main ideology behind this method entirely
vests in assisting a radiologist to have clear examination and
diagnosis of the patients suffering from brain tumor.
Step 2 : The dimensions of the image are fixed with a limit ranging between
0 and 255.
Step 3 : Voxel values of the MR brain image are processed, so that they lie
between the lower and the upper limit of the specified range.
Step 4 : The seed point initialization is done for the tumor part in the input
image. With a minimal manual interaction, a voxel within the
tumor region is randomly chosen as the initial seed point.
Step 5 : The seed point value acts either as a median value or as a center
value to segregate the tumor region from the other bounded tissue
and edema regions.
145
Step 6 : The voxel values having minimum distance (De) and slight
intensity variations when compared with the initialized seed point
value are grown around it. This describes the working principle of
the algorithm. The region which is developed around the seed point
value (P) is continuously grown till the specified spatial intensity
levels are completely grouped.
Step 7 : The process of region growing is done until a clear level of tumor
region is exactly carved out from the other tissue regions (purely
based on spatial intensities between P and other neighbouring
voxels).
The novelty of this methodology is explicated as the Region

Growing algorithm and it is quite independent of the threshold value usually
expressed in conventional techniques. Due to the inclusion of PSO (Modified
PSO or MPSO) based FCM the threshold value used in conventional Region
Growing algorithms does not have any impact upon the proposed
combinational Region Growing algorithm. The suggested Region Growing
algorithm depends only upon the spatial intensity variations and not upon the
thresholding factor. Due to the difficulties experienced in obtaining live
clinical datasets with proper ethical standards, images of four patients and
other images from Harvard brain web database have been used to analyse the
functioning of the proposed methodology.
146
Input Images
PSO Based FCM

results
PSO based FCM

with
Region Growing
results

Figure 7.2 & 7.3 is completely attained from the patient suffering
from PNET (Primitive Neuro Ectodermal tumor). Tumor region present in
axial, coronal and sagittal axes of the MR brain is uniquely identified.
Compression of ipsilateral ventricular system is observed with patient 1. T2
FLAIR Axial with Contrast Enhancement, T2 Coronary and T1 Coronary
with Contrast Enhancement were used in clinical practices to diagnose PNET.
The proposed algorithm succeeded in identifying the tumor regions in all the
three images of Figure 7.2.
3D Multi Planar Reconstruction with Contrast Enhancement was

also used to diagnose PNET in patient 1. The proposed PSO based FCM with
Region Growing algorithm effectively segmented the tumor region present in
3D MPR with Contrast Enhanced images of Figure 7.3. MPR images are
often used to examine the anatomy of brain. A clear structure of Gyri and
147
Sulci is visualized in the input images of Figure 7.3. The ability of

segmentation possessed by the proposed algorithm is eventually proved in
Figure 7.3.
Input Images
PSO Based FCM results
PSO based FCM with

Region Growing results

Primitive Neuro Ectodermal tumor.
Image type: MPR (Multi Planar Reconstruction) 3D contrast enhanced
images
Figure 7.4 is completely derived from the patient suffering from

meningioma. Calcified moderately enhancing meningioma is notified in the
right parietal convexity with a dural tail. In this challenging case of tumor
identification, the proposed PSO based FCM with Region Growing algorithm
efficiently carved out the tumor region bounded between the tissues of grey
matter. Flair Axial, T1 Sagittal and T1 Coronary with Contrast Enhancement
were used by the radio surgeon to identify the tumor region located in the
meninges of the patient‘s brain.
148
Input Images
PSO based FCM with

Meningioma
After injecting gadolinium contrast agent, a heterogeneous

enhancement was notified in right ganglio capsular region and thalamus. Due
to the presence of intense edema, effacement of ipsilateral ventricular system
was also observed by the radiologist. The high grade Astrocytoma present
amid dense edema region is exactly identified and it is clearly depicted in
Figure 7.5.
Input Images

149
PSO based FCM with

high grade Astrocytoma
The above Figure shows the effectiveness of the proposed

methodology in resolving challenging clinical datasets where computer aided
diagnosis is extensively preferred by the radio surgeon. T2 Axial, T1 Sagittal
and T1 Coronary Contrast Enhanced image sequences were used to identify
the tumor region of patient 3.
Input Image
PSO based FCM with

high grade astrocytoma.
Image type: MPR (Multi Planar Reconstruction) 3D contrast enhanced image

150
A 3D MPR with Contrast Enhancement was also observed to

identify the vulnerability of the patient towards high grade Astrocytoma. The
proposed algorithm succeeded in segmenting the tumor region stated in
Figure 7.6. Even though continuous intensity variation is observed in the
tumor region of Figure 7.6, the proposed PSO based FCM with Region
Growing algorithm segmented the tumor region impressively. This proves
that the algorithm is less sensitive towards seed point initialization and small
intensity variations.
Patient 4 affected by supratentorial PNET was subjected to radio

therapy. The images of Figure 7.7 & 7.8 show the segmentation of the tumor
region remaining after radio therapeutic procedures. On injecting gadolinium
contrast agent, a heterogeneous enhancement was observed in the areas of
right parieto-occipital, cortical and subcortical regions. Larger volumes of
cerebral edema resulting in effacement of Sulci and right lateral ventricle
were visualized during the diagnosis. The tumor region bounded between the
dense edema portions of the first two images of Figure 7.7 was identified,
whereas, the third image of Figure 7.7 gives an eminent partition between the
tumor and the edema filled regions. Effective tumor identification is noticeable in
the third image. T2 Axial, T1 Sagittal with Contrast Enhancement and T1
Coronal image sequences were used to image the tumor region.
Input Images

151
PSO based FCM with

Figure 7.7 Segmented tumor region of patient 4 (age: 3) affected by

Input Images
PSO based FCM with


Image type: MPR (Multi Planar Reconstruction) 3D images
A 3D MPR sequence was used to examine the impact of the tumor

region and its edema portion upon the structures of Gyri and Sulci. Both the
tumor and edema regions covering the tissue structures of the brain were
exactly identified using the proposed algorithm and it is illustrated in Figure
7.8.
152
Figure 7.9, 7.10, 7.11, 7.12 & 7.13 are contrast enhanced and were
entirely obtained from Harvard Brain web database. These images were used
to validate the performance of the proposed PSO based FCM merged with
Region Growing algorithm.
Input
Images
PSO
Based
FCM
results
PSO based
FCM with
Region
Growing
results
Figure 7.9 Segmented tumor region of the patient suffering from

Pituitary Adenomas
Figure 7.9 was derived from the patient suffering from pituitary
adenomas. T1 axial images with contrast enhancement offered by gadolinium
were used to diagnose the patient. The proposed methodology successfully
identified the different sized tumor region with variations in intensity levels.
The successful identification of tumor region is depicted in Figure 7.9.
153
Input Images
PSO based FCM with

Figure 7.10 Segmented tumor region of the patient suffering from Nerve
Sheath Tumors
Figure 7.10 is obtained from the patient suffering from nerve sheath
tumors. Different tumor sizes are successfully identified using the proposed
methodology.
PSO Based FCM PSO based FCM with

Input Images
results Region Growing results
154
Figure 7.11 Segmented tumor region of the patient suffering from low
grade Glioma
In T1 axial view of Figure 7.11, the tumor region affecting the

patient is confirmed as low grade Glioma. Even under the influence of slight
noise distortions, the proposed algorithm clearly identifies the tumor region of
varying sizes.
PSO Based FCM PSO based FCM with

Input Images
results Region Growing results
155
Figure 7.12 Segmented tumor region of the patient suffering from high
grade Glioma
Variable sizes of the tumor region of the patient suffering from high
grade Glioma are perfectly identified with the aid of the proposed PSO based
FCM with Region Growing algorithm. In spite of the presence of white
layered regions in the tissues of brain very near to the tumor region, the
proposed algorithm effectively identifies the tumor region amid severe
complications. The effectiveness can be clearly understood by viewing the
segmentation results of Figure 7.12.
Figure 7.13 is attained from the patients suffering from low grade
Glioma and metastatic bronchogenic carcinoma. T2 - w Contrast Enhanced
156
image sequence is used to observe the cystic tumor region present in the left
occipital region of the brain affected with low grade Glioma.
Input Images
PSO based FCM with

Figure 7.13 Segmented tumor region of the patients suffering from

Glioma and Metastatic Bronchogenic Carcinoma
The first image of Figure 7.13 shows the tumor region of the patient
affected with low grade Glioma. The next two images, likely to be T2 - w
axial and T1 – w axial with contrast enhancement depict the patient suffering
from metastatic bronchogenic carcinoma. The suggested PSO bases FCM
with Region Growing algorithm effectively segments the tumor region and is
proved in Figure 7.13. The tumor region situated in the second image of
Figure 7.13 is densely surrounded by edema region. Using the proposed
algorithm, tumor region is clearly distinguished from the edema portions.
157
Figure 7.14 Magnetic Resonance Spectroscopic (MRS) Image

representing the diagnosis of patient 1 (age: 5) suffering
from Primitive Neuro Ectodermal tumor
The patient of age five, suffering from primitive neuro ectodermal

was subjected to the scan procedures denoted in Figure 7.2, and finally
underwent diagnosis with the aid of MRS imaging technique. From Figure
7.14, a radiologist could determine the activeness of the cancer cells involved
in the production of tumor. For an effective quantification and analysis of
tumor region, MRS based imaging is considered and extensively adopted in
clinical practices. The activity of the cancer cell is observed using the choline
to creatine ratio. This describes the seriousness of the tumor growth. T2 – w
FLAIR axial, T1 – w contrast enhanced sagittal and T2 – w coronal are shown
in Figure 7.14, where tumor identification and quantification are done
simultaneously. Similar tumor identification procedure is exhibited using the
proposed PSO based FCM with Region Growing methodology and is
described in Figure 7.2. The proposed technique gives out an initial level
158
ideology for a radiologist in diagnosing the patients suffering from different

types of tumor. It can also be used in both pre and post radio therapy and
radio surgical procedures as an initial level sketching of the tumor affected
region and to have a clear discrimination between the tumor and the edema
regions.
Segmented result
PSO Based FCM from Region Proposed
Input Image
results Growing Algorithm
algorithm
Figure 7.15 Individual results from PSO based FCM and Region
Growing algorithms for comparison
Figure 7.15 is a comparison result obtained by segmenting the

