robust medical image segmentation for accurate computer...

Mehdi Hassan

2015

Department of Computer and Information Sciences

Pakistan Institute of Engineering and Applied Sciences

Nilore, Islamabad, Pakistan

Robust Medical Image Segmentation for

Accurate Computer Aided Diagnosis

This page intentionally left blank.

Thesis Approval Form

Student’s Name: Mehdi Hassan Department: DCIS

Registration Number:03-7-1-027-2010 Date of Registration:26.01.2010

Thesis Title: Robust Medical Image Segmentation for Accurate Computer Aided Diagnosis

RECOMMENDATION (if any) by:

General comments (attach additional sheet if required)

When the final thesis defense of the student has been concluded and all other requirements have been

met, I

a. Do Recommend that the candidate be certified to the faculty for the degree of

Doctor of Philosophy

b. Do Recommend that the candidate be certified to the faculty for the degree of

Doctor of Philosophy subject to the minor correction in the thesis.

c. Do Recommend that the candidate should reappear in the oral defense

d. Do NOT Recommend that the candidate be certified to the faculty for the degree

of Doctor of Philosophy

Examiners Signatures

1. Internal Examiner (Name & Affiliation):



4. Supervisor (Name & Affiliation):

5. Head of Department (Name):

6. Dean (Research) :

Approved by:

Head of the Department (Name) :___________________________ Signatures/Date ______________

Distribution:

1. Original to be placed in Student’s Personal file in the office of Dean (Research)

2. Copy to be included in the thesis prior to final submission.

Thesis Submission Approval

This is to certify that the work contained in this thesis entitled Robust Medical

Image Segmentation for Accurate Computer Aided Diagnosis, was carried out by

Mehdi Hassan, and in my opinion, it is fully adequate, in scope and quality, for the

degree of Ph.D. Furthermore, it is hereby approved for submission for review and

thesis defense.

Supervisor: _____________________

Name: Dr. Asmatuallah Chaudhry

Date: 21 May, 2015

Place: Directorate of MIS, PAEC,

Islamabad.

Co-Supervisor: __________________

Name: Dr. Asifullah Khan

Date: 21 May, 2015

Place: PIEAS, Islamabad.

Head, DCIS: __________________

Name: Dr. Javid Khurshid

Date: 21 May, 2015

Place: PIEAS, Islamabad.

Robust Medical Image Segmentation for

Accurate Computer Aided Diagnosis

Mehdi Hassan

Submitted in partial fulfillment of the requirements

for the degree of Ph.D. May, 2015

Department of Computer and Information Sciences

Pakistan Institute of Engineering and Applied Sciences Nilore, Islamabad, Pakistan

ii

Dedications

Dedicated to my parents for their utmost love, prayers and

encouragement

&

My wife (Shamsa Batool) for all of her love and support

iii

Declaration of Originality

I hereby declare that the work contained in this thesis and the intellectual content of

this thesis are the product of my own work. This thesis has not been previously

published in any form nor does it contain any verbatim of the published resources

which could be treated as infringement of the international copyright law. I also

declare that I do understand the terms ‘copyright’ and ‘plagiarism,’ and that in case of

any copyright violation or plagiarism found in this work, I will be held fully

responsible of the consequences of any such violation.

__________________

(Mehdi Hassan)

21 May, 2015

PIEAS, Islamabad.

iv

Copyrights Statement

The entire contents of this thesis entitled Robust Medical Image Segmentation for

Accurate Computer Aided Diagnosis by Mehdi Hassan are an intellectual property

of Pakistan Institute of Engineering & Applied Sciences (PIEAS). No portion of the

thesis should be reproduced without obtaining explicit permission from PIEAS.

v

Acknowledgements

First of all, praise is due to Allah whose worth cannot be described by speakers,

whose bounties cannot be counted by calculators and whose claim (to obedience)

cannot be satisfied by those who attempt to do so, whom the height of intellectual

courage cannot appreciate, and the divings of understanding cannot reach; He for

whose description no limit has been laid down, no eulogy exists, no time is ordained

and no duration is fixed. He brought forth creation through His Omnipotence,

dispersed winds through His Compassion, and made firm the shaking earth with

rocks. Due to His showering of special blessings on me and honoring me with

strength and determination to accomplish PhD research work. He guided me during

every phase of my life as well as this PhD research work.

I am especially very grateful to Dr. Asmatullah Chaudhry and Dr. Asifullah

Khan for supervising me all heartedly and inspiring me for conducting and

completion of this research. I would also like to express my thanks to all beloved

teachers (especially to Dr. Abdul Majid, Dr. Abdul Jalil, and Dr. Ghulam Raza, who

give their precious time whenever I needed. I am very thankful to Dr. Nisar A.

Memon, QUEST, Nawabshah who motivated me to start my PhD studies at PIEAS.

I am grateful to my loving parents, siblings, cousins and other relatives, who

helped me and prayed for me to successfully accomplish PhD degree. I am also

thankful to all my friends and colleagues (especially Dr. Fakhar e Alam, Dr. Shozab

Mehdi, Khurram Jawad, Muhammad Tahir, Adnan Idris, Muhammad Aksam Iftikhar,

Iqbal Mirza, Syed Gibran Javed, Dr. Zia-ur-Rehman, Dr. Mattiullah Shah, Taskeen

Raza, Muhammad Shafiq, Safdar Ali) without them it would have been difficult for

me to stay and complete PhD research work at PIEAS.

Outside PIEAS, I am very thankful to Shakeel Ahmed, Tasuwar Hussain,

Aqeel Ahmed, and Raheel Ahmed who are always valuable throughout my life. I am

also thankful to my friends Haji Muhammad, Abdul Rasheed Safdar, Dr. Syed Hamad

Raza Naqvi, and Abdullah Aman Khan for their encouragement and support.

vi

I would like to extend my gratitude to all the anonymous reviewers of my

research papers. Without their valuable comments and suggestions it might not be

possible to complete my research on-time.

Beside these, how can I forget Pattern Recognition Lab, PIEAS where I have

spent days and nights to complete my PhD research? Pattern Recognition Lab

provided computational resources for experiments. Last but not least, I would like to

extend my thankfulness to Higher Education Commission (HEC) of Pakistan, for

providing financial support under Indigenous 5000 PhD program (PIN # 074-1729-

Ps4-078).

Mehdi Hassan

PIEAS, Islamabad.

vii

Table of Contents

Dedications ........................................................................................................................ ii

Declaration of Originality ................................................................................................ iii

Copyrights Statement ........................................................................................................iv

Acknowledgements ............................................................................................................ v

Table of Contents ............................................................................................................ vii

List of Figures....................................................................................................................xi

List of Tables .................................................................................................................... xv

Abstract ........................................................................................................................... xvi

List of Publications ...................................................................................................... xviii

List of Symbols ................................................................................................................ xx

List of Abbreviations .................................................................................................... xxii

Chapter 1 : Background and Goals of Study .................................................................... 1

1.1 Causes of Atherosclerosis................................................................................... 2

1.2 Symptoms of Cerebrovascular Accidents ......................................................... 2

1.3 Cerebrovascular Disease Diagnosis ................................................................... 3

1.4 Research Objectives and Contributions ............................................................. 5

1.5 Organization of Thesis ........................................................................................ 6

Chapter 2 : Literature Survey and Related Concepts ....................................................... 8

2.1 Segmentation of Carotid Artery Ultrasound Images ........................................ 8

2.2 Segmentation Algorithms ................................................................................. 11

2.2.1 Fuzzy C-means Clustering ........................................................................ 11

2.2.2 Spatial Fuzzy C-means Clustering ........................................................... 12

2.2.3 Radial Basis Function Network ................................................................ 13

2.2.4 Fuzzy Radial Basis Function Networks ................................................... 14

2.2.5 K-means Clustering ................................................................................... 15

2.2.6 Self-Organizing Maps (SOM) .................................................................. 15

2.3 IMT Measurements ........................................................................................... 15

2.4 Classification of Carotid Artery Ultrasound Images ...................................... 16

viii

2.5 Carotid Artery Ultrasound Images Datasets.................................................... 17

2.6 Performance Evaluation ................................................................................... 18

2.6.1 Clustering Performance Evaluation ......................................................... 18

2.6.2 Classification Performance Evaluation .................................................... 20

Chapter 3 : Carotid Artery Image Segmentation using Modified Spatial Fuzzy C-

means and Ensemble Clustering ..................................................................................... 22

3.1 The Proposed Spatial Fuzzy C-means Modified (sFCMM)........................... 22

3.1.1 Image Pre-processing ................................................................................ 23

3.1.2 Feature Extraction ..................................................................................... 23

3.1.3 Feature Selection ....................................................................................... 28

3.1.4 The Proposed Spatial Fuzzy C-means Modified Clustering Algorithm 29

3.1.5 The Proposed Ensemble Clustering Scheme ........................................... 30

3.1.6 Post-processing Operations ...................................................................... 31

3.2 Classification of Carotid Artery Ultrasound Images ...................................... 31

3.3 Experimental Results and Discussions ............................................................ 32

3.3.1 Performance Assessment of Proposed Technique on Phantom

Ultrasound Images .................................................................................................... 33

3.3.2 Performance Assessment of the Proposed Technique at Real Carotid

Artery Ultrasound Images ........................................................................................ 33

3.3.3 Performance Analysis of Ensemble Clustering based on Fuzzy C-means

34

3.3.4 Ensemble Clustering Assessment based on sFCM ................................. 37

3.3.5 Performance Analysis of sFCMM Based Ensemble Clustering............. 37

3.3.6 Clustering Performance Analysis of K-mean & SOM Based Ensemble ...

.................................................................................................................... 38

3.3.7 Classification of Carotid Artery Ultrasound Images ............................... 42

Chapter 4 : Robust Information Gain based Fuzzy C-Means Clustering and

Classification of Carotid Artery Ultrasound Images...................................................... 46

4.1 The Proposed Clustering Algorithm ................................................................ 46

4.1.1 The Proposed Information Gain Based FCM Clustering Algorithm ..... 46

4.1.2 Illustration of the Proposed IGFCM approach ........................................ 48

4.1.3 Clustering Quality Measures .................................................................... 51

4.2 Decision system for Segmented Carotid Artery Ultrasound Images ............. 51

4.2.1 Probabilistic Neural Networks Classifier................................................. 52


ix

4.3.1 Scenario-I: Segmentation Performance of the Proposed Algorithm ...... 54

4.3.2 Scenario-II: Segmentation and Decision Performance of IGFCM on

Carotid Artery Ultrasound Images .......................................................................... 61

Chapter 5 : Robust Segmentation of Carotid Artery Ultrasound Images based on

Neuro Fuzzy GA and Expectation Maximization .......................................................... 70

5.1 The Proposed RSC-US Technique................................................................... 70

5.1.1 Label Initialization of Pixels ..................................................................... 71

5.1.2 The Expectation Maximization Step ........................................................ 72

5.1.3 Feature Extraction ..................................................................................... 74

5.1.4 Neuro Fuzzy Classifier.............................................................................. 76

5.1 Decision Making System for Carotid Artery Ultrasound Images .................. 78

5.2.1 Classification Performance Measures ...................................................... 79


5.3.1 Performance Analysis of the Proposed Segmentation Technique.......... 83

5.3.2 SVM based Decision Making System ..................................................... 84

Chapter 6 : Robust Fuzzy RBF Network Based Segmentation and Decision

Making System for Carotid Artery Ultrasound Images ................................................. 90

6.1 The Proposed Robust Fuzzy RBF Network Approach ................................... 90

6.1.1 Targets Outputs ......................................................................................... 93

6.1.2 Input Features ............................................................................................ 94

6.1.3 Training of RFRBFN Clustering Approach ............................................ 95


6.2.1 Performance Comparison on Synthetic Image ........................................ 97

6.2.2 The proposed RFRBFN Segmentation Performance on Brain MRI ...... 99

6.2.3 Segmentation of Carotid Artery Ultrasound Images ............................. 105

6.2.4 Decision System for Carotid Artery Ultrasound Images ...................... 107

Chapter 7 : Automatic Active Contour Based Segmentation and Classification of

Carotid Artery Ultrasound Images ................................................................................112

7.1 The Proposed Approach ................................................................................. 112

7.1.1 Alignment of Carotid Artery Ultrasound Images .................................. 112

7.1.2 Snake Initialization .................................................................................. 114

7.1.3 Separation of Objects from Background................................................ 114

7.1.4 Segmentation of Carotid Artery Ultrasound Images ............................. 116

7.2 Classification of Carotid Artery Ultrasound Images .................................... 117

x

7.2.1 IMT Feature Extraction........................................................................... 118

7.2.2 Classification Performance Measurements............................................ 118

7.3 Experimental Results and Discussions .......................................................... 119

Chapter 8 : Conclusions and Future Directions............................................................126

8.1 Modified Spatial Fuzzy C-means and Ensemble Clustering ........................ 126

8.2 Robust Information Gain Based FCM Clustering ........................................ 127

8.3 Robust Segmentation of Carotid Artery Ultrasound Images using Neuro

Fuzzy and Expectation Maximization....................................................................... 127

8.4 Robust Fuzzy RBF Network Based Segmentation ....................................... 128

8.5 Automatic Active Contour Based Segmentation of Carotid Artery

Ultrasound Images...................................................................................................... 128

8.6 IMT Measurements and Medical Decision Systems .................................... 129

8.7 Future Recommendations ............................................................................... 129

References.......................................................................................................................131

xi

List of Figures

Figure 1-1 The ten most leading causes of death in United States [3] ........................... 2

Figure 1-2 (a) Carotid artery location in the head and neck, (b) the normal carotid

artery without any blockage, (c) the diseased carotid artery having plaque [1]. ............ 3

Figure 2-1 The sample carotid artery ultrasound image ................................................ 18

Figure 3-1 Block diagram of the proposed scheme ....................................................... 24

Figure 3-2 Experimental results: (a) the original phantom image; (b) FCM segmented

image; (c) segmented image using sFCM; (d) the proposed sFCMM segmented image

(e) segmented image using K-means (f) segmented image using SOM (g) proposed

ensemble scheme segmented image and (h) segmented image using sFCMLSM

technique. .......................................................................................................................... 34

Figure 3-3 Experimental results: (a) the original carotid artery image; (b) histogram

equalized image; (c) FCM segmented image; (d) segmented image by sFCM; (e) the

proposed sFCMM segmented image (f) K-means segmented image (g) SOM

segmented image (h) the proposed ensemble segmented image (i) morphological

corrected image, (j) sFCMLSM technique segmented image. ...................................... 35

Figure 3-4 (a) and (c) Carotid artery image is segmented using the proposed ensemble

clustering approach and (b) and (d) images segmented by sFCMLSM approach. ...... 40

Figure 3-5 Proposed approach segmentation of carotid artery ultrasound image using

(a) bilateral and (b) median filtering as pre-processing step for noise reduction......... 41

Figure 3-6 Performance measures versus increasing number of features. ................... 41

Figure 3-7 (a) Magnified IMT measurement section of an original carotid artery

ultrasound image (b) Magnified IMT measurement section of a segmented carotid

artery ultrasound image using the proposed scheme. .................................................... 41

Figure 3-8 (a) IMT measurements of a normal carotid artery image (b) IMT

measurements of abnormal carotid artery....................................................................... 42

Figure 3-9 ROC curve analysis of true and false positive rates using MLBPNN

classifications. .................................................................................................................. 42

Figure 4-1 The block diagram of the proposed IGFCM approach............................... 47

Figure 4-2 A sample pixel and corresponding local neighborhoods ........................... 50

Figure 4-3 Segmentation of noisy synthetic image (noise variance 0.01): (a) noisy

image segmented by (b) FCM, (c) sFCM, (d) sFCMM, (e) FLICM, and (f) the

proposed IGFCM algorithm. ........................................................................................... 55







xii

Figure 4-6 The Wolf image segmentation, (a) The Wolf image (original), (b) FCM

segmentation, (c) sFCM segmentation, (d) sFCMM segmentation, (e) FLICM

segmentation, and (f) the proposed IGFCM segmentation............................................ 56

Figure 4-7 Segmentation of noisy Wolf image (noise variance 0.01), (a) noisy Wolf

image, (b) FCM segmentation, (c) sFCM segmentation, (d) sFCMM segmentation,

(e) FLICM segmentation, and (f) the proposed IGFCM segmentation ....................... 57

Figure 4-8 CT liver image segmentation, (a) original CT liver image, (b) FCM


segmentation, and (f) the proposed IGFCM segmentation............................................ 60

Figure 4-9 Graphical comparison of clustering quality measures (a) PC and (b) CE for

the CT Liver image .......................................................................................................... 60

Figure 4-10 Carotid artery US image segmentation, (a) original carotid artery US

image, (b) FCM segmentation, (c) sFCM segmentation, (d) sFCMM

segmentation, (e) FLICM segmentation, and (f) the proposed IGFCM segmentation

........................................................................................................................................... 63

Figure 4-11 IGFCM segmentation at noise free and noisy carotid artery ultrasound

image corrupted through Gaussian noise of various intensities. ................................... 64

Figure 4-12 (a) PC and (b) CE measures of 300 carotid artery ultrasound images ..... 65

Figure 4-13 Average PC and CE measures for 300 carotid artery ultrasound images 65

Figure 4-14 IMT measurements of a normal and abnormal carotid artery................... 67

Figure 4-15 ROC curve plotted against true positive rate vs. false positive rate for

PNN, KNN, and MLBPNN classifiers. .......................................................................... 67

Figure 4-16 Effect of segmentation techniques on the classification of carotid artery

images ............................................................................................................................... 68

Figure 5-1 Graphical representation of the proposed RSC-US approach..................... 71

Figure 5-2 Initial targets marked as artery wall, area inside the artery and the

background tissues. .......................................................................................................... 72

Figure 5-3 GMM parameters estimation using EM algorithm ...................................... 75

Figure 5-4 Behavior of RMSE curve of NFC training at different epochs .................. 77

Figure 5-5 NFC generated FIS used for segmentation of carotid artery ultrasound

images. .............................................................................................................................. 78

Figure 5-6 Graphical representation of fuzzy membership functions for selected

feature 1 (a) and feature 2 (b) respectively. .................................................................... 79

Figure 5-7 a) One of the longitudinal original carotid artery ultrasound images with

marked plaque, b) selected ROI c) segmented by the proposed RSC-US scheme, d)

FCM segmentation, e) K-means and f) Magenta represents initial and green represents

final segmentation of sFCMLSM technique. ................................................................. 80

Figure 5-8 Sample noise free and noisy longitudinal carotid artery ultrasound images

segmented by the proposed RSC-US approach.............................................................. 82

Figure 5-9 (a) One of the original longitudinal carotid artery ultrasound images, b)

image corrupted by Gaussian noise of variance 0.10 with marked ROI c) the proposed

RSC-US technique segmented image; d) FCM segmentation e) K-means and f)

sFCMLSM segmentation. ................................................................................................ 82

xiii

Figure 5-10 Segmentation quality comparison of different techniques at various noise

levels ................................................................................................................................. 85

Figure 5-11 Computational time (in Sec.) of the proposed scheme across different

number of extracted features. .......................................................................................... 85

Figure 5-12 (a) IMT measurement of a normal and (b) abnormal carotid artery......... 86

Figure 5-13 Performance comparisons of KNN, MLBPNN and SVM classifiers using

RSC-US segmentation at various classification quality measures ................................ 87

Figure 5-14 ROC curves of SVM, MLBPNN and KNN. .............................................. 88

Figure 5-15 Comparison of classification accuracy for various segmentation

techniques ......................................................................................................................... 88

Figure 6-1 Schematic diagram of the proposed RFRBFN technique ........................... 93

Figure 6-2 (a) original synthetic image b) Noisy image (Gaussian noise intensity of

0.008) c) Image segmented by FCM d) RBF segmented image e) Fuzzy RBF

segmentation image and f) The proposed RFRBFN segmented image. ....................... 98

Figure 6-3 Misclassification error rE at various Gaussian noise levels ...................... 99

Figure 6-4 Cross validation error (in term of mean square error) of FCMJ vs ......... 99

Figure 6-5 RMSE of the proposed RFRBF network training for MR T1-weighted

image ............................................................................................................................... 101

Figure 6-6 (a) Original brain MR image (b) image with 0.01 Gaussian noise (c) FCM

segmentation (d) Image segmented by RBF technique (e) Fuzzy RBF segmented

image(f) The proposed RFRBFN segmented image. ................................................... 101

Figure 6-7 (a-c) Original brain MR images, (d-f) respective ground truths, (g-i)

RFRBFN based segmentation and (j-l) difference between ground truth and the image

by the proposed RFRBFN approach. ............................................................................ 102

Figure 6-8, Column (a) Original brain MR images, column (b) the noisy version of the

original brain MR images, column (c) the proposed RFRBFN approach segmented

images, column (d) the ground truth of the respective brain MR images. ................. 104

Figure 6-9 (a) Original carotid artery ultrasound image selected ROI (b) Noisy carotid

artery image of intensity 0.05 (salt & pepper) (c) FCM segmented image (d) image

segmented by RBF (e) FRBF network segmentation and (f) the proposed RFRBFN

segmented image. ........................................................................................................... 105

Figure 6-10 Left column: The original carotid artery ultrasound images with marked

plaque and right column: the proposed RFRBFN segmented images. ....................... 107

Figure 6-11 IMT measurements of normal and abnormal carotid arteries ................. 108

Figure 6-12 The ROC curves for different MLBPNN classifier (with different number

of hidden layer neuron) .................................................................................................. 110

Figure 7-1 Flow chart of the proposed approach ......................................................... 113

Figure 7-2 a) The original carotid artery ultrasound image, b) the cropped and median

filtered ROI, c) automatic snake initialization window, d) segmented carotid artery

ultrasound image using active contour method. ........................................................... 120

Figure 7-3 Column a) segmentation results using our proposed automatic snake

initialization approach and column b) images segmented by manual snake

initialization. ................................................................................................................... 121

xiv

Figure 7-4 IMT measurement of a normal carotid artery ultrasound image .............. 122

Figure 7-5 ROC curves of SVM, KNN and PNN classifiers showing true positive

(TP) and false positive (FP) rates at different thresholds ............................................ 124

Figure 7-6 Performance comparison of KNN, PNN and SVM classifiers at various

classification performance measures. ........................................................................... 124

xv

List of Tables

Table 3.1 Feature selected by WEKA using GA for segmentation of carotid artery

ultrasound images ............................................................................................................. 29

Table 3.2 Clustering performance measures of the proposed approach on phantom

ultrasound image. ............................................................................................................. 34

Table 3.3 The clustering quality performance comparison of various techniques....... 37

Table 3.4 Performance comparison of the proposed ensemble clustering approach

based on FCM, sFCM, sFCMM, K-means and SOM approaches. ............................... 39

Table 3.5 Performance comparison of the proposed approach segmentation of carotid

artery ultrasound images with other techniques by utilizing all thirty five extracted

features. ............................................................................................................................. 39

Table 3.6 The Performance comparison of segmentation using median and bilateral

filtering for image denoising. .......................................................................................... 40

Table 3.7 Classification performance measure of the MLBPNN ................................. 43

Table 3.8 The temporal comparison of the proposed approach at reduced and whole

extracted features set. ....................................................................................................... 43

Table 4.1 Clustering quality comparison of the proposed IGFCM and other

segmentation techniques .................................................................................................. 61

Table 4.2 Performance comparison of PNN-based decision system with other

classifiers .......................................................................................................................... 67

Table 4.3 The effect of segmentation technique at classification of carotid artery

ultrasound images ............................................................................................................. 68

Table 5.1 Average performance comparison of the proposed RSC-US and other

techniques over 300 carotid artery ultrasound images ................................................... 84

Table 5.2 Classification performance comparisons of SVM, KNN and MLBPNN

using the proposed RSC-US segmentation technique ................................................... 86

Table 5.3 Effects of segmentation on object classification ........................................... 88

Table 6.1 Parameters used for training of network for brain MR images .................. 100

Table 6.2 Parameters used for training of the network for carotid artery ultrasound

images ............................................................................................................................. 108

Table 6.3 The MLBPNN classification performance measures on 200 carotid artery

ultrasound images segmented by the proposed RFRBFN technique .......................... 109

Table 6.4 Effects of hidden layer neurons on classification accuracy of carotid artery

ultrasound images ........................................................................................................... 110

Table 7.1 Various IMT measurements of upper and lower wall of carotid artery in

term of mm ...................................................................................................................... 121

Table 7.2 Classification performance comparison of the various classifiers based on

proposed technique segmentation. ................................................................................ 122

xvi

Abstract

Image processing is being successfully applied in many areas medical research such

as computer aided diagnosis, tumor imaging and treatment, angiography, and carotid

artery plaque detection. For medical image analysis, segmentation is an intermediate

step to segregate region of interest from the background. The ultimate goal of

segmentation is to identify the part of the data array that makes up an object in the

real world. Many imaging modalities are in practice for disease diagnosis. Among

those, owing to noninvasive nature, ultrasound imaging provides an invaluable tool

for disease diagnosis. Major limitations faced by ultrasound imaging modality include

low quality, inherent noise, and wave interferences. Consequently, a substantial effort

from radiologists is required to extract constructive information about a particular

disease. In this regard, an efficient and accurate computer aided diagnostic system for

ultrasound images is highly desirable for disease (plaque) diagnosis.

Carotid arteries are vital arteries that supply oxygen rich blood to the brain.

Carotid artery stenosis is the process of narrowing the carotid artery due to the

presence of atherosclerosis. The plaque may partially or fully block the blood flow to

the brain and the probability of cerebrovascular stroke becomes high. Ultrasound

imaging is used for detection of plaque in carotid artery. Due to lower quality and

other degradations, segmentation of carotid arteries ultrasound images becomes a

challenging task.

In this thesis, several segmentation techniques are proposed, which

successfully segment the carotid artery ultrasound images. Firstly, we have proposed

spatial fuzzy c-means modified (sFCMM) clustering technique and also investigated

effectiveness of ensemble clustering. The proposed sFCMM technique assigns weight

to each pixel in a sub-window according to the pixel’s contribution. The proposed

scheme required image pre-processing for noise reduction and hence segmentation

has been performed on filtered image. In another approach, we propose information

gain based fuzzy c-means clustering (IGFCM) algorithm that avoids the pre-

processing step and still yields better results compared to sFCMM technique. The

IGFCM approach exploits the concept of information gain to automatically update the

xvii

fuzzy membership function and cluster centeriods. However, from IGFCM segmented

images, it has been observed that some of the pixels of arterial walls are mislabeled by

IGFCM. In order to overcome this problem, a semi-supervised clustering approach

named robust segmentation and classification of ultrasound images (RSC-US) has

been proposed to segment carotid artery ultrasound images.

The RSC-US approach is composed of three phases. In the first phase, the

fuzzy inference system (FIS) is generated. In second phase, carotid artery ultrasound

images are segmented based on the generated FIS. Finally, a decision making system

has been designed to segregate the segmented images into normal or abnormal

subjects. The RSC-US approach did not utilize the spatial information of pixel’s

which plays a vital role in segmentation. Consequently, the spatial information has

also been explored and a new approach named robust fuzzy radial basis function

networks (RFRBFN) has been proposed to segment carotid artery ultrasound images.

The RFRBFN segments the carotid artery ultrasound images with high precision. Due

to the Lagrange function and a smoothing parameter, the RFRBFN might be

computationally expensive. Finally, an automatic active contour based segmentation

technique for carotid artery ultrasound images is proposed. This technique can

successfully segment natural scene as well as medical images.

xviii

List of Publications

Journal Articles:

Mehdi Hassan, Asmatullah Chaudhry, Asifullah Khan, M. Aksam Iftikhar,

“Robust Information Gain Based Segmentation and Classification of Carotid

Artery Ultrasound Images”, “Computer Methods and Program in Biomedicine”

Vol. 113, pp. 593-609, 2013. Impact Factor 1.555

Mehdi Hassan, Asmatullah Chaudhry, Asifullah Khan, Jin Young Kim, “Carotid

Artery Ultrasound Image Segmentation using Modified Spatial Fuzzy c-means and

Ensemble Clustering”, “Computer Methods and Program in Biomedicine”, Vol

108, pp 1261-1276, 2012. Impact Factor 1.555

Asmatullah Chaudhry, Mehdi Hassan, Asifullah Khan, Jin Young Kim,

“Automatic Active Contour Based Segmentation and Classification of Carotid

Artery Ultrasound Images”, “Journal of Digital Imaging” Vol. 26, pp 1071-1081,

2013. Impact Factor 1.10

Asmatullah Chaudhry, Mehdi Hassan, Asifullah Khan, Jin Young Kim, Tran Anh

Tuan, “Automatic Carotid Artery Image Segmentation using Snake Based Model”,

“Journal of Korean Navigation Institute, Vol. 17, pp. 115-122, 2013”. Impact

Factor 0.278

Asmatulalh Chaudhry, Asifullah Khan, Anwar M. Mirza, Asad Ali, Mehdi

Hassan, Jin Young Kim, “Neuro Fuzzy and Punctual Kirging based Filter for

Image Restoration”, Applied Soft Computing, Vol. 13, 2012. Impact Factor

2.526.

Jan Alam, Mehdi Hassan, Asifullah Khan, Asmatullah Chaudhry, “Robust Fuzzy

RBF Network Based Segmentation and Intelligent Decision Making System for

Carotid Artery Ultrasound Images” “Neurocomputing” Vol. 151, pp. 745-755,

2015. Impact Factor 2.005

Asmatullah Chaudhry, Mehdi Hassan, Asifullah Khan, Robust Segmentation of

Carotid Artery Ultrasound Images using Neuro Fuzzy and Expectation

Maximization: Employing Intima-Media Thickness and SVM for Disease

Prediction, Submitted in Information Sciences, Journal, 2015.

Proceedings in National/International Conferences:


Tuan, Image clustering using Improved Spatial Fuzzy C-means. (ACM, ICUIMC,

12).

Mehdi Hassan, Asmatullah Chaudhry, Asifullah Khan, Kashif Riaz, An

Optimized Fuzzy C Means Clustering with Spatial Information for Carotid Artery

Image Segmentation. (IEEE, IBCAST, 2011).

xix

Mehdi Hassan, Asmatullah Chaudhry, Asifullah Khan, M. Aksam Iftikhar, Jin

Young Kim, Medical Image Segmentation Employing Information Gain and

Fuzzy C-means Algorithm, IEEE, ICOSST, Lahore, 2013.

Book Chapter:


Tuan, “Automatic Segmentation and Decision Making of Carotid Artery

Ultrasound Images”, Advances in Intelligent Systems and Computing, ISI

Indexed 2012.

xx

List of Symbols

nu nth order moment about mean

Standard deviation 2

Variance

ip z Probability of z for ith intensity value

iju Fuzzy membership function

iv Center of ith cluster.

kM kth classifier/cluster

|

k

iM

y c xP

Conditional probability of y on c value for x instance

ns The average distance of all objects from cluster centeriods.

Set of all possible points

x Observed/Pixel value

X Set of observed data points

Unknown log likelihood parameter need to be estimated

,L The log likelihood function

J Objective function

cW Weights between hidden and output node c

Smoothing parameter

, ,pL pM pHZ Z Z Low, medium and high linguistic property sets

p Z Input vector for low, medium and high linguistic property

sets

Normalization factor

pT Z Obtained target vector

rE Error rate

mcpN Number of misclassified pixels

Class probability separated by specific threshold

Morphological erosion

Morphological dilation

E v Energy function of snake

, ,M C K Mass, damping and stiffness matrices q External force

,s tv Element along snake contour

2

2

d

dt

u Second order snake function derivative

Observed value mean

xxi

EI Expected Information f b Morphological opening operation

f b Morphological closing operation Wavelet function

cS Wavelet scale parameter

HX Entropy of xp

HY Entropy of yp

IMC Information measure of correlation

IG Information gain

S Sigmoid function

cO Output node c response

Learning rate

W Weights change between hidden and output layers

The objective function

xp , xp Marginal probability along rows and columns

xxii

List of Abbreviations

TIA Transient Ischemic Attack

MRA Magnetic Resonance Angiogram

CTA Computed Tomography Angiogram

CAD Computer Aided Diagnostics

IMT Intima Media Thickness

FCM Fuzzy C-Means

sFCM Spatial Fuzzy C-Means

sFCMM Spatial Fuzzy C-Means Modified

MGH Moments of Gray Level Histogram

ROI Region of Interest

MLBPNN Multi-Layer Backpropagation Neural Networks

FFT Fast Fourier Transform

SVM Support Vector Machine

KNN K Nearest Neighbors

PSNR Peak Signal to Noise Ratio

GLCM Gray Level Co-occurrence Matrix

CWT Continuous Wavelet Transform

GA Genetic Algorithm

SOM Self-Organizing Maps

DBI Davies Bouldin Index

sFCMLSM Spatial Fuzzy C-Means Level Set Method

IGFCM Information Gain Based Fuzzy C-Means

PNN Probabilistic Neural Networks

ROC Receiver Operating Characteristic

FLICM Fuzzy Local Information C-Means

AUC Area Under the Curve

RSC-US Robust Segmentation and Classification of Ultrasound

Images

EM Expectation Maximization

GMM Gaussian Mixture Model

NFC Neuro Fuzzy Classifier

FIS Fuzzy Inference System

RFRBFN Robust Fuzzy Radial Basis Function Networks

1

Chapter 1 : Background and Goals of Study

Carotid arteries supply oxygenated blood to the brain. These vital arteries reside at

each side of the lower neck, below the jaw. The carotid arteries are further divided

into internal and external carotid arteries. The internal carotid artery supply blood to

the brain, where thinking, sensory, personality, speech and motor functions reside.