second image available in Figure 7.15, attained from the patient suffering
from supratentorial Primitive Neuro Ectodermal Tumor. Tumor region mixed
up with the edema portion and normal tissue region is segmented using the
Region Growing algorithm and there is a leakage of detected tumor region,
clearly visible in the third image of Figure 7.15. Partial identification of tumor
region is characterized by PSO based FCM algorithm and it is observed from
the second image of Figure 7.15. Exact identification of tumor region without
any subsequent leakage is offered by the proposed algorithm and it is
visualized in the fourth image of Figure 7.15. It is conclusive that the
proposed PSO based FCM with Region Growing algorithm overrides both
PSO based FCM and Region Growing algorithms in all the aspects of tumor
identification.
159
Table 7.1 and Figure 7.16 illustrate that the proposed algorithm has
low MSE values when compared to PSO based FCM algorithm. MSE is
defined as Mean Square Error obtained by subtracting the input and the output
values. So, the proposed methodology is capable of producing segmented
images with low MSE values. Low MSE values depict lower error rates
produced by the algorithm. Higher MSE values indicate maximum
deformation in the segmented results [82]. 41.617 is the average MSE value
offered by Govindaraj Vishnuvarthanan and Murugan Pallikonda Rajasekeran
[45], whereas, 0.000088 is the average MSE value produced by the proposed
methodology. Table 7.1 and Figure 7.17 indicate the comparison of PSNR
values. PSNR is used for quality estimation in image processing where
thresholding plays a primitive role Sathya and Kayalvizhi [123]. Normally,
the value of PSNR should be considerably high. High PSNR values indicate
less sensitivity of the output image towards noise signals. The segmented
images obtained using the recommended algorithm produces high PSNR
values when compared to Region Growing and PSO based FCM algorithms.
The proposed methodology produces PSNR value above 60dB, which
exemplifies excellent segmentation results. The proposed PSO based FCM
merged with Region Growing algorithm offers 81.457 as an average PSNR
value and is far better than the PSNR value offered by the methodology
recommended by Sathya and Kayalvizhi [123], where thresholding requires
continuous tuning.
160
Table 7.1 Comparison parameters utilized to evaluate the efficiency of

the proposed PSO based FCM merged with Region Growing
Algorithm
Computational Memory
PSNR
Input Images MSE TC DOI time in Requirement
in dB
Seconds in Bytes
T2 FLAIR Axial with CE 8.38E-05 88.8992 0.9773 0.9885 22.87 5.3098e+014
T2 Coronary 2.06E-04 84.9964 0.9894 0.9947 29.58 4.8294e+014

T1 Coronary with CE 1.72E-04 85.7813 0.9879 0.9939 23.61 4.8221e+014
MPR with CE 1.38E-04 86.7262 0.9773 0.9885 30.50 4.8162e+014
MPR with CE 2.07E-04 84.9611 0.9798 0.9898 22.87 4.8225e+014
Flair Axial 1.04E-04 87.9558 0.9978 0.9989 29.58 6.8894e+014
T1 Sagittal CE 9.33E-05 88.4333 0.9985 0.9992 23.61 6.8910e+014
T1 Coronary CE 1.33E-04 86.9084 0.996 0.998 30.50 6.8893e+014
T2 Axial 8.88E-05 88.649 0.9793 0.9896 29.33 6.7532e+014
T1 Sagittal CE 1.02E-04 88.0262 0.992 0.996 32.68 6.8725e+014
T1 Coronary CE 9.58E-05 88.3182 0.9837 0.9918 23.35 5.3098e+014
MPR with CE 8.91E-05 88.631 0.9754 0.9876 22.74 4.8294e+014
T2 Axial 8.23E-05 88.9791 0.7697 0.8699 37.51 4.8221e+014
T1 Sagittal CE 1.11E-04 87.6599 0.9194 0.958 21.33 4.8162e+014
T1 Coronal 9.61E-05 88.3034 0.9187 0.9576 15.39 4.8225e+014
T1 MPR 8.24E-05 88.9714 0.8581 0.9237 31.64 6.8894e+014
T1 MPR 8.09E-05 89.0525 0.8289 0.9064 32.84 6.8910e+014
T1 Axial with CE 0.0013 76.8675 0.9936 0.9968 30.89 4.4388e+013
T1 Axial with CE 0.0013 76.8706 0.9909 0.9954 29.81 4.4336e+013
T1 Axial with CE 0.0018 75.5881 0.9976 0.9988 31.92 4.4484e+013
T1 Axial with CE 0.0018 75.535 0.9894 0.9947 30.76 4.4377e+013
T1 Axial with CE 0.0018 75.6887 0.9894 0.9947 29.44 4.4321e+013
T1 Axial with CE 0.0012 77.2633 0.9909 0.9954 29.39 4.4293e+013
T1 Axial with CE 0.0012 77.4271 0.986 0.993 31.59 4.4240e+013
T1 Axial with CE 0.0012 77.1913 0.9842 0.992 25.97 4.4192e+013
T1 Axial with CE 0.0011 77.5965 0.9801 0.9899 31.17 4.4139e+013
T1 Axial with CE 0.0017 75.8265 0.9894 0.9947 31.70 4.4268e+013
T1 Axial with CE 0.0017 75.8134 0.986 0.993 29.70 4.4202e+013
161
T1 Axial with CE 0.0017 75.8118 0.9822 0.991 31.45 4.4159e+013

T1 Axial with CE 0.0016 76.1079 0.9756 0.9877 20.99 4.4388e+013
T1 Axial with CE 0.0016 76.1316 0.9779 0.9888 31.63 4.4336e+013
T1 Axial with CE 0.0016 76.0106 0.9732 0.9864 32.51 4.4484e+013
T1 Axial with CE 0.0016 76.2184 0.9597 0.9794 30.20 4.4377e+013
T1 Axial with CE 0.0015 76.2323 0.9681 0.9838 24.48 4.4321e+013
T1 Axial with CE 0.0015 76.3924 0.9756 0.9877 27.99 4.4293e+013
T2 Axial with CE 0.0014 76.5719 0.7487 0.8563 80.20 5.5813e+013
T2 Axial with CE 0.0018 75.7003 0.9258 0.9615 36.48 5.5787e+013
T1 Axial with CE 0.0012 77.2709 0.7183 0.8361 26.67 5.2370e+013
Dice Overlap Index
Jaccard Index, also called the Tanimoto Coefficient Index, is used in

order to prove the segmentation accuracy of the algorithm. Higher the value
of Jaccard index better is the segmentation result of the proposed algorithm
[30]. The proposed PSO – FCM merged with Region Growing algorithm
offers higher Jaccard Index values when compared with the other
segmentation algorithms. The Jaccard index value proposed by Sultan
Aljahdali and Zanaty [21] is 0.604, whereas, the proposed algorithm produces
0.952 as an average Jaccard index value.
The time required for processing the proposed methodology is much

higher when compared with the other soft computing algorithms. Table 7.1
and Figure 7.21 clearly depict computational or processing time. Accuracy in
identifying the tumor region plays a major role in clinical diagnosis. Offline
examination of pathologies present in human brain could extensively support
radio therapeutic procedures. Based on this motivation, a semi – automated
algorithm which includes the working principles of optimization, clustering
and region based growing is proposed through this paper. A detailed analysis
of pathologies with minimal user interaction with purported efficiency in
identifying the tumor region is observed through the proposed methodology.
162
Inclusion of all three techniques has deliberately incremented the time

consumption level for processing the input MR brain images. While
considering the segmentation efficiency, higher time consumption comes into
lower level consideration.
Mean Square Error

2.00E-02
1.80E-02
1.60E-02
1.40E-02
MSE values
1.20E-02
1.00E-02
MSE
8.00E-03
6.00E-03
4.00E-03
2.00E-03
0.00E+00
PSO based FCM with PSO based FCM Region growing
Algorithm
Algorithms
Figure 7.16 Comparison of MSE values
Peak Signal to Noise Ratio in dB

90
80
Peak Signal to Noise Ratio
70
60
50 PSNR in dB
40
30
20
10
0
Algorithm
Algorithms
Figure 7.17 Comparison of PSNR values

163
From Figure 7.16 and Figure 7.17, it is clearly understandable that

the proposed PSO – FCM merged with Region Growing algorithm
produces better MSE and PSNR values when compared with Region
Growing and PSO based FCM algorithms. 0.0008, 81.45 are the average
MSE and PSNR values produced by the proposed methodology. Region
growing algorithm produces 0.0168 and 66.17 as the average MSE and
PSNR values. FCM based PSO produces 0.0176 and 66.027 as the average
MSE and PSNR values. For a better segmentation algorithm, PSNR value
should lie between 40 decibels and 60 decibels. But the proposed seed
based region growing algorithm is capable of producing PSNR values of
above 60 dB. This proves the efficiency of PSO based FCM merged with
Region Growing algorithm. Segmented image with better PSNR values can
be extensively used in image fusion and compression techniques. Both
MSE and PSNR are interrelated with each other. MSE is used to observe
the error rates and deformation in the segmented image.
DOI value is calculated using Jaccard Index values. For better

image segmentation, the DOI values should be higher. The proposed PSO
based FCM with Region Growing algorithm produces higher DOI values in
identifying the tumor region on comparison with Region Growing and PSO
based FCM algorithms.
164
Jaccard Tanimoto Coeffecient Index

1.2
0.8
TC Values
TC
0.6
0.4
0.2
0
PSO based FCM with PSO based FCM Region growing Algorithm
Algorithms
Figure 7.18 Comparison of Jaccard Tanimoto Coefficient (TC) Index

values
Dice Overlap Index

1.2
0.8
DOI values
DOI
0.6
0.4
0.2
0
PSO based FCM with PSO based FCM Region growing Algorithm
Algorithms
Figure 7.19 Comparison of Dice Overlap Index (DOI) values
Figure 7.18 and Figure 7.19 define the efficiency of the proposed
algorithm in terms of Jaccard Tanimoto Coefficient Similarity Index (TC) and
Dice Overlap Index (DOI). The average TC and DOI values produced by the
proposed algorithm are 0.9529 and 0.9744. On the other hand, 0.2157 and
165
0.3530 are the average TC and DOI value produced by PSO based FCM
algorithm. Region growing algorithm produces 0.9419 and 0.9638 as the
average TC and DOI value. When expressed in percentage, 95.29 and 97.44
are the average TC and DOI values produced by the proposed methodology.
DOI value for segmenting the MR images proposed by Ulas Bagci et al [150]
is 88.15, which is quite lesser than the DOI value produced by the proposed
algorithm.
The memory required for processing the algorithms is inferred in

Table 7.1. When compared with the other two algorithms, the proposed
algorithm utilizes higher memory space for processing. Usually, memory
required for algorithm execution is expressed in terms of bytes. The entire
process was executed using a desktop equipped with Pentium 4 processor
2.27 GHz and 2 GB RAM. In efficient offline clinical diagnosis, memory
requirement and time consumption levels are matters to be ignored.
Figure 7.20 explicates the overall memory required by the processor

to execute the algorithms in a MATLAB platform. The average memory
required by the processor to execute the proposed region growing algorithm is
2.84E+14 Bytes. In addition, 1.97E+14 and 1.62E+14 bytes are the average
memory required by the processor for executing PSO based FCM and Region
Growing algorithms. Memory space occupancy is not a constraint for a
detailed tumor analysis.
166
Memory Requirement in Bytes

3.00E+14
Memory requirement values
2.50E+14
2.00E+14
Memory Requirement in MB
1.50E+14
1.00E+14
5.00E+13
0.00E+00
PSO based FCM PSO based FCM Region growing
with Region Growing Algorithm
Algorithm
Algorithms
Figure 7.20 Comparison of memory requirement values in terms of Bytes
The average computational time required for processing the

algorithms is illustrated in Figure 7.21. It is clearly identifiable that the
proposed algorithm requires average time duration of 29.87 seconds for
processing 38 input MR brain images, whereas, PSO based FCM and Region
Growing algorithms require 28.98 and 5.53 seconds for processing the input
MR brain images. Time factor plays a major role in clinical diagnosis.
Normally, large numbers of patient data are analyzed by a radiologist. The
main aim of developing the segmentation algorithms is to reduce the
tediousness in diagnosing the pathologies and save precious life. The
segmentation algorithms should assist the radiologist by offering effective
segmentation. The proposed PSO based FCM with region growing algorithm
satisfies this criteria.
167
Time in Seconds
35
29.87061747 28.98055
30
25
Time duration
20
15
Time in Seconds
10
5.530628351
5
0
Algorithm
Algorithms
Figure 7.21 Comparison of computational time in terms of Seconds
SI and OF values of an algorithm should be nearly 1 [115]. EF value

of a segmentation algorithm should be close to 0 [115]. Further, algorithms
producing SI and OF values above 0.7 can be considered as better
segmentation algorithm [115]. The proposed methodology outperforms the
conventional PSO based FCM and Region Growing algorithms in terms of SI,
OF and EF, which is clearly stated in Table 7.2. The above said values were
derived by comparing the functionary of the proposed methodology with an
expert manual segmentation or Gold Standard [65].
Similarity Index Overlap Extra

Algorithm
(SI) Fraction (OF) Fraction (EF)
PSO based FCM 0.9333 0.9211 0.0263
Region Growing
0.9589 0.9722 0.0556
algorithm
Proposed Algorithm 0.9600 0.9730 0.0541
168
Comparison of the proposed methodology with the other techniques

using OF or sensitivity values is illustrated in Figure 7.22. Since, T1 – w, T2
– w, FLAIR and MPR type brain images are used, the proposed technique
could be compared with the other techniques mentioned in Figure. 7.22. It is
highly commendable that the proposed PSO based FCM with Region
Growing algorithm outperforms the other competitive algorithms in terms of
sensitivity. The average sensitivity value proposed by Vijayakumar
Chinnadurai and Gharpure Damayanti Chandrashekhar [157] ranges between
92.47 and 94.37. The proposed algorithm offers 96% as the average
sensitivity value.
Comparison of SI or OF values
1.2
0.96
1 0.88
0.82
0.75
0.8 0.65 0.68
Values
0.62
0.6 0.51
0.4
0.2
0
K.V.Leemput et al (2001)
proposed algorithm
Hassan Khotanlou et al
Petronella Anbeek et al
B. Johnston et al (1996)
A.O. Boudraa et al (2000)
A.P. Zijdenbos et al (2002)
Rasoul Khayat et al (2008)
(2011)
(2005)
Algorithms
Figure 7.22 Overall comparison of sensitivity values with an

approximation offered by Rasoul Khayati et al [115].
169
Comparison of FIS, SOM - FKM, PSO - FCM and PSO -

FCM with RGA
120
100 SOM - FKM

FIS
80
Values
PSO - FCM
60
PSO - FCM with RGA
40
20
0
PSNR in dB TC DOI E. TIME in
seconds
Figure 7.23 Comparison of PSO based FCM improved by RGA with

other soft computing algorithms
From Figure 7.23, it is clearly understandable that the proposed PSO

based FCM merged with RGA has better values in all terms. It offers an
average PSNR value of 80 decibels, 95.29% as average TC value and 97.44%
as average DOI value. Time consumption is quite higher due to the
combination three different techniques. The proposed algorithm could be
extensively used in effective clinical diagnosis.
Table 7.3 Overall comparison of SI, OF and EF values

Algorithm
(SI) (OF) (EF)
FIS 0.7119 0.6364 0.1515
SOM - FKM 0.9189 0.8718 0.0256
PSO - FCM 0.9333 0.9211 0.0263
Proposed
0.9600 0.9730 0.0541
Algorithm
170
Table 7.3 represents the overall comparison of Similarity Index (SI),

Overlap Fraction (OF) and Extra Fraction (EF) values between FIS, SOM –
FKM, PSO – FCM and proposed PSO – FCM merged with RGA algorithm.
The recommended PSO based FCM merged with RGA algorithm surpasses
the other competitive algorithms by producing better SI and OF values, which
is of utmost consideration for determining the accuracy of a medical image
segmentation algorithm.
Comparison of MSE values

45
40
35
30
MSE values
25
20
MSE
15
10
5
0
FIS SOM - FKM PSO - FCM PSO - FCM
with RGA
Algorithms
Figure 7.24 Comparison of MSE values derived from PSO based FCM
improved by RGA with other soft computing algorithms
From Figure 7.24, PSO based FCM merged with RGA requires
2.81E+14 bytes as an average memory requirement value for processing 38
input MR images. The memory required by the proposed algorithm is quite
higher when compared with normal PSO based FCM algorithm, but lower
when compared to the memory occupied by FIS and SOM – FKM algorithms.
Utilization of three different techniques for segmentation has led to this
memory consumption levels. While making an analysis, by consideration of
segmentation accuracy, memory requirement is preferred to have the least
significance.
171
The deformation rate of the segmented tumor region has also been
considerably reduced and it is justified in Figure 7.24, which implies the
averaged values of MSE. Time consumption becomes a matter to be ignored
when the segmentation efficiency levels are drastically improved.
MEMORY REQ in bytes

3.50E+14
3.00E+14
2.50E+14
MEMORY REQ
2.00E+14
Bytes
in bytes
1.50E+14
1.00E+14
5.00E+13
0.00E+00
FIS SOM - FKM PSO - FCM PSO - FCM
with RGA
Algorithms
Figure 7.25 Comparison of memory requirement values
Comparison of SI, OF and EF values

1.2
1
Similarity Index (SI)
0.8
Paremeters
Overlap Fraction (OF)

0.6 Extra Fraction (EF)
0.4
0.2
0
FIS SOM - FKM PSO - FCM Proposed Expert
Algorithm Segmentation
(Gold
Standard
Images)
Methodologies
Figure 7.26 Comparison of soft computing techniques with Gold

standards using Similarity Index, Overlap Fraction and
Extra Fraction
172
From Figure 7.26, the values of Similarity Index (SI) and Overlap
Fraction are quite higher in PSO – FCM with RGA algorithm. Extra Fraction
(EF) is found to be quite lower in SOM – FKM algorithm. Taking into
account, the proposed PSO based FCM merged with region growing
algorithm has a remarkable efficiency in analyzing the pathologies and
segmenting the tissue regions present in the input MR brain images. PSO
based FCM with RGA has made a benchmark in tumor identification and
tissue segmentation on comparison with other soft computing techniques and
manual expert segmentations.
The proposed PSO based FCM merged with region growing

algorithm is efficient in terms of PSNR, MSE, Jaccard Index, DOI, SI, OF
and EF values in comparison with PSO based FCM and Region Growing
algorithms. T1 - w, T2 – w, FLAIR and MPR type MR brain images are
segmented using the proposed algorithm, which affirms that the proposed
methodology has the capability to segment heterogeneous type tumors in all
the axes of MR brain image sequences. The level of tumor infiltration is
clearly distinguished from the edema portion. Large volume of clinical
datasets can be analyzed efficiently. Thus the algorithm assists the radiologist
in diagnosing the tumor region with minimized manual operation. Future
work of the algorithm includes segmenting MRS images (Magnetic
Resonance Spectroscopy) and identifying the volumes of white Matter, Grey
matter and Cerebrospinal Fluid regions.
173
CHAPTER 8
CONCLUSION AND FUTURE WORK
The membership parameters assignment for the FIS algorithm

requires manual intervention. One advantage of FIS algorithm is that, it does
not require frequent tuning of threshold for performing the segmentation
process. The usage of FIS is limited in segmenting T2 – W and T2 – W
contrast Enhanced MR brain images. Other modalities of MR brain images,
including the images with strong bias fields (Intensity inhomogeneity) cannot
be properly segmented by FIS algorithm. Thus, FIS could be used as an initial
level sketching tool for identifying the pathologies and could be used in
minimal levels for diagnosing procedures. Computational complexity and
higher memory space occupation experienced by FIS is considerably reduced
by the introduction of SOM based FKM technique.
The recommended hybrid SOM – FKM algorithm excels FIS

algorithm in the process of segmenting multimodal MR brain images. SOM
based FKM requires minimal manual intervention to segment the tissues of
brain, especially GM, WM and CSF regions, along with the identification of
different types of tumor at different locations. SOM based FKM excels in
segmenting T2 – Weighted (with and without CE), FLAIR type images (with
and without CE). MPR type MR brain images are also excellently segmented
using SOM based FKM technique. Diagnosing the patients within reduced
time duration is greatly achieved through the combination of SOM and FKM
174
techniques. Improving the values of DOI, TC and PSNR produced by SOM –