Whereas, the external carotid artery is used to supply blood to human face, neck and

scalp [1].

Carotid artery stenosis is the process of narrowing the artery due to the

presence of atherosclerosis. A plaque is built up in carotid artery because of the

atherosclerosis which is formed by the excess amount cholesterol and other fatty

materials. This plaque may partially or fully block the blood flow to the brain and the

probability of cerebrovascular accident (stroke) became very high [1, 2].

Cerebrovascular accident occurs when blood supply to the brain is stopped or

the blood vessel bursts. As, there is no mechanism in which brain can store oxygen,

hence the brain solely depend on the internal carotid artery which provides oxygen-

rich blood [2]. Cerebrovascular attack can occur due to the following reasons:

Due to the presence of plaque, carotid artery becomes narrow

When plaque ruptures, small pieces of plaque drift to the brain

Carotid artery becomes narrowed, due to the formation of blood clots

Figure 1-1 shows the ten most leading causes of deaths in United States.

Among those, cerebrovascular accident is the fourth leading cause of deaths in United

States. A total of 129476 deaths have been reported because of cerebrovascular

accidents in United States in year of 2010 only [3]. There are two common types of

cerebrovascular accidents (strokes), Ischemic and Hemorrhagic [4]. Ischemic stroke

mostly occurs in two ways: first, formation of clots somewhere in the body and

wedged into the blood vessel supplying oxygenated blood to the brain. If stroke

occurs in this way it is known as Embolic stroke. Secondly, a blockage or plaque

Chapter 1: Background and Goals of Study

2

forms directly into the carotid artery, if stroke occurs due to formation of plaque in

carotid artery, it is called Thrombotic accident.

We have considered the second case in our research to identify the plaque in

the carotid artery. Location and cross section of normal and abnormal carotid artery is

shown in Figure 1-2(a). Similarly, Figure 1-2(b) shows a normal carotid artery

whereas; plaque is formed and reduced blood flow to the brain is visible from Figure

1-2(c). The consequence of plaque may cause a serious brain stroke. An early

detection of plaque in carotid artery may prevent from such serious strokes.

1.1 Causes of Atherosclerosis

Arthrosclerosis is a disease in which plaque has buildup inside the arteries. The exact

causes of atherosclerosis are not known. Studies have revealed that the atherosclerosis

is a gradual process and it may starts from childhood. With the passage of time,

development of atherosclerosis became very fast. Common reasons of atherosclerosis

can be found in [2].

1.2 Symptoms of Cerebrovascular Accidents

Given below are some symptoms, which might be used as a warning for

cerebrovascular strokes. Among those, the most important warning sign is Transient

Ischemic Attack (TIA). TIA occurs when the blood vessels partially block the blood

flow to the brain. The symptoms of TIA can be found in [5].

Figure 1-1 The ten most leading causes of death in United States [3]


3

Figure 1-2 (a) Carotid artery location in the head and neck, (b) the normal carotid

artery without any blockage, (c) the diseased carotid artery having plaque [1].

1.3 Cerebrovascular Disease Diagnosis

Various modalities are available for the examination and cerebrovascular disease

diagnosis. The following are the techniques currently being used for cerebrovascular

disease diagnosis [1].

Blood Screening: Medical experts can examine different factors which may

cause the cerebrovascular stroke. Blood screening is required to check blood

sugar level, infection and blood clotting time of a patient.

Carotid Angiogram: In this technique, a dye is injected to the blood and then

X-ray of head is taken which helps experts to find the possible blockage of

carotid artery. In this way, a detailed view of the brain and neck arteries can be

obtained. The patients may feel allergic reaction using carotid angiogram

technique.

Computed Tomography Scan: A computed tomography (CT) scan is

performed if some type of cerebrovascular stroke or TIA has already occurred.

In this process, X-ray images of the body are acquired from different angles

and computer is used to combine the acquired images to view it on computer

screen in 2D or 3D formats. Using CT-scan imaging, the radiologist injects a

dye into blood vessel of the patient to highlight the carotid arteries in images.


4

The CT scan test may expose the affected areas of the brain. High radiation

risk factor is involved in CT scan technique.

Magnetic Resonance Angiogram (MRA): MRA is a procedure similar to

magnetic resonance imaging (MRI). Using MRA imaging, high frequency

magnets and radio waves are used to acquire the blood vessels images. MRA

is an effective method used for plaque detection in the carotid artery. Using

this imaging technique, information about carotid and other vertebral arteries

and presence of plaque in the carotid artery can be obtained. The MRA might

be harmful for those patients having kidney problems.

Computed Tomography Angiogram (CTA): CTA can be used to obtain

high resolution 3D images to analyze the carotid artery. In this procedure,

intravenous contrast material is injected to get high resolution images. Using

this imaging approach, the carotid arteries are examined that either the blood

supply to the brain is normal or not.

Magnetic Resonance Imaging: To assess plaque in carotid artery, MRI

approach can also be used. MRI uses magnetic field and radio waves to

acquire the carotid artery images.

Carotid Artery Ultrasound: It is one of the most common tests to assess the

carotid artery. In this procedure, high frequency sound waves are generated

and hence used to create inside image of carotid artery. It is a popular

technique because of its non-invasive nature. In this imaging technique,

neither patients are exposed to radiation nor do patients feel discomfort. With

the help of carotid artery ultrasound imaging, expert can examine if plaque

exists in carotid artery which may reduce the oxygenated blood flow to the

brain.

The above mentioned techniques are in practice for carotid artery disease

diagnosis. In this thesis, ultrasound imaging modality has been chosen for plaque

detection. There are certain reasons for selection of ultrasound imaging modality

which includes, its non-invasive nature, not exposed to radiation, comfortable to

patients, common and affordable to people [1, 2, 5].

On the other hand, ultrasound imaging has some limitations like low quality,

presence of noise and wave interferences etc. Due to these limitations, it is very


5

challenging to analyze and interpret the ultrasound images. It needs considerable

efforts from the radiologists to analyze these images for disease diagnosis. Firstly, as

the volume of medical images is growing day by day, manual interpretation and

analysis of these images in not feasible. Secondly, the variability of opinion across the

human experts is another major problem. Hence, a computer aided diagnostic (CAD)

technique is highly desirable to analyze the bulk of medical images. The objective of

CAD system is to provide additional support to the radiologists. Specifically, effective

computer algorithms are required to separate the region of interest (ROI) and the

structure of an organ automatically. These computer algorithms usually segment the

medical images into specified number of homogenous regions. Segmentation plays

vital role in image analysis such as anatomical structure [6], treatment planning [7],

and computer aided surgery [8].

1.4 Research Objectives and Contributions

There are two main objectives of this research for diagnosis of carotid artery plaque

using carotid artery ultrasound images. First and most important is segregation of

arterial walls of carotid artery from background tissues. This separation requires very

effective image segmentation algorithms which accurately segregate the arterial walls

from background tissues. Successful plaque identification requires accurate intima

media thickness (IMT) measurements from segmented carotid artery images. A great

care is needed to measure the IMT values because the plaque in carotid artery has to

be identified by the IMT measurement. Second objective of this thesis is to propose

an intelligent decision making system based on IMT values which classifies the

segmented images into normal and abnormal subjects.

To achieve these objectives, we have proposed several image segmentation

techniques named spatial fuzzy c-means modified and ensemble clustering,

information gain based fuzzy c-means, expectation maximization, neuro fuzzy,

genetic algorithm and automatic active contour based segmentation. Post processing

technique such as classification highly depends upon quality segmentation. The

proposed approaches outperformed the other state of the art segmentation techniques

like FCM, K-means, sFCM, SOM, FLICM, sFCMLSM.


6

In classification, we have employed multilayer backpropagation neural

networks (MLBPNN), support vector machine (SVM), K-nearest neighbors (KNN),

and probabilistic neural networks (PNN) to detect the plaque in carotid artery. High

classification accuracy has been achieved by employing these classification systems.

Further, the effect of segmentation at classification has also been investigated. It has

been observed that accurate segmentation has a high impact on classification stage.

The research contributions of the thesis in the field of medical image analysis and

disease diagnosis are as follows.

Carotid artery segmentation technique named spatial fuzzy c-means modified

(sFCMM) and ensemble clustering has been proposed. Intelligent decision

making system has also proposed to separate out the normal and abnormal

subjects.

Robust information gain based FCM segmentation approach is proposed for

carotid artery ultrasound image segmentation. The proposed technique

outperforms state of the art segmentation approaches.

Segmentation and classification approach is proposed based on EM, neuro

fuzzy, GA and SVM for disease prediction.

Robust Fuzzy RBF network segmentation and intelligent classification

technique for carotid artery ultrasound images is proposed.

Automatic active contour based segmentation approach for carotid artery

ultrasound images is proposed. The segmented images are then classified into

normal and abnormal subjects.

The above mentioned proposed techniques offer quality segmentation and

high classification accuracy on the obtained datasets. Detailed description of the

datasets is presented in Section 2.5.

1.5 Organization of Thesis

The rest of the thesis is organized as follows: Chapter 2 describes a comprehensive

literature survey and related theory of segmentation and classification of carotid artery

ultrasound images. Detail of the datasets used in this research work has also described

in Chapter 2. In Chapter 3, carotid artery segmentation using a modified spatial fuzzy


7

c-means and ensemble clustering approach is elaborated. In Chapter 4, information

gain fuzzy c-means clustering has been presented which is used for segmentation and

classification of carotid artery ultrasound images. In Chapter 5, a new robust approach

based on expectation maximization, neuro fuzzy classification and genetic algorithm

for segmentation is described. Chapter 6 presents a novel robust fuzzy radial basis

function network for segmentation and classification of carotid artery ultrasound

images approach. Whereas, in Chapter 7, a new automatic deformable model based

segmentation and classification of carotid artery ultrasound images is presented.

Finally, conclusions and future recommendations are provided in Chapter 8.

8

Chapter 2 : Literature Survey and Related

Concepts

In this chapter, we will present a detailed literature survey about the segmentation,

IMT measurements, and classification of medical images especially carotid artery

ultrasound images. The algorithms which have been used for segmentation of carotid

artery ultrasound images will be discussed with their pros and cons. Intima-media

thickness which is very important for the classification of carotid artery segmented

images will also be discussed. Finally, various classifiers used for identification of

plaque in carotid artery ultrasound images will also be discussed in this chapter.

2.1 Segmentation of Carotid Artery Ultrasound Images

Segmentation is the process of splitting an image into different homogenous regions.

Carotid artery ultrasound images based disease diagnosis is active area of research

and have attracted the attention of many researchers. Main task of segmentation is to

segregate the overlapping organs in medical images into non-overlapping regions. In

medical imaging, multiple tissues are contributing for the formation of a single pixel

so that they blur the boundaries of objects; hence this blurring is called a partial-

volume effect. Owing to the partial-volume effect, soft clustering is a good choice

because the clusters are allowed to overlap in a soft/fuzzy clustering. That is, one

pixel may belong to more than one class with a different degree of membership. In

fuzzy clustering, pixels have different memberships in multiple regions; thus fuzzy

clustering is more informative about original image [9].

Chuang et al. [10] have proposed spatial fuzzy c-means (sFCM) clustering

approach to segment medical images. The sFCM has utilized the spatial information

of the pixel under consideration and incorporates into conventional fuzzy c-means

(FCM). To make the model robust to noise, the objective function of the FCM has

been modified and spatial information has been incorporated into FCM objective

function. The sFCM approach assigns an equal weightage to all pixels of window and

it may produce inhomogeneous clustering. Chaudhry et al. [11] have proposed

Chapter 2: Literature Survey and Related Concepts

9

modification to sFCM for medical image segmentation named it spatial fuzzy c-

means modified (sFCMM). The Euclidean distance based weights have been assigned

to every pixel in sub-window according to pixel contribution. The advantage of

weight assignment is that it becomes robust and produces more homogenous

clustering as compare to sFCM.

Iscan et al. have proposed medical image segmentation technique using

incremental neural networks based on moments of gray level histogram (MGH) and

2D-continous wavelet transform [12]. It has used a high dimensional feature vector

and works only at the small noise levels. Vasantha et al. [13] have proposed

segmentation scheme in which most important features are selected by greedy

stepwise and genetic algorithm among a large set of hybrid features. Carotid artery

ultrasound image segmentation technique based on RANSAC and cubic splines has

been reported in [14]. The technique is computationally expensive and only 50

brightness mode (B-mode) ultrasound images used for evaluation.

Spatial fuzzy clustering and level set methods based medical image

segmentation is reported by Li et al. [15]. They have incorporated partial differential

equations into spatial fuzzy c-means became computationally expensive. Yu et al. [16]

have reported a medical image segmentation technique based on object density. The

marker controlled watershed segmentation approach has been utilized to the object of

interest which is used to identify the region of interest (ROI). Improper selection of

ROI leads to misleading results and over segmentation is one of the major drawbacks

of watershed segmentation.

Mao et al. [17] have used deformable model for extraction of carotid artery

lumen. The images using deformable models are segmented with high accuracy. The

snake based models have been reported in literature to extract the contour of carotid

artery using ultrasound images [18, 19]. However, the snake based/deformable

models have a major limitation in which user intervention is required for snake

initialization. A new Doppler based scheme is proposed by Hovda et al. [20] which

has been used to segment the image by utilizing the blood or tissue characteristics.

Likelihood ratio function was applied for classification however, very difficult to

validate the assumptions of the model and thus the system becomes more complex.


10

Canny edge detector based technique has been proposed by Hamou et al. [21]

for segmentation of carotid artery ultrasound images. The proposed technique is based

on three parameters, upper and lower boundary thresholds, and standard deviation of

Gaussian smoothing kernel. This method is used to wipe out the insignificant details

from the map generated by their technique. Similarly, carotid artery contour

extraction technique is reported by Adel-Dayem et al. [22]. The carotid artery image

pixels have been segmented into three classes, named the arterial wall, an area inside

the artery and background tissues, by employing uniform quantizer. Edges among

these three clusters have been extracted using the morphological edge detector. For

minimizing the effect of noise, they have employed some pre and post-processing

techniques as well. However, the basic limitation of the technique is its low sensitivity

to noise and it may not differentiate the small variation in intensity especially within

the arterial walls.

Segmentation of carotid artery ultrasound images based on watershed

segmentation is reported in [23, 24]. Threshold selection is crucial to region merging

stage which is based on the average of neighboring area pixels. Kamel et al. [25] have

proposed a technique using fuzzy region growing for segmentation of carotid artery

ultrasound images. The technique generates a fuzzy connectedness map of image

which is computationally expensive.

An integrated approach based on snake based model for segmentation of

carotid artery ultrasound images have been proposed by Loizou et al. [26]. Bastida-

Jumilla et al. [27] have reported the common carotid artery segmentation scheme

using frequency implementation for deformable models. Improper initialization by the

inexperienced user may lead toward false results. Chaudhry et al. [28] have proposed

an automatic active contour based segmentation of carotid artery ultrasound images.

The proposed approach overcomes the limitation of manual initialization of snakes

and thus their approach is fully automatic.

Golemati et al. [29] have proposed Hough transform based segmentation of

carotid artery ultrasound images. Hough transform based techniques are useful to

detect lines and circles, however the carotid artery vessels are curvy and Hough

transform based techniques may not accurately segment such images. Moreover the


11

authors tested their approach at 10 B-mode ultrasound images and have performed a

small scale evaluation of stenosis level.

Abdel Dayem and El Sakka [30] have proposed an integrated segmentation

approach which is based on multi-resolution and watershed techniques. Their

objective was to speedup segmentation process. The scheme has been employed to

decompose an image into a pyramid by applying wavelet transform. The proposed

scheme is suitable for noise free images but the noisy images might not be well

segmented by their approach. Dynamic programming based carotid artery ultrasound

image segmentation techniques have been reported in [31, 32]. The dynamic

programming performed well at noise free images, but the computational cost of

dynamic programming is significantly high.

The basic objective of the above mentioned techniques is to segment the

carotid artery ultrasound images in an efficient way. However, every technique has its

own limitations and thus it becomes challenging to segment carotid artery ultrasound

images with high precision.

2.2 Segmentation Algorithms

Detail of the segmentation/clustering algorithms used in our experiments are given

below. Some of our proposed approaches are extension of these algorithms. These

techniques are being successfully used to segment the medical images especially,

carotid artery ultrasound images. Additionally, these state of the art algorithms are

also used to compare the results of the proposed scheme. Detail of each algorithm is

as under:

2.2.1 Fuzzy C-means Clustering

Fuzzy c-means clustering (FCM) [33, 34] is an unsupervised technique that has been

successfully applied to clustering, features analysis and classifier designing in fields

such as image segmentation, medical imaging, astronomy and geology [35]. Similar

to other clustering algorithms, FCM is also used to group similar data points into a

same cluster. The clustering is achieved iteratively, minimizing the cost function, to

decide the fate of pixel to which cluster of pixel may belong.


12

FCM may be one of the candidate techniques for clustering in areas where the

objects regions are overlapped. Image pixels are highly correlated and probability is

high that one pixel may belong to more than one clusters with different degree of

memberships. In such a case, fuzzy clustering algorithm like FCM might be a better

choice for segmentation.

The FCM algorithm assigns pixels to clusters based on their fuzzy membership

values. It strives to minimize the following cost function:

2

1 1

N Cm

ij j i

j i

J u x v

(2-1)

where iju shows the membership of pixel jx to ith cluster jx and represents

the set of points that an image is composed. C and N represent total number of

clusters, and iv is centroid of ith cluster. The constant m is also known as degree of

fuzziness and is usually set to 2 for most applications.

The cost function of FCM is minimized iteratively by updating the cluster

centroid. Fuzzy membership values are assigned to pixels based on their distance

from the center of clusters. The smaller the distance of pixel under consideration from

cluster centroid, higher will be the degree of membership to that cluster and vice

versa. The following two mathematical expressions are used to update the fuzzy

membership functions and cluster centers, respectively [36]:

2

1

1

1ij

mCj i

k j k

u

x v

x v

(2-2)

1

1

N

ij j

j

i Nm

ij

j

u x

v

u

(2-3)

2.2.2 Spatial Fuzzy C-means Clustering

In an image, due to high pixel correlation, the probability of pixels belonging to the

same cluster becomes high. The effect of neighboring pixels has not been utilized in


13

conventional FCM clustering. Using FCM algorithm, a noisy pixel may mislabeled

because of the abnormal feature data. To overcome this problem, Chuang et al. [10]

proposed modification in the basic FCM framework and named it spatial fuzzy c-

means clustering (sFCM) algorithm. In sFCM, the spatial information of pixel is

utilized and an equal weight is assigned to each pixel in sub-window to decide the

pixel’s fate. The advantage of incorporation of spatial information is that the sFCM

algorithm becomes more robust to noise and produces a more homogenous clustering

as compared to FCM algorithm.

The sFCM incorporates spatial information into the basic FCM framework for

clustering.

j

ij ik

k NB x

h u

(2-4)

where jNB x shows a square window centered on jx . Like the membership function

in FCM, ijh shows the probability of pixel belonging to ith cluster. Large spatial

function value shows that majority of the pixels belong to same cluster. Spatial

function has been incorporated into the FCM membership function and new

membership function will become as follows:

'

1

p qij ij

ij cp q

kj kj

k

u hu

u h

(2-5)

2.2.3 Radial Basis Function Network

Unlike the simple neural network, radial basis function has been used in the hidden

units of network as described in [37]. The functions used in the network are strictly

positive and radially symmetric with a unique maxima at its center. The hidden layer

weights were optimized using K-means algorithm [38].

2

1

N

kN p k

p k

J z v

(2-6)

where represents a set of input data points, N is the total number of hidden units,

and kv corresponds to cluster centers of hidden units. Here, we need to find kv that


14

minimizes Equation 2-6. If we use Gaussian as a base function, the response of the

hidden units can be calculated using the following expression:

22

exp

p k

k

z v

k

(2-7)

where 1 2, ,...,p nz z z z corresponds to input data samples. Response of each output

node is simply a linear combination of kernel functions:

c c j

j k

O W

(2-8)

where cO is the response output of node c and cW represents the weights between the

hidden layers and the output of node c .

2.2.4 Fuzzy Radial Basis Function Networks

Fuzzy Radial Basis Function Network [38] combines FCM and RBF approaches.

Weights of hidden layers (cluster centriod kv ) are updated using FCM, instead of K-

means clustering. In fuzzy RBF network, fuzziness has been incorporated at input and

output layers. To compute pku locally, a modified RBF network was used [38]. Using

this modified architecture, the output of each hidden layer is calculated by the

Equation 2-9.

2

1

1

q

p

k

p k

p

z v

(2-9)

In FRBF, weights between hidden and output layers are optimized using

gradient descent algorithm. Response of each output node is calculated using the

following expression:

c ck kp

k N

O W u

(2-10)


15

2.2.5 K-means Clustering

K-means clustering is an unsupervised technique of clustering. The K-means is a

commonly used algorithm in computer vision for image segmentation [39] . It is an

iterative approach to update cluster center or mean to classify the whole data.

2.2.6 Self-Organizing Maps (SOM)

Self-organizing maps (SOM) is learning algorithm that produces low dimension input

space for training samples. Kohonen has developed the technique with self-

organizing for a network of adaptive elements [40]. The SOM algorithm follows two

basic principles; matching and finding that determine the winner neuron by the

minimum Euclidean distance to the input and update the position of neurons inside

the cluster.

2.3 IMT Measurements

IMT measured values are used to identify the presence of plaque in carotid artery [41,

42]. High IMT value is associated with high risk of cerebrovascular stroke. Keeping

the importance of IMT measurements, it is essential that IMT should be measured

accurately. For accurate IMT measurements, it is necessary that the arterial walls

should be separated accurately from background tissues.

IMT is the width including lumen-intima and the media adventitia [43]. As the

atherosclerosis and other fatty materials narrow the carotid artery, hence due to the

presence of plaque blood supply to the brain is reduced. IMT is one of the important

measures being successfully used for identification of plaque in the carotid artery. The

atherosclerotic changes are reflected by IMT measurements and thus can be used to

predict the cerebrovascular accident (stroke). The certain factors should be considered

while measuring IMT, i.e. age, gender, body to mass index (BMI), high blood

pressure and diabetic patients [44-50]. Classification of the segmented carotid artery

is vital because patient’s rehabilitation process may start. The decision system for

carotid artery plaque detection can be formed based on the IMT values.


16

2.4 Classification of Carotid Artery Ultrasound Images

The main objective of the classification is to divide an input dataset into various

categories. The input data need to be classified on the basis of classifier’s training. In

our case, the carotid artery ultrasound images segmented by the proposed techniques

need to be classified into normal or abnormal subjects. A number of classifiers (linear

and non-linear) have been reported in the literature to classify the input data into

specified number of classes. The classification phase results are highly dependent

upon the accurate IMT measurements. It is possible only when the carotid artery

ultrasound images are segmented correctly.

Texture based classification of carotid artery plaque is proposed by

Christodoulou et al. [51]. A set of statistical and gray level dependence feature

vectors has been formed to classify the carotid artery plaque. The technique might be

computationally expensive because of large feature vector dimensions. K-nearest

neighbors (KNN) based classification of carotid artery plaque is reported in [52]. A

large dimension of feature set including shape, texture, statistical and morpholog have

been extracted and given as an input to KNN classifier. Similarly, KNN based

classification of carotid artery plaque have been investigated in [26]. A robust

decision system based on KNN for identification of plaque in carotid artery has been

proposed by Latifoglu et al. [53]. The fast Fourier transform (FFT) based features are

extracted and important features are selected using principal component analysis

(PCA). Similarly, spatial gray level dependence matrices features were extracted and

consequently used for plaque classification employing SVM classifier [54].

Multilayer back-propagation neural networks (MLBPNN) based classification

of carotid artery plaque has been reported in [32]. The probability of MLBPNN

approach to trap into a local minima is high. Support vector machine (SVM) and

probabilistic neural network based classification of atherosclerosis tissues have been

reported in [55]. The authors have achieved the classification accuracy of upto 90%.

The post-processing techniques (classification in this case) highly depend

upon quality segmentation. The efficiency of decision system can be examined by

high classification accuracy. It is essential for computer aided diagnostic (CAD)

system to distinguish the abnormal subjects from normal ones.


17

2.5 Carotid Artery Ultrasound Images Datasets

It is very important for any segmentation algorithm to test its worth at a significant

amount of dataset. To the best of our knowledge, we could not find any standard

dataset of carotid artery ultrasound images. We have formed a real carotid artery

ultrasound images dataset and made collaboration with Radiology department, Shifa

International Hospital Islamabad, Pakistan. We have frequently visited and have

arranges meetings with the medical experts to understand the basic phenomenon and

problems are faced by experts for identification of carotid artery plaque. After a

careful study, to coup with those problems, we have proposed several new algorithms

for solution of those problems. The detail of the datasets used in the whole study is

given below.

In Shifa International Hospital Islamabad, Toshiba Xario XG Ultrasound

machine is being used for carotid artery examination. The machine is equipped with

Linear Probe Transducer and the frequency of range 7-8 MHz has employed for

carotid artery imaging. The carotid artery data was originally recorded videos taken

for 10 seconds. These videos have been then converted into frames using video

Decompiler. The original images are the size of 800x600 pixels with a resolution of

72 pixels per inch (PPI). The images were cropped from borders to separate area of

interest and converted into gray scale before further processing has been applied. The

size of cropped images reduced to 350x380 pixels.

The whole dataset consists of 350 carotid artery ultrasound images. A total of

57 patients’ data (including males and females) was obtained. The age of the patients

was in the range of 35-76 years. Overall mean and standard deviation of the patients’

age is 61.75 and 7.67 years respectively.

With the help of medical experts, images in the dataset were labeled as normal

or abnormal subjects. The benefit labeling the images was to test the classification

performance based on the proposed segmentation algorithms. Training of classifiers

needs class labels for learning the patterns effectively and based on the learned

patterns, it might be able to classify the test images with high accuracy. The dataset is

obtained step by step depends upon the arrival of patients to the hospital for carotid

artery examination. This is why different number of images has been used for


18

evaluation of each proposed segmentation algorithm. A sample longitudinal carotid

artery ultrasound image is shown in Figure 2-1.

2.6 Performance Evaluation

The first objective of the thesis is to propose robust and accurate segmentation

algorithm and the second one is to design an intelligent decision making system for

segregation of normal and abnormal subjects. It is mandatory to evaluate the

performance of the proposed segmentation and classification technique. Clustering

and classification have their own performance evaluation measures, firstly, the

clustering quality measures used to evaluate the performance of proposed approaches

have been described. Secondly, classification/decision making performance measures

have been described.

Figure 2-1 The sample carotid artery ultrasound image

2.6.1 Clustering Performance Evaluation

Davies Bouldin Index (DBI)

DBI is frequently used clustering quality measure. It has advantages such as easiness

of producing results, robustness to high class specific noise, and lower computational

requirements [56, 57]. DBI is the ratio of sum of within cluster scatter to between

cluster separations. It can be calculated by the following expression:


19

1

( ) ( )1DBI max

( , )

nn i n j

i i j

Q s Q

n s Q Q

s

(2-11)

where n represents number of clusters, ns is an average distance of all objects from

cluster to the cluster centroid and ,i js Q Q is the distance between cluster centers.

The ratio will be small, if clusters are compact and far from each other. It means that

the smaller the DBI is the better the clustering quality is [58, 59].

Partition Coefficient (PC)

The PC index proposed by Bezdek [60] is a fuzzy clustering quality measure and used

to measure the amount of overlap between clusters. PC index is used to evaluate the

presence of fuzziness in partitions/clusters. Maximal value of PC represents the best

possible clustering [59]. It can be calculated by the following expression:

2

1 1PC

n c

ijj i

u

n

(2-12)

where iju shows the fuzzy membership of data point j to cluster i .

Classification Entropy (CE)

Similar to PC measure, CE is another measure to evaluate the fuzziness of

partitions/clusters. Minimal value of CE shows better clustering [59, 61]. Classification

entropy can be computed as:

21 1

1CE log ( )

c n

ij iji j

u un

(2-13)

where iju is the fuzzy membership of data point j belonging to cluster i .

To check the clustering quality the above mentioned quality measures have been

calculated for each of the segmented carotid artery ultrasound images.


20

2.6.2 Classification Performance Evaluation

In statistical prediction, to check the effectiveness of the classifier, different types of

cross validation techniques are in practice. Among those, Jackknife technique is a

popular one. It gives unique results for a given dataset. It is being used by the analysts

to validate the accuracy of prediction. In our work, we have used Jackknife validation

technique to examine the quality of classifier. According to Jackknife test, N-1 data

samples are used for training and one data sample for testing. The class of the test

pattern is predicted by the classifier based on the N-1 training data samples. The

sampling process is repeated for N times and the class of each test sample is

predicted. The true positive (TP) and true negative (TN) are the number of correctly

classified positive and negative test samples. False positive (FP) and false negative

(FN) numbers represent those images whether are they incorrectly classified or not,

respectively. Following measures are used to evaluate the performance of a classifier

[59].

Accuracy

The accuracy measure is used to check the overall usefulness of the classifier. It can

be expressed by the following equation.

Accuracy 100TP TN

TP FP TN FN

(2-14)

Sensitivity

This measure is used to check if, either the classifier is able to predict the positive

class patterns. Sensitivity of a classifier can be calculated by the following equation.

SensitivityTP

TP FN

(2-15)

Specificity

The specificity is used to validate how much ability of a classifier has to recognize the

negative class patterns. Specificity can be calculated by the following expression.

SpecificityTN

TN FP

(2-16)


21

Mathew Correlation Co-efficient (MCC)

MCC is used to measure the classifier performance for binary classes having the value

range from -1 to 1. MCC value +1 means perfect classification having no mistake, zero

MCC value means a random prediction and -1 MCC value means that classifier never

predict a correct label. MCC measure can be obtained using the following expression:

MCC

)

TP TN FP FN

TP FN TP FP TN FN TN FP

(2-17)

F-Score

The F-score considers both precision and recall to measure the validation of a

classifier. Its value ranges from 0 to 1 and is a weighted average of the precision and

recall. Value 0 and 1 shows worst and best scores, respectively. The F-score is

calculated by the following formula:

precision recallF.Score 2

precision recall

Precision= ,Recall=TP TP

TP FP TP FN

(2-18)

Negative Predicted Value (NPV)

The NPV measure is used for evaluation of diagnostic performance. High value of

NPV indicates high probability that a patient is free from the disease against which

the test was conducted. NPV is calculated by the following equation [62]:

NPV 100TN

FN TN

(2-19)

ROC Curve

ROC curve is a graphical representation of true positive vs. false positive rates. It is

one of the effective measures in disease diagnostic tests. It is widely being exploited in

radiological tests to evaluate the performance of a medical test. ROC is used to show

the performance of binary classifier by varying threshold. ROC curve is obtained from

the true rate and false positive rate (1-specificity).

22

Chapter 3 : Carotid Artery Image Segmentation

using Modified Spatial Fuzzy C-means and

Ensemble Clustering

In this chapter, first we thoroughly discuss the proposed clustering approach, which

is used to segment carotid artery ultrasound images. The proposed clustering approach

is an extension of spatial fuzzy c-means (sFCM) developed by Chuang et al. [10]. In

addition, an intelligent decision making system has been designed for segregation of

carotid artery ultrasound images into of normal or abnormal subjects. Enhanced

classification performance has been achieved based on the proposed approach

segmentation in comparison to the existing state of the art approaches.

3.1 The Proposed Spatial Fuzzy C-means Modified (sFCMM)

The proposed sFCMM is used to overcome the limitation of sFCM [10]. The sFCM

technique assigns equal weightage to all pixels in sub-window which may produces

noisy clusters. The objective function of the sFCM has been modified such that it

includes a weightage for every pixel of window has proposed. The advantage of the

new membership function is that it is tolerant to noise and offers homogenous

clustering compare to some other states of the art clustering techniques. In sFCMM,

we have considered spatial information of pixel under consideration, which contribute

more to decide the fate of pixel. Detail of the proposed sFCMM can be found in

Section 3.1.4. In addition to this, we have explored the ensemble clustering of

majority voting.

The proposed scheme consists of four stages:

a. Pre-processing

b. Feature extraction and selection

c. Image segmentation

d. IMT measurements and classification of segmented carotid artery ultrasound

images

Chapter 3: Carotid Artery Image Segmentation using Modified Spatial Fuzzy C-means and Ensemble

Clustering

23

Each and every phase of the proposed scheme is explained below and the block

diagram of the proposed scheme is depicted in Figure 3-1.