FKM algorithm has to be done. Time duration for processing the images and
MSE values can also be further reduced. The proposed SOM - FKM
algorithm can be extensively used in pre and post radio surgical applications.
Traditional techniques described in related works are capable of

segmenting either T1 or T2 weighted and FLAIR type images. But, it is
clearly identified that the proposed PSO based FCM algorithm is capable of
segmenting all T1, T2, MPR and FLAIR type brain images with all sequential
axis such as coronal, sagittal and axial. Also, images with low radio frequency
signals are segmented using PSO – FCM, which proves the robustness of the
algorithm. Future work lies with improving Jaccard Index and DOI values,
with reduction in memory requirement for processing the algorithm. PSO
plays a prominent role in offering cluster values to the FCM algorithm, based
on which the segmentation process is carried out. Thus a new approach to
segment MR brain images using PSO based FCM algorithm is proposed
through this research. Further, the convergence rate of FCM algorithm is
reduced. Concurrently, segmentation accuracy of FCM algorithm is
prominently improved with the inclusion of PSO technique.
The proposed PSO based FCM merged with Region Growing

algorithm is efficient in terms of PSNR, MSE, Jaccard Index, DOI, SI, OF
and EF values on comparison with PSO based FCM, FIS and SOM based
FKM algorithms. T1 - w, T2 – w, FLAIR and MPR type MR brain images are
efficiently segmented using PSO – FCM with region growing algorithm,
which affirms that the proposed methodology has the capability to segment
heterogeneous type tumors in all the axes of MR brain image sequences. The
level of tumor infiltration is clearly distinguished from the edema portion.
Large volume of clinical datasets can be analyzed efficiently. Thus
PSO – FCM with region growing algorithm assists the radiologist in
175
diagnosing the tumor region with minimized manual operation and is

considered to be the optimum medical images segmentation algorithm
through this research work. Future work of the algorithm includes segmenting
MRS images (Magnetic Resonance Spectroscopy) and identifying the
volumes of white Matter, Grey matter and Cerebrospinal Fluid regions.
8.2 CLINICAL ASSISTANCE
Normal tissue regions present in an MR brain image can be

segmented by the above - said soft computing algorithms. Also, tumor region
bounded between the tissues of the brain can be efficiently segregated.
Further, tumor infiltration levels and edema covered portions are intellectually
viewed with the support offered by the oft computing algorithms suggested
through this research. Larger data volumes and complicated clinical cases can
be easily processed with lesser manual intervention with the usage of afore -
mentioned methodologies. Finally, segmentation of multimodal brain images
of all axes, with the identification of heterogeneous tumor regions is made
available for clinical oncology.
8.3 CONCLUSION
Amid the presence of several soft computing techniques, this thesis

recommends that PSO based FCM with Region Growing Algorithm could be
profusely used in offline clinical practices for effective patient diagnosis.
8.4 SCOPE FOR FUTURE WORK
1. Volumes of tumor and volumes of normal tissue regions can

be assessed, where minimal exploration has been made.
176
2. Various combinations of optimization and soft computing

technologies could be attempted to derive better segmentation
efficiency.
3. Computational time involved in executing the algorithms

could be further reduced.
4. Segmentation of functional MRI (f - MRI) and Positron

Emission Tomography (PET) images can also be done using
the above said methodologies, so as to serve different levels of
medical image applications.
5. The same methodologies can also be used to identify and

analyze the pathologies found in kidney, liver, lungs and so
on.
6. The proposed soft computing algorithms can be impinged in

the Field Programmable Gate Array (FPGA) of a clinical MRI
scanner, so that the aberrant regions and tissues present in the
brain could easily be visualized.
177
APPENDIX 1
BASICS OF AN MRI SCANNER
Magnetic Resonance Imaging (MRI) helps in obtaining a structural

three dimensional image of the internal parts of human body. It works based
on the principle of Nuclear Magnetic Resonance (NMR) and it provides a
better augmentation in analysing organs such as heart, lungs, pelvis, abdomen
and soft tissues in brain.
80% of our human body is made of hydrogen atoms. Hydrogen atom

has a strong positive nucleus comprising proton particle. This proton has its
own spin velocity, spin direction and axis of rotation. The parallel alignment
of the hydrogen nuclei is set by a super conducting magnet cooled with
helium liquid. The hydrogen nuclei, at this state, are subjected towards a radio
frequency radiation which disturbs the analogous alignment of the hydrogen
nuclei. As the nuclei absorb some quantity of radio frequency signal, they
shift from lower energy level state to higher energy level state. As soon as the
radio frequency radiation is stopped the nuclei get back to their original
position with some emission of relative energy. This emitted energy is
produced as a result of shifting of nuclei from excited state of energy to
ground state, which is recorded and further processed by using a signal
processor. This procedure gives out a spatial image of the internal structure of
human body.
178
The same procedure is used repeatedly to obtain 3D (MR) brain

images. In an MRI scanner, two dimensional Fourier transforms is done over
the obtained image structures. A simple setup of an (MRI) scanning system is
shown in Figure A.1.1.
Vacuum Helium for cooling

creation the magnets
Super
device
conducting
Speed Linear Acceleration magnet
accelerating guide
electron gun Radiation
focusing
Power
head
Waveguide supply to RF
coils
Patient lying bed
Modulating Control
device panel
Figure A 1.1 Basic MRI Machine

179
REFERENCES
1. Acharya. D.P., Panda. G., Mishra. S and Lakhshmi. Y.V.S., Bacteria

foraging based independent component analysis, in: Proceedings of
the international Conference on Computational Intelligence and
Multimedia Applications, vol. 2 (2007), 527–531.
2. Ahmed. M.N, Yamany.S.M, Mohamed. N, Farag. A.A and Moriarty.

T, A modified fuzzy c-mean algorithm for basis field estimation and
segmentation of MRI data, IEEE Transactions on Medical Imaging,
21 (3) (2002), 193–199.
3. Albert Huang, Rafeef Abugharbieh, Roger Tam, and Alzheimer‘s

Disease Neuroimaging Initiative, A Hybrid Geometric–Statistical
Deformable Model for Automated 3-D Segmentation in Brain MRI,
IEEE Transactions on Biomedical Engineering, Vol. 56, No. 7
(2009).
4. Alhoniemi. E, Himberg. J, Parhankagas. J and Vesanta. J, SOM

Toolbox for Matlab v2.0, <http://www.cis.hut.fi/ projects/
somtoolbox>, (2005).
5. Alirezaie. J, M. Jernigan and C. Nahmias, Automatic segmentation

of cerebral MR images using artificial neural networks, IEEE
Transactions on Nuclear Science, 45 (4) (1998), 2174–2182.
6. Aminah Abdul Malek et al, Region and Boundary Segmentation of

Micro calcifications using Seed-Based Region Growing and
Mathematical Morphology, International Conference on
Mathematics Education Research, Elsevier Ltd,
doi:10.1016/j.sbspro.2010.12.088, (2010), 1877- 0428.
7. Amir Ehsan Lashkar, A Neural Network based Method for Brain

Abnormality Detection in MR Images Using Gabor Wavelets,
International Journal of Computer Applications, 4 (7) (2010), 0975
– 0987.
180
8. Andac Hamamci, Nadir Kucuk, Kutlay Karaman, Kayihan Engin

and Gozde Unal, Tumor-Cut: Segmentation of Brain Tumors on
Contrast Enhanced MR Images for Radiosurgery Applications, IEEE
transactions on medical imaging, vol. 31 (2011), 790 - 804.
9. Andres Ortiz, Antonio A. Palacio, Juan M. Gorriz, Javier Ramírez

and Diego Salas-Gonzalez, Segmentation of Brain MRI Using SOM-
FCM-Based Method and 3D Statistical Descriptors, Computational
and Mathematical Methods in Medicine, Article ID 638563, <
http://dx.doi.org/10.1155/2013/638563>, (2013).
10. Anthony Bianchi, Bir Bhanu, Virginia Donovan and Andre Obenaus,
Visual and Contextual Modeling for the Detection of Repeated Mild
Traumatic Brain Injury, IEEE Transactions on Medical Imaging,
Vol. 33, No. 1(2014), 11 – 22.
11. Ardeshir Goshtasby. A, 2-D and 3-D Image Registration for

Medical, Remote Sensing, and Industrial Applications, Wiley-
Interscience publication, ISBN 0-471-64954-6.
12. Arnaldo Mayer and Hayit Greenspan, An Adaptive Mean-Shift

Framework for MRI Brain Segmentation, IEEE Transactions on
Medical Imaging, Vol. 28, No. 8 (2009).
13. Arsuaga. E and Díaz. F, Topology preservation in SOM,

International Journal of Mathematical and Computer Sciences, 1 (1)
(2005), 19–22.
14. Artem Mikheev, Gregory Nevsky, Siddharth Govindan, Robert

Grossman and Henry Rusinek, Fully Automatic Segmentation of the
Brain from T1-Weighted MRI Using Bridge Burner Algorithm,
Journal of Magnetic Resonance Imaging, 27 (2008), 1235–1241.
15. Atiq Islam, Syed M. S. Reza and Khan M. Iftekharuddin,

Multifractal Texture Estimation for Detection and Segmentation of
Brain Tumors, IEEE Transactions on Biomedical Engineering, Vol.
60, No. 11 (2013), 3204 – 3215.
16. Ayelet Akselrod-Ballin, Meirav Galun, Moshe John Gomori, Ronen

Basri and Achi Brandt, Atlas Guided Identification of Brain
Structures by Combining 3D Segmentation and SVM Classification,
R. Larsen, M. Nielsen and J. Sporring (Eds.): MICCAI 2006, LNCS
4191(2006), 209–216.
181
17. Ayse Demirhan and Inan Guler, Combining stationary wavelet

transform and self-organizing maps for brain MR image
segmentation, Engineering Applications of Artificial Intelligence, 24
(2011), 358–367.
18. Baki Hekimoglu, Hasan Aydin, Volkan Kızılgoz, Idil Güneş Tatar,
Ali Rıza Ugan and İrem Paker, Identification and Localization of
Prostate Cancer with Combined Use of T2-Weighted, Diffusion
Weighted MRI and Proton MR Spectroscopy, Correlation with
Histopathology, Advances in Molecular Imaging, 3 (2013), 23-36.
19. Ben George. E and Karnan. M, MR Brain Image Segmentation using

Bacteria Foraging Optimization Algorithm, International Journal of
Engineering and Technology, 4 (5) (2012).
20. Berengere Aubert-Broche, Mark Griffin, G. Bruce Pike, Alan C.

Evans, and D. Louis Collins Twenty New Digital Brain Phantoms
for Creation of Validation Image Data Bases, IEEE Transactions on
Medical Imaging, Vol. 25, No. 11 (2006), 1410 – 1416.
21. Bernis R, MRI Brain Segmentation, Matlab Central, File Exchange.

G. Dougherty, Digital Image Processing Applications, Cambridge
University Press (2009).
22. Bezdek. J. C, Pattern Recognition with Fuzzy Object Function

Algorithms, New York, 1981.
23. Bouchet A, Pastore J and Ballarin. V, Segmentation of medical

images using fuzzy mathematical morphology, JCS, 7(3) (2007),
256-62.
24. Boudraa A.O., Dehakb S.M.R., Zhu Y.M., Pachai C., Bao Y.G. and
Grimaud , Automated segmentation of multiple sclerosis lesions in
multispectral MR imaging using fuzzy clustering, Computers in
Biology and Medicines, 30(2000), 23–40.
25. Brain Extraction tool (BET), FMRIB Centre, Nuffield Department

of Neurosciences, University of Oxford, <http:// www. fmrib. ox. ac.
uk/analysis/research/bet/>.
26. Burt. P and Adelson, E, The Laplacian Pyramid as a Compact Image

Code, IEEE Transactions on Communications, 31 (4) (1983), 532–
540.
182
27. Callaghan P, Principles of Nuclear Magnetic Resonance

Microscopy, Oxford University Press, ISBN 0-19-853997-5 (1994).
28. Canny. J, A computational approach to edge detection, IEEE Trans.