3.1.1 Image Pre-processing

Ultrasound imaging has several types of degradations such as low quality, speckle

noise, wave interferences and subject movements. It is convenient to minimize these

degradations as much as possible in pre-processing phase. The advantage of image

pre-processing phase is that one can get better segmented images. To increase the

dynamic range, histogram equalization has been applied on carotid artery ultrasound

images [63]. Several approaches have been reported in literature to minimize the

degradations due to noise such as median, bilateral, and antistrophic diffusion filtering

[64-67].

To reduce the effect of noise, we have employed median and bilateral filtering

on carotid artery ultrasound images. The PSNR of both filtered images have been

computed and a noise removal filter is selected having maximum PSNR values.

Median filter with PSNR value 35.51db has been chosen as compared to bilateral

filtering having PSNR value of 28,12db. Median filter is easy to implement technique

and preserve the image details. Impact of median and bilateral filters at segmentation

has also investigated.

3.1.2 Feature Extraction

The objective of feature extraction phase is to formulate a feature vector for every

pixel. In this work, three different types of feature extraction strategies named as

moments of gray level histogram (MGH), continuous wavelet transforms (CWT) and

gray level co-occurrence matrix (GLCM) have been employed. These combined

features are hence used for segmentation of carotid artery ultrasound images. The

extracted features have advantages to represent the medical images [63, 68]. Brief

description of each feature extraction strategy is given below.


Clustering

24

Figure 3-1 Block diagram of the proposed scheme

Moments of Gray Level Histogram

MGH is a frequently used statistical technique of feature extraction which is based on

intensity histogram to represent the medical images. The histogram based features

have better capability to represents the normal and abnormal tissues in medical

images [69]. Nine MGH features are extracted from the carotid artery ultrasound

image [63, 68]. The following expression nu is used to find the nth order moment

about the mean:

where z represents the intensity and ip z is the intensity histogram in the specified

region, m is average intensity level, and L is the total possible intensity level.

The mathematical expressions to find MGH features are given below:

1

0

Mean FM1 . ( )L

i ii

z p z

(3-2)

1

0

( ) . ( )L

n

n i ii

u z m p z

(3-1)


Clustering

25

2Standard Deviation FM2 u (3-3)

1Smoothness FM3 1

1

(3-4)

13

0

Third moment FM4 ( ) . ( )L

i ii

z m p z

(3-5)

12

0

Uniformity FM5 ( )

L

i

i

p z

(3-6)

1

2

0

Entropy FM6 ( ).log ( ( ))

L

i i

i

p z p z

(3-7)

13

0

FM7 ( )

L

i

i

p z

(3-8)

14

0

FM8 ( )

L

i

i

p z

(3-9)

15

0

FM9 ( )

L

i

i

p z

(3-10)

These features are extracted to segment the carotid artery ultrasound images.

Window size of 5x5 has been used and dimension of the image remains same for the

feature extraction strategy. The 7 , 8 , and 9th th th moments have been utilized in

literature for segmentation of medical images [12].

2D-Continuous Wavelet Transform

Since, the medical images used in our experiment are non-stationary in nature and it is

convenient to analyze non-stationary signals using continuous wavelet transform [70].

In pattern recognition 2D-CWT offers robust processing and analysis. Rather than

pixel evaluation, 2D-CWT evaluates the whole image [12]. To overcome the

computational complexity and speed up CWT, the concept of cover-map has been

introduced [71]. The wavelet function used in this study is defined as:


Clustering

26

, ,

1cu s

c c

x ux r

s s

(3-11)

where u is translation and cS is scale parameters, respectively. and r are rotation

angle and rotation matrix with respect to angle, respectively. We have exploited

the scale parameter only for feature extraction, keeping default value for rotation and

translation parameters (i.e. zero).

We have obtained a total of nine features at different scales, which outputs

different filtered images, corresponding to lower frequency bands. Related to 2D-

CWT (Gaussian wavelet) feature extraction strategy, scale parameter is very

important. Smooth areas become visible at higher scale values and high frequency

components (edges) become clear at lower scale values. A number of scale parameter

values ranging 1 to 10 have been utilized to extract features from carotid artery

ultrasound images. Through experiment, it has been observed that eight out of ten

scale values are well enough to segment the images and the remaining redundant scale

parameter values contribute less for segmentation. We have empirically found that

these scale parameter value (i.e. 1.0, 1.6., 2.6, 3.9, 4.0, 5.0, 5.4 and 7.0) produces

better segmentation results.

Gray Level Co-Occurrence Matrix

The GLCM feature extraction strategy is a statistical approach finding in which pixels

spatial relationship is explored also known as gray level spatial dependence matrix.

By default, the spatial relationship of a pixel is defined as the pixel under

consideration and its immediate right neighboring pixel (horizontally adjacent). But

the relationship can be specified between two pixels into diagonal, off-diagonal etc.

In resultant GLCM, an element (J, K) represents the sum of the number of times a

pixel J occurred in a specified spatial relationship with a pixel K of the input image

[13]. A fixed size window of 5x5 has been used for GLCM feature extraction strategy.

From statistical viewpoint, when using a small window size the extracted spatial

information may not be reliable. Alternatively, if window size becomes too large, it

may lead to erroneous textural information. A total of 17 GLCM features have been

extracted from the image and mathematical description of each feature is given below:


Clustering

27

1

2

0 1 1| |

Contrast ( , )

g g gN N N

n i ji j n

n p i j

(3-12)

i

( , ) ,

Correlation

x y

j

X y

i j p i j

(3-13)

2Sum of Squares= ( ) ( , )

i j

i p i j (3-14)

2

1Inverse Difference Moment ( , )

1 ( )i j

p i ji j

(3-15)

2

2

Sum Average ( )

gN

x y

i

ip i

(3-16)

2

2

2

Sum Variance ( _ ) ( )

gN

x y

i

i sum entropy p i

(3-17)

2

2

Sum Entropy ( ) log ( )

gN

x y x y

i

p i p i

(3-18)

Entropy , log ,

i j

p i j p i j (3-19)

Difference Variance variance of x yp (3-20)

1

0

Difference Entropy log

gN

x y x y

i

p i p i

(3-21)

1/2

Information Measure of Correlation IMC :

11

Max ,

2 1 exp 2.0 2

, log ,

i j

HXY HXYIMC

HX HY

IMC HXY HXY

HXY p i j p i j

(3-22)

where HX and HY are the entropy of xp and yp respectively.

1 , log x y

i j

HXY p i j p i p j , 2 logx y x y

i j

HXY p i p j p i p j


Clustering

28

1 1,

20 0

Homogeneity

N Ni j

i j

p

i j

(3-23)

1 1

,

0 0

Dissimilarity

N N

i j

i j

p i j

(3-24)

where ,, P i jRp i j is ,i j th the normalized gray tone spatial dependence matrix

entry, and gN is the total unique gray levels in the image.

The autocorrelation, cluster prominence, cluster shade, maximum probability features

have also been extracted from the given image.

3.1.3 Feature Selection

Feature selection is used to reduce the feature vector dimension and improve accuracy

and computational time [13]. Redundant and irrelevant features are excluded from

feature space and only those features having high contribution are selected for

segmentation. High dimensional data needs more computational time and valuable

resources. It is highly required to select the most relevant features to save valuable

resources and computational time. Feature selection is usually done using search

methods. Many search methods have been reported in literature such as, forward

selection, genetic algorithm, greedy search, sequential backward selection, and

particle swarm optimization.

Feature selection has been performed by employing Genetic Algorithm (GA),

an evolutionary approach, which works on the principle of survival of the fittest. GA

is being successfully used to optimize complex problems. The fittest or best

individuals are selected after each generation. These individuals, after crossover and

mutation, produce new offspring, which are supposed to be better than their parents.

In this research work, hybrid extracted features are provided as input to GA for

selection of most important features using WEKA machine learning tool [72]. GA

parameters are set empirically and the following configuration parameters of GA have

been used for feature selection. Except the following (probability of crossover=0.8,

number of generations=100, Population size=100 and probability of mutation=0.1),

default GA parameters of WEKA has been used. The selected features are shown in


Clustering

29

Table 3.1. It is obvious that segmentation based on the selected features reduces the

computational time and resource requirement.

Table 3.1 Feature selected by WEKA using GA for segmentation of carotid artery

ultrasound images

Feature extraction

techniques

Total number of extracted

features

No of selected

features

MGH 09 02

2D-CWT 09 02

GLCM 17 01

3.1.4 The Proposed Spatial Fuzzy C-means Modified Clustering Algorithm

The proposed sFCMM is an extension of sFCM. The sFCM algorithm has a limitation

of an equal weightage assignment to each pixel in a sub-window. The result of the

equal weightage might be inhomogeneous clustering. To overcome, we have modified

the sFCM membership function and incorporated a weight factor with each pixel in

the sub-window. The weightage is assigned based on the contribution of the pixel. For

this purpose, Euclidean distance based weight has been assigned to each data pixel in

sub-window. The advantage of the Euclidean distance based weightage is that it may

better preserve the image details (edges). Whereas, an equal weightage assignment of

pixels in the sub-widow may smoothen the image detail. The membership function

and its cluster centers are computed using the Equations (2-1) and (2-3), respectively.

After incorporation of the spatial information based on the Euclidian distance the

weight vector become as under:

j

ij il il

l S x

q w u

(3-25)

where, S represents a square window centered at jx and ilw is the Euclidian weight of

every pixel from the centered pixel. This distance based weightage, assigned to every

pixel in the sub-window, gives advantage to sFCMM over FCM and its variants

because, it utilizes the spatial information of pixel in better way [73]. Thus modified


Clustering

30

membership function of the proposed algorithm after including distance based

weightage can be represented as follows:

1

p rij ij

ij cp r

kj kj

k

u qH

u q

(3-26)

The parameters p , r , and other conditions remain same as in spatial fuzzy c-means.

3.1.5 The Proposed Ensemble Clustering Scheme

The main idea behind the ensemble technique is to combine several individual

algorithm results to obtain a segmented image that outperforms every individual

algorithm. Ensemble technique has many advantages; first, it also reduces the risk of

picking the wrong algorithm, second, it also reduces the risk of staying trapped into

local minima, third, one may not be able to obtain the optimal algorithm, hence

combining several classifiers may result in better clustering. Whereas, combining of

several classifiers produce a better approximation of the optimal solution [74]. In this

research work, after obtaining optimized image features by Genetic algorithm and

segmenting by the various algorithm, an ensemble clustering of majority voting

scheme using Expression 3.27 has been applied. Ensemble clustering of majority

voting can be written as:

( )

class argmax ( ( ( ), ))

i

k ic dom y k

x g y x c

(3-27)

where ky x is the classification (segmentation) of the kth classifier and ,g y c is

an indicator function which is defined as:

1 y = c( , )

0 y cg y c

(3-28)

and ky x is defined as:

( )

( ) ( )arg maxk

i

Mk ic dom y

x y c xy P

(3-29)


Clustering

31

where kM denotes classifier k and ( )k

iMy c xP

denotes the probability of y

obtaining the value c for given instance x .

3.1.6 Post-processing Operations

Some isolated pixels might be wrongly segmented into other clusters and requires to

be removed from the final segmented image. For this purpose, post-processing

techniques have been applied at the segmented images. In order to remove isolated

pixels, morphological opening and closing operations have been applied on

segmented image, which help to measure IMT values more accurately. The opening

and closing operations with structuring elements of size 5x5 have performed by

employing the following expressions [59].

( ) f b f b b (3-30)

( ) f b f b b (3-31)


The segmented carotid artery ultrasound image needs to be classified into normal and

abnormal subjects. Quality of segmentation has a high impact on classification. For

classification, artificial neural networks (ANN) is widely being used in many areas and

has been demonstrated by many researchers in the field of classification and pattern

recognition [75-77], image restoration [78], and machine vision [79]. Consequently,

we have employed ANN to classify the segmented images based on measured IMT

values. In this work, we have used multilayered perceptron with back-propagation

neural network (MLBPNN) algorithm for classification of segmented images into

normal or abnormal.

Neural Networks Architecture

The network configuration used for classification based on IMT values is given below.

Input Layer: Five neurons

Hidden Layer: Fifty neurons

Output Layer: Two neuron

Activation Function: Binary sigmoidal function.

No. of Epochs: 500


Clustering

32

The neural network architecture has high impact on classification accuracy.

Each layer of neural network consists of different processing neurons. Among these

layers, numbers of hidden layer neurons are most important. Neural network with

small number of hidden layer neurons may affect the network in a way that it may not

have sufficient degree of freedom to learn the patterns correctly. Network may over-

fit or it may take a long time to converge, if hidden layer neurons are high [80].

In order to find an optimal number of hidden layer neurons a validation set

error has been used. In this process, various combinations of input and hidden layer

neuron are compared. This process has been repeated for number of times with one

leave out concept and the model with minimum mean sum of prediction error (MSPE)

has been selected [80, 81]. Enhanced classification accuracy has been achieved by

utilizing the above mentioned architecture obtained by the aforementioned procedure.

Training and Testing of MLBPNN

IMT values have been measured from segmented carotid artery ultrasound images and

given as input to the MLBPNN classifier. For training and validation, 10-fold cross

validation technique has been utilized.

3.3 Experimental Results and Discussions

The proposed scheme has been used to segment the synthetic as well as real carotid

artery ultrasound images. Optimized features obtained by WEKA using GA, hence,

used as input to clustering algorithm. Ensemble clustering approach of majority voting

has also been developed to segment carotid artery ultrasound image. All of the

computations have been performed on Intel Core i7 PC with Matlab 7.12 (2011a). The

dataset detail used in this research work is described in Section 2.5. The images have

been categorized into normal and abnormal with the help of medical experts.

MLBPNN is used for classification of the segmented images. Classification

performance of MLBPNN has been assessed against already labeled images.


Clustering

33

3.3.1 Performance Assessment of Proposed Technique on Phantom Ultrasound

Images

In this section, the performance of the proposed sFCMM and ensemble clustering

approaches is evaluated on ultrasound phantom images of size 240-by-300. Out of 35

hybrid features, GA has selected 21 features, which have been used for segmentation.

The selected features include four MGH, seven 2D-CWT and GLCM features

respectively. Phantom image has three clusters, namely liquid medium, dark and light

colored phantom objects.

Figure 3-2(a-f) shows the original ultrasound phantom image, segmentation

results of FCM, sFCM, the proposed sFCMM, SOM and K-means respectively. The

segmented image by the the proposed ensemble clustering approach is shown in

Figure 3-2(g), whereas Figure 3-2(h) shows result of image segmentation using

sFCMLSM. It can be observed from Figure 3-2(g) that ROI is successfully separated

from background area, using the proposed technique. On the other hand, sFCMLSM

technique misclassified pixels and hence ROI has not successfully been extracted

from background area.

Clustering quality measures are also computed from segmented images and

compared with the other state of the art techniques. Performance of various

segmentation techniques on phantom image are shown in Table 3.2. It is evident from

Table 3.2 that the proposed ensemble clustering offers quality clustering with

minimum DBI value as compared to the other clustering techniques. The PC and CE

measures require fuzzy membership function and ensemble clustering do not have

fuzzy membership function, hence these measures are not calculated for ensemble

clustering approach as well as K-means and SOM approaches.

3.3.2 Performance Assessment of the Proposed Technique at Real Carotid

Artery Ultrasound Images

The carotid artery ultrasound image consists of three clusters, area inside the artery,

arterial wall and background tissues. Figure 3-3(a) shows one of the original carotid

artery images, whereas Figure 3-3(b) shows the histogram equalized image.


Clustering

34

Table 3.2 Clustering performance measures of the proposed approach on phantom

ultrasound image.

Techniques DBI PC CE

FCM 0.4031 0.9626 0.0719

sFCM 0.4247 0.9917 0.0135

The proposed sFCMM 0.4705 0.9957 0.0135

K-means 0.3864 -- --

SOM 0.3863 -- --

Proposed Ensemble Technique 0.3006 -- --

(a) (b) (c) (d)

(e) (f) (g) (h)

Figure 3-2 Experimental results: (a) the original phantom image; (b) FCM

segmented image; (c) segmented image using sFCM; (d) the proposed sFCMM

segmented image (e) segmented image using K-means (f) segmented image using

SOM (g) proposed ensemble scheme segmented image and (h) segmented image

using sFCMLSM technique.

3.3.3 Performance Analysis of Ensemble Clustering based on Fuzzy C-means

In this study, five different clustering algorithms for segmentation of carotid artery

ultrasound images and final segmented image was obtained using ensemble clustering

of majority voting scheme. In clustering different image pixels, tie is resolved by

assessing the DBI measure of each clustering algorithm. The new pixel is clustered by


Clustering

35

the algorithm, which achieves the lowest DBI values. One thing should be noted is that

from this point onward pixel clustering with ties would be referred to as “mismatched

pixel”.