Pattern Analysis and Machine Intelligence, Vol 8 (1986), 679-714.
29. Caselles. V, Kimmel. R and Sapiro. G, Geodesic active contours,

International Journal of Computer Vision, 22(1) (1997), 61-79.
30. Chaplot S, Patnaik L.M and Jagannathan N.R, Classification of

magnetic resonance brain images using wavelets as input to support
vector machine and neural network, Biomed. Signal Processing
Control, (2006), 86-92.
31. Chih-Tang Chang, Jim Z. C. Lai and Mu-Der Jeng, A Fuzzy K-

Means Clustering Algorithm Using Cluster Center Displacement,
Journal of information science and engineering, 27 (2011), 995-
1009.
32. Christos G. Xanthis, Ioannis E. Venetis, Chalkias A. V and Anthony

H. Aletras, MRISIMUL: A GPU-Based Parallel Approach to MRI
Simulations, IEEE Transactions on Medical Imaging, Vol. 33, No. 3
(2014), 607 – 617.
33. Colm Elliott, Douglas L. Arnold, Louis Collins. D, and Tal Arbel,
Temporally Consistent Probabilistic Detection of New Multiple
Sclerosis Lesions in Brain MRI, IEEE Transactions on Medical
Imaging, Vol. 32, No. 8 (2013), 1490 – 1503.
34. Coroso J, Sharon E and Dube S, Efficient multilevel brain

segmentation with integrated Bayesian model classification, IEEE
Transactions on Medical Imaging, (2010).
35. Datta. T, Misra. I.S, Mangaraj. B.B and Imtiaj. S, Improved adaptive
foraging algorithm in optimization of antenna array for faster
convergence, Progress in Electromagnetics Research, 1 (2008),
143–157.
36. Dave. R.N, Characterization and detection of noise in clustering,

Pattern Recognition Letters, 12 (1991), 657–664.
183
37. Dimitrios Zarpalas, Polyxeni Gkontra, Petros Daras and Nicos

Maglaveras, Gradient-Based Reliability Maps for ACM-Based
Segmentation of Hippocampus, IEEE Transactions on Biomedical
Engineering, Vol. 61, No. 4 (2014), 1015 – 1026.
38. El-Hachemi Guerrout, Ramdane Mahiou and Samy Ait-Aoudia,

Medical Image Segmentation on a Cluster of PCs using Markov
Random Fields, International Journal of New Computer
Architectures and their Applications (IJNCAA), The Society of
Digital Information and Wireless Communications (SDIWC) 3(1)
(ISSN: 2220-9085) (2013), 35-44.
39. Ertan Cetingul. H, Margaret J. Wright, Paul M. Thompson and Rene

Vidal, Segmentation of High Angular Resolution Diffusion MRI
Using Sparse Riemannian Manifold Clustering, IEEE Transactions
on Medical Imaging, Vol. 33, No. 2 (2014), 301 – 317.
40. Fan. L and Tian. D., A brain MR images segmentation method based
on SOM neural network, in: IEEE International Conference on
Bioinformatics and Biomedical Engineering, (2007).
41. Fletcher-Health LM, Hall LO, Goldgof DB and Murtagh FR,

Automatic segmentation of non-enhancing brain tumors in magnetic
resonance images, Artificial Intelligence, 21 (2001), 43-63.
42. Fogel. L.J, Zurada. J.M, Marks. R.J. II, and Goldberg. C,
Evolutionary programming in perspective: The top-down view, In
Computational intelligence: Imitating life, edited by J. M. Zurada, R.
J. Marks II, and C. J.Robinson, IEEE Press, Piscataway, NJ, (1994).
43. Forcade Nicolas, Le Guyader, Carole Gout, Christian (July),

Generalized fast marching method: applications to image
segmentation, Numerical Algorithms, 48 (1-3) (2008), 189–211,
<doi:10.1007/s11075-008-9183-x>.
44. Friedman J, Hastie T and Tibshirani R, Additive logistic regression:

a statistical view of boosting. Ann. Statistics, Medical Image
Analysis, 38 (2000), 337–407.
45. Govindaraj Vishnuvarthanan and Murugan Pallikonda Rajasekeran,

Segmentation of MR Brain Images for Tumor Extraction Using
Fuzzy, International Journal on Current Medical Imaging and
Reviews (CMIR), 9(1) (2013), 2-6.
184
46. Govindaraj, V. and Murugan, P. R., A complete automated

algorithm for segmentation of tissues and identification of tumor
region in T1, T2, and FLAIR brain images using optimization and
clustering techniques, Int. J. Imaging Syst. Technol, doi:
10.1002/ima.22108, 24 (2014), 313–325.
47. Guler, Demirhan. A and Karakis. R, Interpretation of MR images

using self-organizing maps and knowledge-based expert systems,
Digital Signal Processing, 19 (2009), 668–677.
48. Hanning Chen, Yunlong Zhu, and Kunyuan Hu, Adaptive Bacterial
Foraging Optimization, Abstract and Applied Analysis, Volume
2011, Article ID 108269, doi:10.1155/2011/108269.
49. Hanning Chen, Yunlong Zhu, and Kunyuan Hu, Cooperative

Bacterial Foraging Optimization, Discrete Dynamics in Nature and
Society, Volume 2009, Article ID 815247, doi:10.1155/ 2009/
815247.
50. Hassan Khotanlou and Mahlagha Afrasiabi, Segmentation of

Multiple Sclerosis Lesions in Brain MR Images using Spatially
Constrained Possibilistic Fuzzy C-means Classification, Journal of
Medical Signals & Sensors, Vol 1, issue 3 (2011).
51. Hassan Khotanlou, Olivier Colliot, Jamal Atif and Isabelle Bloch,
3D brain tumor segmentation in MRI using fuzzy classification,
symmetry analysis and spatially constrained deformable models,
Fuzzy Sets and Systems,160 (2009), 1457–1473.
52. Hayit Greenspan, Amit Ruf, and Jacob Goldberger, Constrained

Gaussian Mixture Model Framework for Automatic Segmentation of
MR Brain Images, IEEE Transactions on Medical Imaging, Vol. 25,
No. 9 (2006), 1233 – 1245.
53. Hollingworth W, Todd CJ, Bell MI, Arafat Q, Girling S, Karia KR

and Dixon AK, The diagnostic and therapeutic impact of MRI: an
observational multi-centre study, Clinical Radiology, 55 (11)
(2000), 825–31.
54. Hossein Mobahi, Shankar Rao, Allen Yang, Shankar Sastry and Yi
Ma, Segmentation of Natural Images by Texture and Boundary
Compression, International Journal of Computer Vision (IJCV), 95
(1) (2011), 86-98.
185
55. Hou. Z, Qian. W, Huang. S, Hu. Q and Nowinski. W.L, Regularized

Fuzzy C-Means Method for Brain Tissue Clustering, Pattern
Recognition Letters, 28 (2007), 1788–1794.
56. Indah Soesanti, Adhi Susanto, Thomas Sri Widodo and Maesadji
Tjokronagoro, MR brain images segmentation based on optimized
fuzzy logic and spatial information, International Journal of Video
& Image Processing and Network Security, IJVIPNS-IJENS, 11(4)
(2011).
57. Ipek Oguz and Milan Sonka, LOGISMOS-B: Layered Optimal

Graph Image Segmentation of Multiple Objects and Surfaces for the
Brain, IEEE Transactions on Medical Imaging, Vol. 33, No. 6
(2014), 1220 – 1235.
58. Islem Rekik, Stephanie Allassonniere, Olivier Clatz, Ezequiel

Geremia, Erin Stretton, Herve Delingette and Nicholas Ayache,
Tumor growth parameters estimation and source localization from a
unique time point: Application to low-grade gliomas, Computer
Vision and Image Understanding, 117 (2013), 238–249.
59. Jiang. Y, Chen. K and Zhou. Z, SOM based image segmentation, in:
Proceedings of the 9th International Conference on Rough Sets,
Fuzzy Sets, Data Mining, and Granular Computing, (2003).
60. Johnston. B, Atkins. M.S, Mackiewich. B and Anderson. M,

―Segmentation of multiple sclerosis lesions in intensity corrected
multispectral MRI‖, IEEE Transactions on Medical Imaging, 15
(1996) 154–169.
61. Juin-Der Lee, Hong-Ren Su, Philip E. Cheng, Michelle Liou, John
A. D. Aston, Arthur C. Tsai and Cheng-Yu Chen et al, MR Image
Segmentation Using a Power Transformation Approach, IEEE
Transactions on Medical Imaging, Vol. 28, No. 6 (2009).
62. Kannan. S. R, Sathya. A, Ramathilagam. S and Devi. R, Novel

segmentation algorithm in segmenting medical images, The Journal
of Systems and Software, 83 (2010), 2487–2495.
63. Kannan. S.R, Ramathilagam. S, Devia. R and Hines. E, Strong

Fuzzy C-Means in Medical Image Data Analysis, The Journal of
Systems and Software, 85 (2012), 2425– 2438.
186
64. Kapur. T, Grimson. W, Wells. I and Kikinis. R, Segmentation of

brain tissue from magnetic resonance images, Medical Image
Analysis, 1 (2) (1996), 109–127.
65. Karan Sikka, Nitesh Sinha, Pankaj K. Singh and Amit K. Mishra, A
fully automated algorithm under modified FCM framework for
improved brain MR image segmentation, Magnetic Resonance
Imaging, 27 (2009), 994–1004.
66. Karim Kalti and Mohamed Mahjoub, Image Segmentation by

Gaussian Mixture Models and Modified FCM Algorithm, The
International Arab Journal of Information Technology, Vol. 11, No.
1 (2014), 11 – 17.
67. Kass. M, Witkin. A and Terzopoulos. D, Snakes: Active contour

models, International journal of computer vision, 1 (4) (1988), 321-
331.
68. Keyvan Kasiri, Kamran Kazemi, Mohammad Javad Dehghani and

Mohammad Sadegh Helfroush, A Hybrid Hierarchical Approach for
Brain Tissue Segmentation by Combining Brain Atlas and Least
Square Support Vector Machine, Journal of Medical Signals &
Sensors, Vol: 3, Issue 4 (2013).
69. Khotanlou. H, Colliot. O, Atif. J and Bloch. I, 3D brain tumor

segmentation in MRI using fuzzy classification, symmetry analysis
and spatially constrained deformable models, fuzzy sets and systems,
(2009).
70. Kohonen, Teuvo, Self-Organized Formation of Topologically

Correct Feature Maps, Biological Cybernetics, 43 (1) (1982), 59–69.
71. Kohonen. T, Self-Organizing Maps, Springer, 2001.
72. Lashkari. A, A neural network based method for brain abnormality

detection in MR images using Gabor wavelets, International Journal
of Computer Applications, 4(7) (2010), 9-15.
73. Leemput. K. V, Maes. F, Vandermeulen. D, Colchester. A and

Suetens. P, Automated segmentation of multiple sclerosis lesions by
model outlier detection, IEEE Transactions on Medical Imaging, 20
(2001), 677–688.
187
74. Lim. Y. K and Lee. S. U, Color image segmentation algorithm based

on the thresholding and the fuzzy c-means techniques, Pattern
Recognition, 23 (1990), 935–952.
75. Lio. P.S, Chen. T.S and Chung. P, A fast algorithm for multilevel
thresholding, Journal of Information Science and Engineering, 17
(5) (2001), 713–727.
76. Lisa Jonasson, Xavier Bresson, Jean-Philippe Thiran, Van J.