(a) (b) (c)

(d) (e) (f)

(g) (h) (i)

(j)

Figure 3-3 Experimental results: (a) the original carotid artery image; (b) histogram

equalized image; (c) FCM segmented image; (d) segmented image by sFCM; (e) the

proposed sFCMM segmented image (f) K-means segmented image (g) SOM

segmented image (h) the proposed ensemble segmented image (i) morphological

corrected image, (j) sFCMLSM technique segmented image.

It can be observed from Table 3.3 that FCM clustering results are utilized for

tie breaking of mismatched pixels because FCM algorithm has achieved minimal


Clustering

36

DBI. It is evident from Table 3.4 that ensemble based on FCM clustering has

minimum DBI among all other mentioned techniques. Figure 3-3(h) shows that

carotid artery ROI is very clear and well separated from the background tissues.

Although, other measures of FCM like PC and CE is lower as compared to sFCM and

sFCMM, but DBI measure of FCM is better than those of other approaches.

Figure 3-3(c) depicts the carotid artery ultrasound image segmented by FCM

whereas Figure 3-3(d) and (e) do those by sFCM and sFCMM clustering algorithms,

respectively. It can be observed from Figure 3-3(c) that there exist some incorrectly

clustered pixels in the FCM segmented image. It is due to the fact that FCM algorithm

is susceptible to noise. The sFCM utilized spatial information, hence it works better as

shown in Figure 3-3(d) compared to FCM algorithm. Further, from Figure 3-3(d) it

can be noticed that some patterns are not clustered accurately by the sFCM algorithm,

especially near the ROI. It is because of assigning equal weightage to all neighboring

pixels in sub-window around the pixel considered. In this connection, the proposed

sFCMM has been utilized to assign the Euclidian distance based weight to the

neighboring pixels in the sub-window. It can be observed, from Figure 3-3(e), that

sFCMM has shown an improved clustering performance over FCM and sFCM

approaches. Carotid artery ultrasound images segmented by K-means and SOM are

shown in Figure 3-3(f) and (g), respectively. Visual inspection shows that K-means

and SOM segmentation results are almost similar.

The proposed ensemble clustering scheme of carotid artery image is presented

in Figure 3-3(h). From visual inspection of the segmentation results, it is evident that

the ensemble clustering strategy is promising compared to other approaches. The

post-processing operation is demonstrated on Figure 3-3(h) and its resultant image is

shown in Figure 3-3(i). Here, one can see that the image is segmented well and region

of interest is successfully separated from background area. Figure 3-3(j) shows the

image segmented by sFCMLSM for comparison purpose. It can be observed that

background tissues are merged with ROI as shown in Figure 3-3(j).

The performance of the proposed approach is also examined using all the

thirty five features without performing any feature selection mechanism. Table 3.3

shows the comparison of the proposed ensemble approach on reduced feature set


Clustering

37

selected by GA, whereas Table 3.5 shows the performance comparison suing thirty

five extracted features. Reduced feature set strategy offers outstanding performance

due to its features orthogonalaty with reduced computational cost.

Table 3.3 The clustering quality performance comparison of various techniques.

Clustering Techniques

Clustering quality measures

DBI PC CE

Hassan

et al.

[73]

Proposed

Strategy

Hassan

et al.

[73]

Proposed

Strategy

Hassan

et al.

[73]

Proposed

Strategy

FCM 0.51 0.41 0.79 0.85 0.38 0.27

sFCM 0.57 0.55 0.94 0.96 0.09 0.06

The Proposed

sFCMM 0.61 0.61 0.97 0.98 0.04

0.03

K-means 0.52 0.42 -- -- -- --

SOM 0.52 0.42 -- -- -- --

Proposed Ensemble

Clustering

0.45 0.30 -- -- -- --

3.3.4 Ensemble Clustering Assessment based on sFCM

We have also assessed the performance of ensemble clustering based on sFCM

algorithm. The PC and CE measures of sFCM are better than that of FCM, but DBI of

sFCM is comparable with the FCM algorithm. When mismatched pixels are replaced

with sFCM clustering results, resultant DBI measure becomes higher than FCM as

shown in Table 3.4. Closely observing the sFCM segmented image, some background

tissues along with region of interest are mislabeled and considered as ROI area. This is

why sFCM based ensemble clustering performance is lower as compared to FCM

clustering.

3.3.5 Performance Analysis of sFCMM Based Ensemble Clustering

In this study five different segmentation algorithms are used and ensemble

performance of these algorithms is analyzed one by one. Performance of ensemble

clustering using sFCMM is analyzed and has less effective as compared with FCM.


Clustering

38

The ensemble performance based on sFCMM has lower as compared with FCM in

term of DBI measure. On the other hand, the proposed sFCMM outperforms the other

techniques in terms of PC and CE measures. Performance evaluation of ensemble

clustering based on the proposed sFCMM is shown in Table 3.4.

3.3.6 Clustering Performance Analysis of K-mean & SOM Based Ensemble

Performance of the ensemble clustering has also been investigated based on SOM and

K-means techniques. Ensemble clustering based on SOM and K-means show less

effectiveness as compared to the FCM based ensemble clustering. K-means and SOM

produces equal DBI, hence ensembles based on these techniques yield almost similar

results. Ensemble performances based on K-means and SOM are shown in

Sample carotid artery ultrasound images segmented by the proposed ensemble

clustering approach are shown in Figure 3-4(a) and (c). Whereas, segmentation results

of FCMLSM technique are shown in Figure 3-4(b) and Figure 3-4(d). From Figure

3-4, it is evident that the proposed ensemble clustering offers better segmented image

results compared with sFCMLSM.

The proposed ensemble clustering approach has outperformed the other

techniques of segmentation. Instead of using thirty five features, extracted from a

carotid artery ultrasound image, only five optimal features have been used for

segmentation. The advantage of feature reduction is that it reduces the dimensionality

and computational time and simultaneously it does not affect the overall clustering

quality which shows the effectiveness of the proposed approach. Table 3.3 represents

the ensemble clustering performance over five selected features, whereas Table 3.5

shows the clustering performance over all thirty five extracted features. The

segmentation performance measure (DBI) vs increasing number of features is also

shown in Figure 3-6. High clustering performance is evident when using five selected

features rather than using all 35 extracted features. There is no significance change in

clustering results in terms of DBI have been observed in Table 3.3 and Table 3.5

when clustering is performed at 35 features or five selected features, respectively.

But, clustering at selected features takes less time and resources. Temporal

comparison of the proposed scheme to segment various images is shown in Table 3.8.


Clustering

39

Table 3.4 Performance comparison of the proposed ensemble clustering approach

based on FCM, sFCM, sFCMM, K-means and SOM approaches.

DBI measure

FCM sFCM sFCMM

K-

mean

s

SOM Ensemble

Clustering

FCM Based Ensemble 0.41 -- -- -- -- 0.30

sFCM Based Ensemble -- 0.55 -- -- -- 0.41

The proposed sFCMM

Based Ensemble -- -- 0.61 -- -- 0.46

K-means Based Ensemble -- -- -- 0.42 0.35

SOM Based Ensemble -- -- -- -- 0.42 0.35

Table 3.5 Performance comparison of the proposed approach segmentation of carotid

artery ultrasound images with other techniques by utilizing all thirty five extracted

features.

Techniques DBI PC CE

FCM 0.41 0.85 0.27

sFCM 0.55 0.96 0.06

sFCMM 0.61 0.98 0.03

K-means 0.42 -- --

SOM 0.42 -- --

Proposed Ensemble Technique 0.30 -- --

Figure 3-5(a) shows the segmentation results of carotid artery ultrasound

image using bilateral filter, whereas Figure 3-5(b) shows the segmented image using

median filter as pre-processing step for noise removal. Visual inspection shows

misclassified patterns reported in Figure 3-5(a), while promising segmentation has

been done in Figure 3-5(b). Table 3.6 shows the segmentation performance

comparison of median and bilateral filtering. The median filtering offers better quality

segmentation as compare with bilateral filtering, hence used as a pre-processing for

noise reduction.


Clustering

40

Table 3.6 The Performance comparison of segmentation using median and bilateral

filtering for image denoising.

Clustering Techniques Median filtered image Bilateral filtered image

DBI PC CE DBI PC CE

FCM 0.4031 0.9626 0.0719 0.5717 0.7722 0.4201

sFCM 0.4247 0.9917 0.0135 0.6325 0.9406 0.1004

The proposed sFCMM 0.4705 0.9957 0.0135 0.6677 0.9754 0.0415

K-means 0.3864 -- -- 0.5779 -- --

SOM 0.3863 -- -- 0.5780 -- --

Proposed Ensemble

Technique 0.3006 -- -- 0.4286 -- --

(a) (b)

(c) (d)

Figure 3-4 (a) and (c) Carotid artery image is segmented using the proposed

ensemble clustering approach and (b) and (d) images segmented by sFCMLSM

approach.


Clustering

41

(a) (b)

Figure 3-5 Proposed approach segmentation of carotid artery ultrasound image using

(a) bilateral and (b) median filtering as pre-processing step for noise reduction.

Figure 3-6 Performance measures versus increasing number of features.

(a) (b)

Figure 3-7 (a) Magnified IMT measurement section of an original carotid artery

ultrasound image (b) Magnified IMT measurement section of a segmented carotid

artery ultrasound image using the proposed scheme.


Clustering

42

(a) (b)

Figure 3-8 (a) IMT measurements of a normal carotid artery image (b) IMT

measurements of abnormal carotid artery

Figure 3-9 ROC curve analysis of true and false positive rates using MLBPNN

classifications.

3.3.7 Classification of Carotid Artery Ultrasound Images

IMT values have been measured from images segmented by the proposed ensemble

clustering approach. The dataset, which we have utilized to assess the performance of

the proposed approach is comprised of 150 images of different patients and ages.

Figure 3-7 shows a magnified ROI of original and segmented image using the

proposed ensemble approach. From Figure 3-7(b), one can observe that the proposed

approach successfully segregated the plaque presented in the carotid artery. With the

consultation of medical expert, IMT measurements have been performed and have


Clustering

43

been used in the development of decision making system. Figure 3-8(a, b) shows the

IMT measurements of normal and abnormal carotid arteries. The IMT mean and

standard deviation of the normal carotid artery are 0.3753 mm and 0.1021 mm,

respectively. On the other hand, the mean of abnormal IMT value of carotid artery

was 0.7561 mm and standard deviation was 0.1647 mm. The measured IMT values

have been given as input to the MLBPNN classifier for training and validation.Table

3.7 shows the classification performance of the MLBPNN classifier. We have

achieved 98.4% accuracy using MLBPNN based on the proposed segmentation

approach, which demonstrates the usefulness of the proposed approach. Figure 3-9

shows the ROC curve, which comprises of true vs. false positive rates. The ROC

curve which is close to vertical axis shows fewer misclassifications and thus validates

the effectiveness of the proposed approach. Classification results obtained through the

proposed technique are compared with other methods based on overall accuracy.

Santhiyakumari et al. [32] have employed MLBPNN for classification and have

reported a maximum of 96% classification accuracy on their dataset. Whereas, we

have employed a straight forward approach to classify the segmented carotid artery

ultrasound images and have obtained 98.4% classification accuracy. Statistical results

indicate that the proposed approach outperforms the other state of the art techniques,

hence successfully used for identification of plaque in carotid artery.

Table 3.7 Classification performance measure of the MLBPNN

Classification validity measures Performance

Accuracy (%) 98.40%

F-Measure 0.9630

MCC 0.9295

Sensitivity 0.9915

Specificity 0.9394

Since, GA has been applied to remove the redundant and irrelevant features, it

is an evolutionary process and takes time for selection the optimal features. Secondly,

to train the neural networks, it also takes a considerable time. However, the proposed

technique segments images in less time at testing phase.


Clustering

44

Table 3.8 The temporal comparison of the proposed approach at reduced and whole

extracted features set.

Images Segmentation time in seconds

Clustering with

feature selection

Clustering without

feature selection

Figure 3-2 34.00 48.00

Figure 3-3 34.30 91.00

Figure 3-4 (a) 33.90 88.00

Figure 3-4 (c) 58.00 90.86

Summary

In this chapter, a new clustering algorithm named sFCMM has been proposed for

segmentation of carotid artery ultrasound images. In the proposed scheme, different

types of features are extracted and important features are selected by GA. The main

objective of features optimization is to reduce feature vector dimensionality which

results in low computational cost and offers higher accuracy. Segmentation results of

the proposed clustering algorithm have been compared with the state of the art

clustering techniques. Ensemble clustering of majority voting has also explored and

results indicate that the proposed ensemble clustering offers quality segmentation. The

carotid artery images segmented efficiently and it has been observed that the ROI is

efficiently separated from background tissues. IMT values have been measured from

the segmented images, hence have been used for classification. We have achieved

98.4% classification accuracy by employing MLBPNN, which shows the effectiveness

of the proposed approach.

In this chapter, we have used pre-processing technique of median filter to

minimize the effects of noise. The limitation of applying pre-processing technique is

that it smoothes the image detail, which results in loss of important information from

the image. However, preservation of image detail is vital for segmentation and post-

processing analysis. In this connection, the next chapter proposes another technique of

segmentation, which does not require any pre-processing step for noise removal and

hence, better segmentation results are obtained.


Clustering

45

Outcomes of the current chapter:

1. Carotid artery ultrasound image segmentation using modified spatial fuzzy c-

means and ensemble clustering, “Computer Methods and Program in

Biomedicine”, Journal, Vol. 108, pp 1261-1276, 2012. (Impact Factor=1.555)

2. Image Clustering using Improved Spatial Fuzzy C-means Modified, ACM,

ICUMIC, Kuala Lumpur, Malaysia, 2012.

3. An Optimized Fuzzy C-means clustering with spatial information for carotid

artery image segmentation, IEEE, IBCAST Islamabad, Pakistan, 2011.

46

Chapter 4 : Robust Information Gain based

Fuzzy C-Means Clustering and Classification of

Carotid Artery Ultrasound Images

In this chapter, a new robust clustering approach namely Information Gain based

Fuzzy C-Means (IGFCM) has been proposed. IGFCM utilizes the concept of

information gain into basic FCM framework and thus avoids any pre-processing step

for noise reduction. The proposed approach has been applied to segment synthetic,

daylight, CT-liver, and real carotid artery images. In order to evaluate the robustness,

segmentation is performed using IGFCM on images degraded with Gaussian noise of

various intensities. The proposed approach successfully segments the noisy and noise

free images. To evaluate the effectiveness of the proposed IGFCM, a significant

amount of real carotid artery images are segmented. Segmentation results of the

proposed IGFCM technique have been compared with other state-of-the-art

techniques. It has been observed from segmentation results both visually and

quantitatively that IGFCM performs better compare with other state-of-the-art

approaches. In addition, an intelligent decision making system has also been proposed

to segregate the normal and abnormal subjects.

4.1 The Proposed Clustering Algorithm

The proposed segmentation algorithm incorporates the information gain concept into

the basic FCM framework. The membership values of FCM algorithm need to be

updated by the procedure of information gain. FCM cost function, membership

functions, and cluster centers can be computed by using the Equations (2-1), (2-2),

and (2-3), respectively.

4.1.1 The Proposed Information Gain Based FCM Clustering Algorithm

The conventional FCM algorithm is susceptible to noise and may produce

inhomogeneous clustering. To coup with the problem effectively, the concepts of

information gain has been incorporated into the basic FCM algorithm. The idea of

Chapter 4: Robust Information Gain based Fuzzy C-Means Clustering and Classification of Carotid


47

entropy was coined by Claude Shannon in his pioneer work on information theory

[82]. It is the measure of uncertainty of a random variable and is calculated by the

following expression.

2entropy log , 1,i i

i

i p p i C (4-1)

where ip is the probability of ith cluster in a certain neighborhood.

Information gain is the measure of goodness of an attribute. It is computed for a

particular class using following expression.

entropy , 1,iIG i EI i C (4-2)

EI represents expected information which is computed from available information

between class entropies and their respective probabilities. EI can be calculated using

Equation (4-3).

1, 1,

en tropy( , )i j

i C j C

EI p p i j

, where, and i j j i (4-3)

where ip and jp are probabilities of cluster i and j, respectively, entropy ,i j is

between class entropy for i and j , computed by the following expression.

1, 1,

2 2entropy , log log i i j

i j C

j

C

i j p p p p

(4-4)

Figure 4-1 The block diagram of the proposed IGFCM approach



48

The proposed IGFCM algorithm is composed of two-phases. In first phase,

input image is segmented by FCM but limited to few iterations to obtain an initial

estimate of true segmentation. In the second phase, for every pixel in FCM-segmented

images, entropy and information gain are computed from its surrounding pixels in a

certain neighborhood (5x5 selected empirically in our case). The proposed technique

then updates the fuzzy membership values of every pixel by swapping them together.