Wedeen and Patric Hagmann, Representing Diffusion MRI in 5-D
Simplifies Regularization and Segmentation of White Matter Tracts,
IEEE Transactions on Medical Imaging, Vol. 26, No. 11 (2007),
1547 – 1554.
77. Liyuan G., Wanhua. C and Haixiang. X., Performance evaluation of

SVM in image segmentation, in: Proceedings of the ICSP 2008
International Conference, 2008.
78. Logeswari. T and Karnan. M, An improved implementation of brain

tumor detection using segmentation based on soft computing,
Journal of Cancer Research and Experiment Oncology (JCREO),
2(1) (2010), 06-14.
79. Logeswari. T and Karnan. M, An Improved Implementation of Brain

Tumor Detection Using Segmentation Based on Hierarchical Self
Organizing Map, International Journal of Computer Theory and
Engineering, Vol. 2, No. 4 (2010).
80. Mariano Rivera, Omar Ocegueda and Jose L. Marroquin, Entropy-

Controlled Quadratic Markov Measure Field Models for Efficient
Image Segmentation, IEEE Transactions on Image Processing, Vol.
16, No. 12 (2007), 3047 – 3057.
81. Marius George Linguraru, Nicholas Ayache, Eric Bardinet, Miguel

Angel Gonzalez Ballester, Damien Galanaud, Stephane Haïk,
Baptiste Faucheux, Jean-Jacques Hauw, Patrick Cozzone, Didier
Dormont and Jean-Philippe Brandel, Differentiation of sCJD and
vCJD Forms by Automated Analysis of Basal Ganglia Intensity
Distribution in Multisequence MRI of the Brain—Definition and
Evaluation of New MRI-Based Ratios, IEEE Transactions On
188
82. Marroquin. J. L, Vemuri. B. C, Botello. S, Calderon. F and

Fernandez- Bouzas, An accurate and efficient Bayesian method for
automatic segmentation of brain MRI, IEEE Transactions on
Medical Imaging, 21(8) (2002), 934–945.
83. Meiyan Huang, Wei Yang, Yao Wu, Jun Jiang, Wufan Chen and
Qianjin Feng, Brain Tumor Segmentation Based on Local
Independent Projection-based Classification, IEEE Transactions on
Biomedical Engineering, DOI: 10.1109/TBME.2014.2325410.
84. Meritxell Bach Cuadra, Claudio Pollo, Anton Bardera, Olivier

Cuisenaire, Jean-Guy Villemure and Jean-Philippe Thiran, Atlas-
Based Segmentation of Pathological MR Brain Images Using a
Model of Lesion Growth, IEEE transactions on medical imaging,
Vol. 23, No. 10 (2004).
85. Mirajkar G and Barbadekar B, Automatic segmentation of brain

tumors from MR images using undecimated wavelet transform and
Gabor wavelets, IEEE XPLORE (2010), 702-705.
86. Mishra. S and Bhende. C.N, Bacterial foraging technique-based

optimized active power filter for load compensation, IEEE
Transactions on Power Delivery, 22 (2007), 457–465.
87. Mishra. S, A hybrid least square-fuzzy bacteria foraging strategy for

harmonic estimation, IEEE Transactions on Evolutionary
Computation, 9 (1) (2005), 61–73.
88. Mohamad Forouzanfar, Nosratallah Forghani and Mohammad

Teshnehlab, Parameter optimization of improved fuzzy c-means
clustering algorithm for brain MR image segmentation, Engineering
Applications of Artificial Intelligence, 23(2010), 160–168.
89. Mohamed. N., Ahmed. M and Farag. A., Modified fuzzy c-mean in
medical image segmentation, in: IEEE International Conference on
Acoustics, Speech and Signal Processing, (1999).
90. Moumen T El-Melegy and Hashim M Mokhtar, Tumor segmentation

in brain MRI using a fuzzy approach with class center priors,
EURASIP Journal on Image and Video Processing, 21 (2014).
91. Murugavalli S and Rajamani.V, A high speed parallel Fuzzy C-

Mean algorithm for brain tumor segmentation, BIME Journal,
Volume (06), Issue (1), 2006.
189
92. Nan Zhang, Su Rua, Stéphane Lebonvallet, Qingmin Liao and

Yuemin Zhu, Kernel feature selection to fuse multi-spectral MRI
images for brain tumor segmentation, Computer Vision and Image
Understanding, 115 (2011) 256–269.
93. Nandha Gopal. N and Karnan. M, Diagnose Brain Tumor Through

MRI Using Image Processing Clustering Algorithms Such As Fuzzy
C Means Along With Intelligent Optimization Techniques, IEEE
Xplore, 978-1-4244-5967-4/2010, (2010).
94. Nandha Gopal. N, Automatic Detection of Brain Tumor through

Magnetic Resonance Image, International Journal of Advanced
Research in Computer and Communication Engineering, 2 (4)
(2013).
95. Ningning Zhou, Tingting Yang, and Shaobai Zhang, An Improved

FCM Medical Image Segmentation Algorithm Based on MMTD,
Computational and Mathematical Methods in Medicine, Volume
2014, Article ID 690349, http://dx.doi.org/10.1155/2014/690349.
96. Nock. R and Nielsen. F, Statistical Region Merging, IEEE

Transactions on Pattern Analysis and Machine Intelligence, Vol 26,
No 11 (2004), 1452-1458.
97. Ong. S, Yeo. N, Lee. K, Venkatesh. Y and Cao. D, Segmentation of

color images using a two-stage self-organizing network, Image and
Vision Computing, 20 (2002), 279–289.
98. Ortiz, Gorriz J. M, Ramirez. J and Salas-Gonzalez. D, Unsupervised

Neural Techniques Applied to MR Brain Image Segmentation,
Advances in Artificial Neural Systems, Article ID 457590, <
http://dx.doi.org/10.1155/2012/457590>, (2012).
99. Ortiz. J, Gorrizb J.M, Ramírezb. J, Salas-Gonzalezb. D and Llamas-

Elvirac. J.M, Two fully-unsupervised methods for MR brain image
segmentation using SOM-based strategies, Applied soft computing,
(2013), 2668–2682.
100. Osher. S and Sethian.J.A, Journal of Computational Physics, 79

(1988), 12-49.
190
101. Ouadfel Salima, Abdelmalik Taleb-Ahmed and Batouche Mohamed,

Spatial Information based Image Clustering with A Swarm
Approach, International Journal of Artificial Intelligence (IJ-AI),
Vol. 1, No. 3 (2012), 149 – 160.
102. Ozkan M, Dawant B.M and Maciunas R.J, Neural network based
segmentation of multi-modal medical images, a comparative and
prospective study, IEEE Transactions on Medical Imaging, 12(3)
(1993), 534-544.
103. Parra C., Iftekharuddin K. and Kozma R., Automated brain tumor
segmentation and pattern recognition using AAN, in: Proceedings of
the 2nd International Conference on Computational Intelligence,
Robotics and Autonomous Systems, (2003).
104. Passino. K. M, Biomimicry of bacterial foraging for distributed

optimization and control, IEEE Control Systems Magazine, 22 (3)
(2002), 52–67.
105. Pearson. K, Contributions to the Mathematical Theory of Evolution.

II. Skew Variation in Homogeneous Material, Philosophical
Transactions of the Royal Society A: Mathematical, Physical and
Engineering Sciences, 186, 343–414.
Bibcode:1895RSPTA.186..343P. doi:10.1098/rsta.1895.0010.
106. Petronella Anbeek, Koen L. Vincken, Glenda S. van Bochove,

Matthias J.P. van Osch, and Jeroen van der Grond, Probabilistic
segmentation of brain tissue in MR imaging, Journal on Neuro
Image, 27 (2005), 795 – 804.
107. Pham D.L, Spatial models for fuzzy clustering, Computer Vision and
Image Understanding, 84 (2001), 285–297.
108. Pham D.L., Unsupervised Tissue Classification in Medical Images

using Edge-Adaptive Clustering, Proceedings of the 25th Annual
International Conference of the IEEE EMBS, Cancun, Mexico,
(2003).
109. Prabhjot Juneja, Philp M. Evans and Emma J. Harris, The Validation
Index: A New Metric for Validation of Segmentation Algorithms
Using Two or More Expert Outlines With Application to
Radiotherapy Planning, IEEE Transactions on Medical Imaging,
Vol. 32, No. 8 (2013), 1481 – 1489.
191
110. Priya. R.K, Thangaraj. C, Kesavadas. C, and Kannan. S, Fuzzy

entropy based MR brain image segmentation using modified particle
swarm optimization, Int J Imag Syst Technol, 23 (2013), 281–288.
111. Rafael C. Gonzalez and Richard E. Woods, Digital Image

Processing. Second Ed.., Pearson Education publication, ISBN 81 –
7808 – 629 – 8, 68 – 70.
112. Raja Kumar. T.C, Arumuga Perumal. S and Krishnan. N, Fuzzy

based contrast stretching for medical image enhancement, ICTACT
journal on soft computing: special issue on fuzzy in industrial and
process automation, ISSN: 2229-6956(online), 02(1) (2012), 233-
236.
113. Rajalakshmi. N and Lakshmi Prabha. V, Brain Tumor Detection of

MR Images Based on Color-Converted Hybrid PSO+K-Means
Clustering Segmentation, European Journal of Scientific Research,
ISSN 1450-216X, Vol.70, No.1 (2012), 5-14.
114. Ramon Casanova, Joseph A. Maldjian and Mark A. Espeland,

Evaluating the Impact of Different Factors on Voxel-Based
Classification Methods of ADNI Structural MRI Brain Images,
International Journal of Biomedical Data Mining, Vol. 1 (2011).
115. Rasoul Khayati, Mansur Vafadusta, Farzad Towhidkhaha and S.

Massood Nabavi, Fully automatic segmentation of multiple sclerosis
lesions in brain MRFLAIR images using adaptive mixtures method
and Markov random field model, Computers in Biology and
Medicine, 38 (2008), 379 – 390.
116. Rechenberg. L, Zurada. J. M, Marks. R. J II and Goldberg. C

(Editors), Evolution strategy in computational intelligence: Imitating
life, IEEE Press, Piscataway, NJ, (1994).
117. Resmi A, Thomas T and Thomas B, A Novel Automatic Method For

Extraction of Glioma Tumour, White Matter and Grey Matter From
Brain Magnetic Resonance Images, Biomedical Imaging and
Intervention Journal, Biomedical Imaging and Intervention Journal,
9(2) 2013.
118. Riad, Atwan. A., El-Bakry. H., Mostafa. R., Elminir. H and
Mastorakis. N., A new approach for segmentation of MR brain
image, in: Proceedings of the WSEAS International Conference on
Environment, Medicine and Health Sciences, (2010).
192
119. Rudolf Mester and Uwe Franke, Statistical Model Based Image
segmentation Using Region Growing, Contour Relaxation and
Classification, Proceedings of SPIE Cambridge Symposium on
Visual Communications and Signal processing, Vol.88 (1988), 616 –
624.
120. Sahoo. P and Soltani. S, A. Wong, Y. Chen, A survey of

thresholding techniques, Computer Vision, Graphics, and Image
Processing, 41 (2) (2010), 233–260.
121. Salas-Gonzalez. D., Schlogl. M., Gorriz. J.M., Ramirez. J and Lang
E.W., Bayesian segmentation of magnetic resonance images using
the α-stable distribution, in: International Conference on Hybrid
Artificial Intelligence Systems, HAIS, (2011).
122. Sarah Parisot, William Wells III, Stéphane Chemouny, Hugues

Duffau and Nikos Paragios, Concurrent tumor segmentation and
registration with uncertainty-based sparse non-uniform graphs,
Medical Image Analysis, 18 (2014), 647–659.
123. Sathya. P.D and Kayalvizhi. R, Amended bacterial foraging

algorithm for multilevel thresholding of magnetic resonance brain
images, Measurement, 44 (2011), 1828–1848.
124. Serge Beucher and Fernand Meyer, The morphological approach to

segmentation: the watershed transformation, In Mathematical
Morphology in Image Processing (Ed. E. R. Dougherty), (1993),
433–481.
125. Shafaf Ibrahim, Noor Elaiza, Abdul Khalid and Mazani Manaf,
Image Mosaicing for Evaluation of MRI Brain Tissue Abnormalities
Segmentation Study, International Journal of Biology and
Biomedical engineering, Issue 4, Volume 5 (2011).
126. Shan. Zu. Y, Guang H. Yue and Jing Z. Liu, Automated histogram-
based brain segmentation in T1-weighted three-dimensional
magnetic resonance head images, Neuro image, 17 (3) (2002), 1587–
1598.
127. Shankar Rao., Hossein Mobahi., Allen Yang., Shankar Sastry and Yi
Ma., Natural Image Segmentation with Adaptive Texture and
Boundary Encoding, Proceedings of the Asian Conference on
Computer Vision (ACCV) (2009), H. Zha, R.-i. Taniguchi, and S.
Maybank (Eds.), Part I, LNCS 5994 (2009), 135-146.
193
128. Shantha Kumar. P and Ganesh Kumar. P, Performance Analysis of

Brain Tumor Diagnosis Based on Soft Computing Techniques,
American Journal of Applied Sciences, 11 (2) (2014), 329-336, <
doi:10.3844/ajassp.2014.329.336>.
129. Shapiro, Linda G. & Stockman, George C, Computer Vision,

Prentice Hall, ISBN 0-13-030796-3.
130. Shen S, Sandham W, Granat M and Sterr A, MRI fuzzy

segmentation of brain tissue using neighborhood attraction with
neural-network optimization, IEEE Trans. Inf. Technol. Biomedical,
9 (2005), 459–467.
131. Siyal. M.Y and Lin Yu, An Intelligent Modified Fuzzy C-Means
Based Algorithm For Bias Estimation and Segmentation of Brain
MRI, Pattern Recognition Letters, 26 (2005), 2052–2062.
132. Smith. S, Brady. M and Zhang. Y, Segmentation of brain images

through a hidden Markov random field model and the expectation-
maximization algorithm, IEEE Transactions on Medical Imaging, 20
(1) (2001).
133. Somasundaram. K and Kalaiselvi. T, Automatic brain extraction

methods for T1 magnetic resonance images using region labelling
and morphological operations, Computers in Biology and Medicine,
41 (2011), 716–725.
134. Somasundaram. K and Kalaiselvi. T, Fully automatic brain

extraction algorithm for axial T2-weighted magnetic resonance
images, Computers in Biology and Medicine, 40 (2010), 811–822.
135. Sultan Aljahdali and Zanaty. E. A, Improving Fuzzy Algorithms for

Automatic Image Segmentation, IEEE Xplore, 978-1-61284-732-
0/11, (2010).
136. Sultan Aljahdali and. Zanaty. E. A, Automatic Fuzzy Algorithms for

Reliable Image Segmentation, IJCA, Vol. 19, No. 3 (2012).
137. Swagatam Das, Ajith Abraham and AmitKonar, Particle Swarm

Optimization and Differential Evolution Algorithms: Technical
Analysis, Applications and Hybridization Perspectives, Springer-
Verlag Berlin Heidelberg, SCI – 116 (2008), 1–38.
194
138. Sun. W, Segmentation method of MRI using fuzzy Gaussian basis

neural network, Neural Information Processing, 8 (2) (2005), 19–24.
139. Tamije Selvy. P, Palanisamy. V and Sri Radhai. M, An Improved

MRI Brain Image Segmentation to Detect Cerebrospinal Fluid Level
Using Anisotropic Diffused Fuzzy C Means, WSEAS transactions on
computers, Issue 4, Volume 12, (2013).
140. Tamilarasi. M and Palanisamy. V, Medical Image Compression

Using Fuzzy C-Means Based Contourlet Transform, Journal of
Computer Science, ISSN 1549-3636, Science Publications, 7(9)
(2011), 1386-1392.
141. Thomas M Hsieh, Yi-Min Liu, Chun-Chih Liao, Furen Xiao, I-Jen
Chiang and Jau-Min Wong, ―Automatic segmentation of
meningioma from non-contrasted brain MRI integrating fuzzy
clustering and region growing, BMC Medical Informatics and
Decision Making, (2011), 11-54.
142. Thomas Schultz and Gordon L. Kindlmann, Open-Box Spectral

Clustering: Applications to Medical Image Analysis, IEEE
Transactions on Visualization and Computer Graphics, Vol. 19, No.
12 (2013), 2100 – 2108.
143. Thomos N, Boulgouris N. V and Strintzis, M. G, Optimized

Transmission of JPEG2000 Streams Over Wireless Channels, IEEE
Transactions on Image Processing, 15 (1) (2006).
144. Tolias. Y. A and Panas. S.M, On applying spatial constraints in

fuzzy image clustering using a fuzzy rule-based systems, The IEEE
Signal Processing Letters, 5 (10) (1998), 245–247.
145. Tripathy. M and Mishra. S, Bacterial foraging based solution to

optimize both real power and voltage stability limit, IEEE
Transactions on Power Systems, 22 (1) (2007), 240–248.
146. Tripathy. M, Mishra. S, Lai.L.L and Zhang.Q.P, Transmission loss

reduction based on FACTS and bacteria foraging algorithm, PPSN
(2006), 222–231.
147. Tsai. Y., Chiang. I., Lee. Y., Liao. C and Wang. K., Automatic MRI
meningioma segmentation using estimation maximization, in:
Proceedings of the 27th IEEE Engineering in Medicine and Biology
Annual Conference, (2005).
195
148. Tuan D. Pham and Klaus Berger, Automated Detection of White

Matter Changes in Elderly People Using Fuzzy, Geostatistical, and
Information Combining Models, IEEE Transactions on Information
Technology in Biomedicine, Vol. 15, No. 2 (2011).
149. Ulagammai. L, Venkatesh. P, Kannan. S. P and Padhy. N.P,

Application of bacteria foraging technique trained and artificial and
wavelet neural networks in load forecasting, Neuro computing, 70
(2007), 2659–2667.
150. Ulas Bagci et al, Joint segmentation of anatomical and functional

images: Applications in quantification of lesions from PET, PET-
CT, MRI-PET and MRI-PET-CT images, Medical Image Analysis,
17 (2013), 929–945.
151. Uma maheswari. J and Radhamani. G, A Hybrid Approach for

DICOM Image Segmentation Using Fuzzy Techniques,
International Journal of Fuzzy Logic Systems (IJFLS), Vol.2, No.3
(2012).
152. Vannier. M.W, Speidel. C.M, Rickman. D.L. Magnetic resonance

imaging multi spectral tissue classification, Journal of NIPS, 10(3)
(1991), 395-407.
153. Vapnik. V, Statistical Learning Theory, Wiley, New York, 1998.
154. Vasuda. P and Satheesh. S, Improved Fuzzy C-Means Algorithm for

MR Brain Image Segmentation, International Journal on Computer
Science and Engineering (IJCSE), Vol. 02, No. 05 (2010), 1713-
1715.
155. Vida Harati, Rasoul Khayati and Abdolreza Farzan, Fully automated
tumor segmentation based on improved fuzzy connectedness
algorithm in brain MR images, Computers in Biology and Medicine,
41(2011), 483–492.
156. Vijay Kumar. G, Biological early brain cancer detection using

artificial neural network, International Journal on Computer Science
and Engineering, 2(8) (2010), 2721-2725.
196
157. Vijayakumar Chinnadurai and Gharpure Damayanti

Chandrashekhar, Neuro – level set system based segmentation in
dynamic susceptibility contrast enhanced and diffusion weighted
magnetic resonance images, Pattern Recognition, 45 (2012), 3501–
3511.
158. Vijayakumar. B and Ashish Chaturvedi, Automatic Brain Tumors