The swapping order of the pixel’s fuzzy membership values is determined by the rank

of its information gain values computed for different clusters. The cluster centroids

are then updated, as in classical FCM. After that, an iteration of classical FCM

proceeds with the new membership values and update cluster centroids. This process

continues until the difference between cluster centroid values in consecutive iterations

reduces below a particular threshold or maximum numbers of iterations have been

reached. The algorithm of IGFCM is presented in Algorithm 4.1.

The defuzzification of membership function ( jku ) yields the final image

segmented by the proposed IGFCM algorithm. Figure 4-1 shows the block diagram of

the proposed scheme. A detailed illustration of algorithmic steps is shown by an

example in Figure 4-2. It should be noted that phase-III is concerned with application

of IGFCM algorithm on classification scenario which will be described in Section

4.3.2.

4.1.2 Illustration of the Proposed IGFCM approach

An example is presented here to demonstrate the working mechanism of the proposed

IGFCM algorithm. Phase-I of the algorithm is straightforward because classical FCM

is well studied and can be easily reproduced. Therefore, only phase-II of IGFCM

algorithm is described in this illustration. We have considered a local neighborhood of

5x5 pixels from the input image (FCM segmented) and apply phase-II steps of the

proposed algorithm. These steps will be repeated for every pixel of the input image

until the convergence criteria has been met; hence, the final IGFCM segmented image

has been obtained.



49

Algorithm-4.1:

Phase-I

Perform the FCM segmentation on the input image.

Compute vkn (initial cluster centroids) for all k and n 1,C

Compute ujk (initial fuzzy membership values) for all j and k 1,C

Phase-I concludes after the image is segmented by few iterations of FCM

algorithm. The output of this phase is fed to Phase-II for further

computation.

Phase-II

Use a 5x5 window and iterate it over FCM segmented image. For each iteration,

repeat steps 1 to 11.

1. Calculate the Probability of each class.

j

j

np

N , 1,j C

where nj represents the number of pixels in the window, which belong to class

j.

N is the total number of pixels (5x5) in the window.

2. Calculate the Entropy of each class entropy( )i using Equation (4-1).

3. Calculate the between class entropy ( , )entropy i j for each two-class

combination using Equation (4.4).

4. Calculate Expected Information EI using Equation (4-3).

5. Calculate Information Gain for each class iIG using Equation (4-2).

6. Compute ,sorted

jk jku sort u ORDER , where sort orders the vector ujk in the

order specified by ORDER. The ordered vector is returned insorted

jku. We have

used descending order in our algorithm.

7. Update New

jku using following expression

, ,

New sorted

jk rank IG IG i ku u 1,,k Cj

where IG sets of information gain values for all classes and rank (IG, i)

returns the rank (i.e. order) of number i within a set of numbers IG.

8. Update knv from Equation (4-1) usingNew

jku .

9. Set New

jk jku u .

10. Use jku and knv to perform FCM iteration.

11. Repeat all the above steps until the stopping criterion is met.



50

Step 1. A sample local neighborhood is shown in Figure 4-2 where the pixel under

consideration is shaded. Total clusters are supposed to be 5 in this case.

Figure 4-2 A sample pixel and corresponding local neighborhoods

The given input fuzzy membership functions in this example is as follows.

0.1 0.15 0.24 0.23 0.28jku

Step 2. Calculate the probability of each class.

1 2 3 4 5

6 5 5 4 5, , , ,

25 25 25 25 25p p p p p

Step 3. Calculate the entropy of each class. As an illustrative example, entropy (1) is

computed as follows.

2

6 6entropy 1 = log = 0.4941

25 25

Therefore, entropy( ) 0.4941 0.4644 0.4644 0.4230 0.4644 , 1,ii C

Step 4. Calculate the entropy between classes for each two-class combination. As an

illustrative example, entropy (1,2) is computed as follows.

2 2

6 6 5 5entropy 1,2 = log log = 0.9585

25 25 25 25

0.9585 0.9585 0.9172 0.9585

0.9585 0.9288 0.8874 0.9288

Therefore, entropy( , ) , , 0.9585 0.9288 0.8874 0.9288

0.9172 0.8874 0.8874 0.8874

0.9585 0.9288 0.9288 0.8874

1,Ci j i j

Step 5. Calculate Expected Information.

11 11 10 11entropy(1,2) entropy(1,3) entropy(1,4) entropy(1,5)

25 25 25 25

10 9 10 9entropy(2,3) entropy(2,4) entropy(2,5) entropy(3,4)

25 25 25 25

10 9entropy(3,5) entropy(4,5)

25 25

EI



51

11 100.9585 0.9585 0.9585 + 0.9172 0.9288 0.9288 0.9288

25 25

90.8874 0.8874 0.8874 3.8606

25

EI

Step 6. Calculate Information Gain for each class. As an illustrative example, IG1 is

computed as follows.

1 entropy(1) 0.7950 3.8606 3.0655IG EI

Therefore, 3.3664 3.3962 3.3962 3.4376 3.3962 IG

Step 7. Compute , ' 'sortedjk jku sort u desc

.

0.28,0.24,0.23,0.15,0.10sorted

jku

Step 8. UpdateNew

jku.

1 2 3 4 5, , , , 3.3664, 3.3962, 3.3962, 3.3962, 3.4376IG IG IG IG IG IG

, 5 3 4 2 1 irank IG IG

1 5Therefore, New sortedk ku u

, 2 3Newk

sortedku u

, 3 4Newk

sortedku u

, 4 2Newk

sortedku u

, 5 1Newk

sortedku u

. . 0.28,0.24,0.23,0.10,0.15Newjki e u

Until this point, we have obtained the new fuzzy membership functions. We can use

these new membership values to continue the process for further calculations (e.g. to

find out new centroids as in step 8 of Phase-II).

4.1.3 Clustering Quality Measures

To check the usefulness of the proposed IGFCM clustering approach quantitatively,

the clustering quality measures partition co-efficient (PC) and classification entropy

(CE) have been computed for every segmented image. Detail description of PC and

CE measures can be found in Section 2.6.

4.2 Decision system for Segmented Carotid Artery Ultrasound

Images

The segmentation performance is critical to disease diagnosis in most computer-aided

diagnostic systems based on medical imaging. Effective segmentation results offer



52

high classification results. In this section, the impact of segmentation using the

proposed IGFCM algorithm has been evaluated on classification of carotid artery

ultrasound images. For this purpose, a decision system based on PNN classifier has

been proposed. In the proposed decision system, IMT values have been measured

from segmented carotid artery ultrasound using IGFCM algorithm. These values

correspond to the number of pixels along columns of a selected region of the

segmented images which belong to the arterial wall [83, 84]. Based on these

measurements, a set of 6 features namely mean, variance, skewness, kurtosis, min and

max [28, 85] has been formed and provided as input to the classifier. Probabilistic

neural network (PNN) has been employed for classification owing to better learning

capability of neural networks. The use of PNN has been justified by comparing its

performance with two other classifiers. Finally, the effect of segmentation has been

analyzed by segmenting the carotid artery ultrasound images with different techniques

including the proposed algorithm. The dataset used in this study is described in

Section 2.5.

4.2.1 Probabilistic Neural Networks Classifier

Probabilistic neural network (PNN) is a popular and frequently used tool for decision

making, which works well both for linear and nonlinear data. It is a feed forward

neural network that utilizes Kernel Fisher discriminant analysis and Bayesian

networks. It employs four layers of neural network, namely, input, hidden, pattern,

and output layers. Compared to multilayer perceptron, PNN is faster, more accurate,

and tolerant to outliers. Moreover, the complexity of the decision surface of PNN

classifier can be controlled by varying the spread parameter. In this case, the spread

parameter value has been empirically found and set to 0.9, which offers high

classification accuracy. The decision surface can approach Bayes optimal classifier

using an optimal value of the spread parameter. PNN classifier has the capability to

operate in parallel without getting feedback from individual input neurons. Finally,

for time variant problems, old patterns can be replaced with new patterns. More

details on PNN can be found in [86].

In this work, IMT features are given as input to the network along with spread

parameter value 0.9 which was set empirically to yield the optimal performance. For

training and validation of PNN, 10-fold cross validation using Jackknife method has



53

been utilized. The detail of 10-fold cross validation and Jackknife approaches can be

found in Section 2.6.

We have used accuracy, sensitivity, specificity, Mathew’s Correlation

Coefficient (MCC), and F-measure to assess the effectiveness of the classifier.

Receiver Operating Characteristics (ROC) curve has been drawn to compare true

positive rate vs. false positive rate. ROC curve, widely used in medical diagnosis

systems, is an effective analysis tool that demonstrates the performance of decision

making system. The area under the curve (AUC) [87] has also been computed from

ROC curve, which shows the effectiveness of the decision system.


Two major experiments have been conducted to access the effectiveness of the

proposed IGFCM algorithm. In first experiment, the proposed IGFCM technique is

applied to segment the various modality images including synthetic, daylight, and CT

liver images. The segmentation results have been obtained for both noisy and noise-

free version of images. Performance of the proposed algorithm has been compared

with some state-of-the-art techniques such as, FCM, sFCM, sFCMM, and FLICM,

both visually and quantitatively. Partition coefficient (PC) and classification entropy

(CE) have been applied as quantitative clustering performance measures.

Whereas, in second experiment, segmentation and classification on real

carotid artery ultrasound images has been performed. These images are segmented by

the proposed IGFCM algorithm prior to classification so that the effect of

segmentation on classification has to be investigated. The proposed decision system,

based on PNN classifier, successfully identifies the segmented carotid artery

ultrasound images as normal or abnormal subjects. Classification results of PNN have

been compared with K-nearest neighbor (KNN) and MLBPNN classifiers. The feature

set feeds input to PNN classifier to obtain the final decision about pathological state

of the subject. A dataset of 300 real carotid artery ultrasound images is used in our

decision making system (Chapter 2). Finally, the superiority of employing the

proposed IGFCM in the decision system has been shown. In particular, the carotid

artery dataset is segmented using different variants of FCM and used in the decision

system in a way similar to IGFCM. The classification results employing different



54

segmentation techniques verified that the proposed IGFCM technique provides

improvement in classification accuracy compared to other segmentation techniques.

We have considered two scenarios to evaluate the performance of the proposed

IGFCM algorithm. Detail of each scenario is described next:

4.3.1 Scenario-I: Segmentation Performance of the Proposed Algorithm

In this scenario, we have applied the proposed IGFAM technique for segmentation of

a synthetic, daylight and CT liver images. To access the robustness of the proposed

approach, Gaussian noise of various intensities has been added and then segmented by

the proposed approach. Validation of the proposed approach to segment the images

from different modalities helps verifying the general-purpose reliability of IGFCM

technique.

Synthetic Image Based Segmentation Performance

We first show the results of IGFCM and variants of FCM to segment the synthetic

image, having three clusters. Gaussian noise of different intensities (e.g. 0.01, 0.02

and 0.03) has been added to the original image. Corresponding noisy images are

shown in Figure 4-3(a), Figure 4-4(a), and Figure 4-5(a), respectively. Figure 4-3,

Figure 4-4 and Figure 4-5(b-f) show the images segmented by different techniques at

various noise levels. Segmentation results obtained by IGFCM for this simple image

are comparable to that of sFCM and sFCMM, however, clear superiority can be

observed compared to FCM and FLICM algorithms at almost all noise levels. Similar

conclusion is drawn from quantitative comparison of segmentation performance

measures shown in Table 4.1.



55

(a) (b) (c)

(d) (e) (f)



proposed IGFCM algorithm.

(a) (b) (c)

(d) (e) (f)






56

(a) (b) (c)

(d) (e) (f)




(a) (b) (c)

(d) (e) (f)

Figure 4-6 The Wolf image segmentation, (a) The Wolf image (original), (b) FCM

segmentation, (c) sFCM segmentation, (d) sFCMM segmentation, (e) FLICM segmentation, and (f) the proposed IGFCM segmentation



57

(a) (b) (c)

(d) (e) (f)

Figure 4-7 Segmentation of noisy Wolf image (noise variance 0.01), (a) noisy Wolf

image, (b) FCM segmentation, (c) sFCM segmentation, (d) sFCMM segmentation,

(e) FLICM segmentation, and (f) the proposed IGFCM segmentation

Segmentation Performance of IGFCM on Daylight Image

Daylight images or natural images also find several applications in computer vision,

such as surveillance and object tracking. Image segmentation is an essential

intermediate processing step in these applications. Performance of the proposed

IGFCM algorithm has been examined on a daylight image, named 'the wolf image',

which is shown in Figure 4-6(a). The objective is to segment the image into three

logical segments. As shown in Figure 4-6(b-f), the images segmented by different

variants of FCM contain misclassified pixels compared with the image segmented by

the proposed IGFCM algorithm. The image has corrupted by adding Gaussian noise

of variance 0.01 and different segmentation techniques have been applied to the noisy

image. Figure 4-7(a-f) show noisy image as well as the image segmented by FCM,

sFCM, sFCMM, FLICM, and IGFCM, respectively. From visual inspection, it can be

observed that the proposed IGFCM algorithm produces quality segmentation with

minimum misclassifications. On the other hand, images segmented by standard FCM

and its variants contain a lot of misclassified patterns. It can be verified that IGFCM is

more robust to noise and produces more homogeneous clustering compared with other

segmentation techniques.



58

Segmentation Performance of IGFCM on CT-Liver Image

Recognizing the importance of segmentation process on medical image based disease

diagnosis, the proposed IGFCM algorithm has been applied to a CT liver image

(extracted region of interest) with a tumor. The proposed IGFCM algorithm, standard

FCM and its variants are also applied to segment CT-Liver image. The objective is to

segment the image into three clusters namely background, tumor and the main liver

tissues. For a segmentation to be ranked as high quality, the tumor must stand out of

other two segments in the image. The original image and the resultant segmented

images obtained after applying different segmentation algorithms on original CT-

Liver image are shown in Figure 4-8(a-f). The over-segmentation is immediately

visible in case of segmentation produced by FCM, sFCM and sFCMM techniques.

Whereas, in case of FLICM segmentation, several misclassified patterns can be

observed in the tumor region as shown in Figure 4-8(e). However, the proposed

IGFCM segments tumor regions with high precision. The quantitative performance

measures have also been computed by applying different segmentation techniques to

noisy versions of CT liver image. The proposed IGFCM outperforms all mentioned

techniques at various noise levels as shown in Table 4.1.

Quantitative Analysis

Table 4.1 compares PC and CE measures for different segmentation techniques

applied to noisy and original images discussed in previous sections. Note that the

quantitative results for carotid artery image correspond to Figure 4-10. The

quantitative results shown in Table 4.1 indicate a better segmentation performance

using IGFCM than standard FCM and its variants. Particularly, there is a significant

improvement in overall segmentation quality when input images contain noise of

different levels. An exception is the case of synthetic image where sFCM and the

proposed approach to produce comparable results. Also, the visual inspection of

Figure 4-3, Figure 4-4 and Figure 4-5(c &f) reveals that the results of sFCM and

IGFCM segmentations are comparable. However, IGFCM performs at other image

modalities. Hence, the general superiority of IGFCM is maintained.

The performance of different segmentation techniques, in terms of PC and CE,

has also been compared graphically. As shown in Figure 4-9, the results have been



59

displayed only for CT Liver image for demonstration purpose. In terms of both PC

and CE measures, superiority of the proposed IGFCM algorithm is immediately

evident from respective graphs. The performance advantage is even more in case of

CE as compared to PC. Hence, it can be concluded reasonably that IGFCM performs

better segmentation than FCM and various variants of FCM, especially in the

presence of noise.

Computational Analysis of the proposed IGFCM Algorithm

The storage and computational complexity of any segmentation technique is an

important factor, especially when two competitor methods have very comparable

segmentation results. In this regard, the proposed IGFCM algorithm does not require

any additional storage capacity as no memory-intensive data structure has been

introduced. Similarly, the computational performance of the proposed algorithm is

also comparable to other employed fuzzy algorithms except FCM. However, FCM

suffers intensively in the presence of noise; therefore, its computational superiority is

not justifiable. All other employed fuzzy algorithms process the input image in a way

similar to IGFCM i.e. they convolve the input image with a window of certain size.

Moreover, due to the information gain in the convergence process, the proposed

IGFCM algorithm converges in less number of iterations than other algorithms. The

computation of information gain is an additional step in the proposed algorithm.

However, the cost of computing information gain is balanced by a faster convergence

rate of IGFCM. Finally, in several offline medical applications, segmentation

accuracy is the main concern rather than computational complexity. Therefore, our

technique is preferable to other techniques as the segmentation accuracy results

presented in the paper verify its superiority.