Segmentation of MR Images using Fluid Vector Flow and Support
Vector Machine, Research Journal of Information Technology, 4(3)
(2012), 108-114.
159. Wei Sun and Yaonan Wang, Segmentation method of MRI using
fuzzy Gaussian basis neural network neural information processing,
Letters and Reviews, 8(2) (2005), 19-24.
160. Wells. W, Grimson. W, Kikinis. R and Jolesz. F, Adaptive

segmentation of MRI data, IEEE Transactions on Medical Imaging,
15 (4) (1996), 429–442.
161. Welstead and Stephen T Fractal, wavelet image compression

techniques, SPIE Publication, ISBN 978-0-8194-3503-3, 155–156.
162. Wen-Liang Hung, Miin-Shen Yang and De-Hua Chen, Parameter

Selection for Suppressed Fuzzy C-Means with an Application to
MRI Segmentation, Pattern Recognition Letters, 27 (2006), 424–
438.
163. William R. Crum, Oscar Camara, and Derek L. G. Hill, Generalized

Overlap Measures for Evaluation and Validation in Medical Image
Analysis, IEEE Transactions On Medical Imaging, Vol. 25, No. 11
(2006), 1451 – 1461.
164. Witkin. A. P., Scale-space filtering, Proc. 8th Int. Joint Conference
on Artificial Intelligence, Karlsruhe, Germany, (1983), 1019–1022.
165. Xiao Han and Bruce Fisch, Atlas Renormalization for Improved
Brain MR Image Segmentation across Scanner Platforms, IEEE
transactions on medical imaging, vol. 26, no. 4 (2007).
166. Xie. J and Tsui. H, Image segmentation based on maximum-

likelihood estimation and optimum entropy distribution (MLE-
OED), Pattern Recognition Letters, 25 (2005), 1133–1141.
197
167. Yan Li and Zheru Chi, MR Brain Image Segmentation Based on

Self-Organizing Map Network, International Journal of Information
Technology, Vol. 11, No. 8 (2005).
168. Yan Xia, Keith Bettinger, Lin Shen and Allan L. Reiss, Automatic
Segmentation of the Caudate Nucleus From Human Brain MR
Images, IEEE Transactions on Medical Imaging, Vol. 26, No. 4
(2007), 509 – 517.
169. Yang. Z and Laaksonen. J, Interactive Retrieval in Facial Image

Database Using Self- Organizing Maps, MVA, (2005).
170. Yaqub. M, Javaid. M. K, Cooper. C and Noble. J. A, Investigation of

the Role of Feature Selection and Weighted Voting in Random
Forests for 3-D Volumetric Segmentation, IEEE Transactions on
171. Yongxin Zhou and Jing Bai, Atlas-Based Fuzzy Connectedness

Segmentation and Intensity Non uniformity Correction Applied to
Brain MRI, IEEE Transactions on Biomedical Engineering, Vol. 54,
No. 1 (2007), 122 – 129.
172. Youyong Kong, Yue Deng and Qionghai Dai, Discriminative

Clustering and Feature Selection for Brain MRI Segmentation, IEEE
Signal Processing Letters, Vol. 22, No. 5 (2015), 573 – 577.
173. Yu J and Yang M.S, A generalized fuzzy clustering regularization

model with optimality tests and model complexity analysis, IEEE
Trans. Fuzzy System, 15 (5) (2007), 904–915.
174. Yu J and Yang M.S, Optimality test for generalized FCM and its
application to parameter selection, IEEE Trans. Fuzzy System, 13 (2)
(2005), 164–176.
175. Zahn. C.T, Graph-theoretical methods for detecting and describing

gestalt clusters, IEEE Transactions on Computers, Vol. 20, No. 1
(1971), 68–86.
176. Zhen Ma, Joao Manuel R.S. Tavares, Renato Natal Jorge and
Mascarenhas. T, A review of algorithms for medical image
segmentation and their applications to the female pelvic cavity,
Computer Methods in Biomechanics and Biomedical Engineering,
13 (2) (2010), 235–246.
198
177. Zhuowen Tu and Xiang Bai, Auto-Context and Its Application to

High-Level Vision Tasks and 3D Brain Image Segmentation, IEEE
Transactions on Pattern Analysis and Machine Intelligence, Vol. 32,
No. 10 (2010).
178. Zijdenbos A.P, Forghani R and Evans A.C, Automatic ‗pipeline‘

analysis of 3-D MRI data for clinical trials: application to multiple
sclerosis, IEEE Transactions on Medical Imaging, 20 (2001), 677–
688, 21 (2002), 1280–1291.
199
LIST OF PUBLICATIONS
International Journal
Published
1. Govindaraj Vishnuvarthanan and Murugan Pallikonda Rajasekeran,

―Segmentation of MR Brain Images for Tumor Extraction Using
Fuzzy‖, International Journal on Current Medical Imaging and
Reviews CMIR, 9(1), 2013. (Impact Factor: 0.708)
2. Govindaraj, V. and Murugan, P. R., ―A complete automated algorithm

for segmentation of tissues and identification of tumor region in T1,
T2, and FLAIR brain images using optimization and clustering
techniques‖, International Journal on Imaging Systems and
Technology, doi: 10.1002/ima.22108, Vol 24, pp.313–325, 2014.
(Impact Factor: 0.786)
Communicated
3. Govindaraj Vishnuvarthanan and Murugan Pallikonda Rajasekaran,

―A Novel Unsupervised Learning Method with Clustering Approach
for Tumor Identification and Tissue Segmentation in Magnetic
Resonance Brain Images‖, Applied Soft Computing, Elsevier.
4. Govindaraj Vishnuvarthanan and Murugan Pallikonda Rajasekaran,

―Tumor Detection in T1, T2, FLAIR and MPR Brain Images Using a
Combination of Optimization and Fuzzy Clustering Improved by Seed
Based Region Growing Algorithm‖, Journal of Medical Systems,
Springer.
International Conferences
5. Saravanan, N.J.; Vishnuvarthanan, G.; Pallikondarajasekaran, M.,

―An effective tree metrics graph cut algorithm for MR brain image
segmentation and tumor Identification‖, Advanced Communication
Control and Computing Technologies (ICACCCT), 2014 International
Conference on , vol., no., pp.1336,1340, 8-10 May 2014 (IEEE
Xplore).
200
CURRICULUM VITAE
Mr.G.VISHNUVARTHANAN was born in Madurai district,
Tamilnadu, India in the year 1986. He completed his Bachelor of Engineering
in Instrumentation and Control Engineering with first class in the year 2007
from Arulmigu Kalasalingam College of Engineering, Virudhunagar,
Tamilnadu and Master of Technology with distinction in the year 2009 from
Bharath University, Chennai, Tamilnadu.
He worked as a Lecturer in Electronics and Communication
Engineering department, Andal Alagar College of Engineering from May
2009 to June 2010. After that, he joined as Assistant Professor in Electrical
and Electronics Engineering department of Vels University, Pallavaram from
June 2010 to May 2011. Later, he continued his academic service as an
Assistant Professor in Electronics and Instrumentation Department of
Kalasalingam Univeristy, Virudhunagar since July 2011. He is currently
pursuing a Ph.D in Electronics and Communication Engineering department
at Kalasalingam University, India. His areas of interest are Medical image
processing and signal processing.

C 5 B 4

Hochgeladen von

Dokumentinformationen

Originalbeschreibung:

Originaltitel

Copyright

Verfügbare Formate

Dieses Dokument teilen

Dokument teilen oder einbetten

Freigabeoptionen

Stufen Sie dieses Dokument als nützlich ein?

Sind diese Inhalte unangemessen?

Copyright:

Verfügbare Formate

C 5 B 4

Hochgeladen von

Copyright:

Verfügbare Formate

TUMOR DETECTION AND TISSUE

SEGMENTATION IN MAGNETIC RESONANCE

in partial fulfillment for the award of the degree

DEPARTMENT OF INSTRUMENTATION AND

This is to certify that no corrections/suggestions were pointed out by

The following members of the Oral Examination Board were present:

l. Dr. M. Pallikonda Rajasekaran,

I hereby declare that the thesis entitled “TUMOR DETECTION AND

Signature of the Candidate

Certified that this thesis titled, “TUMOR DETECTION AND

TISSUE SEGMENTATION IN MAGNETIC RESONANCE BRAIN

IMAGES USING FUZZY AND OPTIMIZATION TECHNIQUES” is the

bonafide work of Mr. VISHNUVARTHANAN G, who carried out the

research under my supervision. Certified further, that to the best of my

dissertation on the basis of which a degree or award was conferred on an

earlier occasion on this or any other scholar.

Signature of the Supervisor

Image processing plays a very important role in analyzing the images

of different standards. Image enhancement, image restoration and image

compression are the types of image processing techniques of which image

segmentation process under Image enhancement technique is utilized

extensively to analyze the medical images.

An image is composed of several groups of pixels. Segmentation is

an image processing technique, in which different regions of an image are

procedures, of which, boundary based segmentation is used widely.

In the presence of several medical imaging techniques, Magnetic

Resonance Imaging (MRI) technique occupies a predominant position in the

area of diagnosing the organs of human body. The functionality of an MRI

frequency signals. The excited hydrogen atom is capable of emitting the

absorbed radio frequency signals, upon which several signal processing

examination and study are therefore required by a radiologist to identify the

pathologies embed within the tissue of brain. Normal Magnetic Resonance

(MR) scanner is capable of producing brain images with bounded tissues,

where unique and segregated views of the tissues are required. A

distinguished view upon the images is manually impossible and can be

subjected to operator - errors. Further, individual segmentation and

identification of tumor affected region in brain are quite unavailable in a

clinical MRI scanner. Due to the presence of afore mentioned hindrances,

automated algorithms capable of both tumor identification and tissue

segmentation processes are proposed through this research work.

With the support of soft computing techniques, automated and

unique segmentation upon the brain tissues, the identification of

heterogeneous type tumors can be effectively done. These functionalities

assist the radiologist in diagnosing process, extensively. Large volumes of

focuses upon the separation of tumor and edema regions.

My first and foremost thanks to the Almighty, for His endless

I am extremely thankful to the chairman, A.K. Group of

research with the patient information and MR brain images. I thank

I wish to express my sincere thanks to all the Teaching Staff

colossal support throughout my academic career.

CHAPTER NO. TITLE PAGE NO.

CHAPTER NO. TITLE PAGE NO.

2.2.5 Region – Growing Methods 15

3. DATA ACQUIREMENT FROM CLINIC AND

CHAPTER NO. TITLE PAGE NO.

3.3.4 MPR Type Images 48

4 SEGMENTATION OF MR BRAIN IMAGES

5. SEGMENTATION OF MR BRAIN IMAGES

CHAPTER NO. TITLE PAGE NO.

6. SEGMENTATION OF MR BRAIN IMAGES

7. SEGMENTATION OF MR BRAIN IMAGES

CHAPTER NO. TITLE PAGE NO.