As an improvement to the proposed IGFCM algorithm from computational

perspective, we conjecture that a parallel implementation similar to [88] is also

possible. In [88], authors have presented a non-rigid 3D medical image registration

framework and a high-speed GPU-based parallel implementation of the framework

has also been proposed. GPU-based implementation of FCM is also a part of this

framework. The proposed IGFCM approach is consistent with FCM implementation

presented in [88]. Parallel implementation of FCM in [88] spreads the processing

across several kernels, each one performing certain processing. Different steps



60

involved in the proposed IGFCM technique can also be parallelized in a similar way.

Ideally, additional kernel(s) may be introduced related to the computation of

information gain. Hence, the GPU-based implementation may drastically increase the

computational performance of the proposed IGFCM algorithm [36].

(a) (b) (c)

(d) (e) (f)

Figure 4-8 CT liver image segmentation, (a) original CT liver image, (b) FCM


segmentation, and (f) the proposed IGFCM segmentation.

(a) (b)

Figure 4-9 Graphical comparison of clustering quality measures (a) PC and (b)

CE for the CT Liver image



61

Table 4.1 Clustering quality comparison of the proposed IGFCM and other

segmentation techniques

Partitioning Coefficient (PC) Classification Entropy (CE)

Original 2 0.01

2 0.02

2 0.03

Original 2 0.01

2 0.02

2 0.03

Synthetic Image

FCM - 0.9667 0.9659 0.9657 - 0.0702 0.0714 0.0721

sFCM - 0.9918 0.9916 0.9915 - 0.0135 0.0140 0.0142

sFCMM - 0.9913 0.9912 0.9911 - 0.0140 0.0143 0.0147

FLICM - 0.8866 0.8446 0.8107 - 0.2485 0.3237 0.3811

IGFCM - 0.9915 0.9914 0.9912 - 0.0138 0.0142 0.0145

The Wolf Image

FCM 0.9660 0.9655 0.9648 0.8800 0.1000 0.1201 0.1441 0.1810

sFCM 0.9831 0.9827 0.9815 0.9807 0.0164 0.0280 0.0296 0.0349

sFCMM 0.9840 0.9588 0.9543 0.9497 0.0195 0.0740 0.0820 0.0903

FLICM 0.9600 0.8724 0.8200 0.7906 0.1032 0.2663 0.3523 0.3984

IGFCM 0.9847 0.9833 0.9823 0.9812 0.0142 0.0200 0.0284 0.0314

Liver CT Image

FCM 0.9251 0.8300 0.8100 0.8000 0.1370 0.3000 0.3000 0.3200

sFCM 0.9355 0.8670 0.8662 0.8627 0.1201 0.2345 0.2357 0.2417

sFCMM 0.9380 0.9041 0.9001 0.8891 0.1559 0.2013 0.2130 0.2388

FLICM 0.9417 0.7384 0.6625 0.6150 0. 1015 0.4843 0.6037 0.6767

IGFCM 0.9451 0.9123 0.9013 0.8910 0.1002 0.1401 0.1500 0.1633

Carotid Artery Ultrasound Image [Figure 4-10]

FCM 0.7850 0.7735 0.7509 0.7297 0.3719 0.3886 0.4109 0.4373

sFCM 0.8461 0.8209 0.7891 0.7722 0.2850 0.3162 0.3409 0.3796

sFCMM 0.8662 0.8409 0.8202 0.7980 0.2150 0.2594 0.2773 0.2862

FLICM 0.8000 0.7566 0.7141 0.6803 0.3833 0.4580 0.5267 0.5799

IGFCM 0.8873 0.8497 0.8364 0.8145 0. 1925 0.2472 0.2633 0.2686

4.3.2 Scenario-II: Segmentation and Decision Performance of IGFCM on


In second scenario, the proposed technique is applied to segment a dataset (Chapter 2)

of 300 carotid artery ultrasound images. Ultrasound is one of the most commonly

used medical imaging modalities. A large number of applications of US in medical

imagery are due to its radiation-safe and cost-effective operation. We have segmented



62

carotid artery ultrasound images using FCM, sFCM, sFCMM, FLICM, and the

proposed IGFCM techniques into three clusters, namely, the arterial wall, an area

inside the artery, and the background tissues. Figure 4-10(a) shows region of interest

(ROI) for one of the original carotid artery ultrasound images. Figure 4-10(b-e) show

that in all cases FCM and its variants have produced a clustering with misclassified

pixels. The IGFCM segmented image on the other hand, produces homogenous

clustering as shown in Figure 4-10(f).

A keen observation of Figure 4-10 reveals that images segmented by variants

of FCM contain most misclassified pixels in the background and arterial walls

regions. Misclassification of arterial wall tissues cannot be tolerated as accurate as

IMT values requires accurate segmentation of arterial wall. On the other hand, the

image segmented by IGFCM has fewer misclassified pixels at arterial walls and

background area compared with other techniques. Figure 4-11(a-l) shows the

proposed approach segmentation at various intensities of Gaussian noise. It can be

seen from Figure 4-11 that the proposed IGFCM has successfully segmented carotid

artery ultrasound image. Further, very few misclassifications show the robustness of

the IGFCM approach.

Figure 4-12(a) and (b) show PC and CE measures for the various carotid

artery images obtained from the database of 300 images. For convenient comparison,

the results have been shown on an interval of 20 images, resulting into a 15 images.

However, average results for all 300 images are also compared for different

segmentation techniques in Figure 4-13. The results in the Figure 4-12 and Figure

4-13 verify that the proposed IGFCM approach outperforms other variants of FCM.

The proposed IGFCM approach also yields more consistent results as evident by the

standard deviation values in Figure 4-13.

A decision system for classification of carotid artery ultrasound images into

normal/abnormal subjects has also been proposed in this work. The classification is

performed based on features computed from IMT values which in turn, are measured

from segmented carotid artery images. Precise IMT measurement, which is based on

accurate segmentation, consequently yields a better decision accuracy. Therefore, for



63

our proposed decision making system, IMT values have been measured from the

carotid artery US images segmented by IGFCM algorithm.

(a) (b) (c)

(d) (e) (f)

Figure 4-10 Carotid artery US image segmentation, (a) original carotid artery US

image, (b) FCM segmentation, (c) sFCM segmentation, (d) sFCMM

segmentation, (e) FLICM segmentation, and (f) the proposed IGFCM

segmentation

(a) Original carotid artery ultrasound

image (b) IGFCM segmentation

(c) Gaussian noise of variance 0.01 (d) IGFCM segmentation



64

(e) Gaussian noise of variance 0.02 (f) IGFCM segmentation

(g) Gaussian noise of variance 0.03 (h) IGFCM segmentation

(i) Original carotid artery image

with marked plaque

(j) IGFCM segmentation

(k) Gaussian noise of variance 0.01 (l) IGFCM segmentation

Figure 4-11 IGFCM segmentation at noise free and noisy carotid artery ultrasound

image corrupted through Gaussian noise of various intensities.

With the consultation of a medical expert, IMT values for upper and lower

arterial walls of carotid artery have been measured from each segmented image.

Figure 4-14 shows the measured IMT curve for one of the normal and abnormal

Plaque

Plaque



65

carotid artery, respectively. IMT mean and standard deviation for the normal carotid

artery are 0.3824 mm and 0.1046 mm, respectively and 0.7630 mm and 0.1702 mm for

the abnormal carotid artery, respectively. A feature vector comprising 6 features (see

Section 6.2.4 ) has been constructed from IMT values and has been fed as input to

PNN classifier for training and validation.

(a) (b)

Figure 4-12 (a) PC and (b) CE measures of 300 carotid artery ultrasound images

Figure 4-13 Average PC and CE measures for 300 carotid artery ultrasound images

Classification performances of PNN, KNN, and MLBPNN using the features

generated from IGFCM have been evaluated. Table 4.2 shows the performance

comparison in terms of various quantitative measures for PNN, KNN and MLBPNN

classifiers. Chaudhry et al. [85] and Santhiyakumari et al. [32] have reported 98.10%,

and 96% classification accuracy, respectively, on their respective datasets. However,



66

PNN classifier achieved 98.40% classification accuracy using the measures extracted

from the proposed IGFCM technique. This accuracy is comparable to 98.4% reported

by Hassan et al.[59]. Hence, the positive influence of our proposed technique on the

classification process is evident.

Figure 4-15 shows ROC curves for PNN, KNN, and MLBPNN classifiers. It

can be observed from the obtained ROC curves that PNN ROC curve is very close to

vertical axis, which signifies fewer misclassifications as compared with KNN and

MLBPNN classifiers. AUC value for PNN has been computed as 0.982, which

indicates a high diagnostic accuracy. These statistical results indicate that the

proposed decision system has effectively discriminated the normal and abnormal

subjects using the features from IGFCM.

Finally, we have also compared the effects of different segmentation

techniques on classification performance. In particular, FCM, sFCM, sFCMM,

FLICM, and IGFCM have been applied to 300 carotid artery images and segmented

images have been classified by PNN, KNN and MLBPNN classifiers. Figure 4-16

shows that highest classification accuracy is obtained by segmenting carotid artery

ultrasound images using the proposed IGFCM and classifying through PNN classifier,

which we have suggested in our proposed decision system.

By critically analyzing the classification results presented in Figure 4-16, one

can observe that the classification accuracy has been improved by 3.27%, 3.80% and

4.31% for PNN, KNN, and MLBPNN classifiers (see Table 4.3), respectively. This

significant classification improvement is due to the accurate segmentation of carotid

artery ultrasound images by the proposed IGFCM technique. As the classification

phase is highly dependent on segmentation and hence correct IMT measurements

which is derived from accurate segmentation. Hence, it can be concluded that the

proposed segmentation based classification scheme can successfully be applied to

identify the presence of plaque in carotid artery.

Hence, the proposed IGFCM algorithm boosts the classification accuracy and can be

used as part of the proposed decision system.



67

Figure 4-14 IMT measurements of a normal and abnormal carotid artery

Table 4.2 Performance comparison of PNN-based decision system with other

classifiers

Classifiers Classification performance

Accuracy MCC F-Score Sensitivity Specificity

PNN 98.40% 0.9600 0.9799 0.9839 0.9762

KNN 98.10% 0.9520 0.9709 0.9742 0.9725

MLBPNN 95.20% 0.8600 0.9244 0.9735 0.8905

Figure 4-15 ROC curve plotted against true positive rate vs. false positive rate for

PNN, KNN, and MLBPNN classifiers.



68

Table 4.3 The effect of segmentation technique at classification of carotid artery

ultrasound images

Classification Techniques

PNN KNN MLBPNN

Segmentation Techniques Accuracy (%)

FCM 95.20 94.37 91.10

sFCM 96.91 95.87 94.27

sFCMM 97.80 95.93 94.71

FLICM 97.84 97.30 95.10

Proposed IGFCM 98.40 98.10 95.20

Figure 4-16 Effect of segmentation techniques on the classification of carotid artery

images

Summary

In this chapter, a robust technique of segmentation, called IGFCM, has been

proposed. The proposed algorithm incorporates the concept of information gain into

fuzzy framework to overcome the shortcomings of conventional FCM algorithm. The

proposed IGFCM algorithm has been applied to segment images from different

modalities and compared with conventional FCM and three different variants of FCM.

Gaussian noise of various intensities is added to images in order to check robustness

of the proposed technique. Quantitative measures (PC and CE) computed from



69

segmented images of different modalities verify the robustness and effectiveness of

the proposed algorithm over other techniques.

Afterward, a decision system has been designed based on PNN classifier using

a dataset of 300 real carotid artery ultrasound images. The proposed decision system

has successfully distinguished the abnormal subjects from normal ones, with accuracy

of 98.40%. In addition to this, the effect of segmentation on classification has also

been investigated. The classification accuracy based on the proposed segmentation

technique has been improved by 4.31% compared with other segmentation techniques

using PNN classifier. The significant improvement in classification accuracy is due to

accurate segmentation of carotid artery ultrasound images by the proposed IGFCM

approach.

In this chapter, a robust clustering technique is proposed to segment carotid

artery ultrasound images. It has been observed from the experimental results of the

proposed IGFCM technique that there still exist some incorrect labeled patterns,

especially, on arterial walls, which is the most important part to be segmented

correctly, since IMT values are measured from these arterial walls. In order to

overcome this limitation, it is desirable to propose a technique, which minimizes the

incorrect clustering in arterial wall tissues. For this purpose, a semi-supervised

clustering technique is developed to correctly cluster the patterns present in carotid

artery ultrasound images.


1. Robust Information Gain based Fuzzy C-Means Clustering and Classification

of Carotid Artery Ultrasound Images, “Computer Methods and Program in

Biomedicine, Journal” (IF=1.555), 2013.

2. Medical Image Segmentation Employing Information Gain and Fuzzy C-Means

Algorithm, IEEE ICOSST, Lahore, 2013.

70

Chapter 5 : Robust Segmentation of Carotid

Artery Ultrasound Images based on Neuro

Fuzzy GA and Expectation Maximization

In this chapter, a new robust segmentation and classification of carotid artery

ultrasound images (RSC-US) is proposed. The proposed hybrid approach employed

neuro fuzzy, expectation maximization (EM) and GA techniques and hence has been

used to segment carotid artery ultrasound images. Initially, class data are selected by

the user and these labels are refined by the expectation maximization algorithm.

Further, fuzzy inference system has been generated based on GA feature selection

mechanism. An intelligent decision making system based on SVM has also been

proposed to classify the segmented carotid artery ultrasound images into normal and

abnormal subjects. Hence, the proposed approach can be used as a secondary

observation to an experienced radiologist.

5.1 The Proposed RSC-US Technique

The prime objective of the proposed RSC-US scheme is to segment the carotid artery

ultrasound images with high accuracy. IMT values have been measured by carrying

out the proposed segmentation approach. Further, SVM classifier has been employed

to classify a real carotid artery images dataset into normal and abnormal subjects.

Due to inherited degradations in ultrasound images, high pixel correlation, and

partial volume effects, it is very hard to identify the correct label for each pixel. For

accurate and robust segmentation, it is highly required to train the model in such a

way that it can segment the image with high accuracy, even in the presence of noise.

The EM algorithm has been used to refine the initial labels of each class

assigned by the user. MGH based image features have been extracted, most relevant

and useful image descriptors are selected by GA. Selected image descriptors and

targets have been given as input to a neural network for generation of FIS. The

generated FIS has further been used to segment the carotid artery ultrasound images.

IMT values have been measured from the proposed RSC-US approach segmented

Chapter 5: Robust Segmentation of Carotid Artery Ultrasound Images based on Neuro Fuzzy GA and

Expectation Maximization

71

images. Based on IMT values, the segmented images have been classified into normal

or abnormal subjects using SVM classifier. Segmentation and classification results are

compared with a few state of the art techniques. The obtained segmentation and

classification results indicate the effectiveness of the proposed approach. Further, the

effect of segmentation at classification has also been investigated. It has been

observed that quality segmentation has high impact at classification phase. The flow

chart of the proposed RSC-US approach is depicted in Figure 5-1. Each phase of the

proposed approach explained in the following sub-sections.

Figure 5-1 Graphical representation of the proposed RSC-US approach

5.1.1 Label Initialization of Pixels

In first phase of the proposed technique, initial labels have been assigned to

pixels by the user. Due to overlapping nature of medical images, user assigned labels

may not be accurate and require further processing. Accurate data labels leads

towards better segmentation, therefore, these initial labels were further processed by

employing EM algorithm. Figure 5-2 shows the sample carotid artery ultrasound

image in which initial class labels have been selected. Carotid artery ultrasound

images mainly have three clusters; the arterial wall, an area inside artery, and the



72

background tissues. It can be observed from Figure 5-2 that the initial selected labels

have the data of other regions. Especially, some background region has been merged

into the artery wall (ROI). Arterial walls are significant for classification; therefore

these walls need to be separated from background with high accuracy.

Owing to narrow region of arterial walls, the probability of merging pixels of

other classes into arterial walls is higher. Because, the other regions like background

tissues and area inside arterial walls have large enough and the probability of merging

these areas is lower. Now, the question is how these initial targets should be refined to

find accurate class labels. For this purpose, EM algorithm has been used to handle this

issue.

Figure 5-2 Initial targets marked as artery wall, area inside the artery and the

background tissues.

5.1.2 The Expectation Maximization Step

The expectation maximization algorithm is widely used to estimate the statistical

models. It has the advantage of being simple, robust, and easy to implement. The

objective of using EM algorithm is to find iterative convergence of a set of input

projections on the most likely reconstruction. The goal is to estimate the parameters of

Gaussian mixture model (GMM) by employing EM algorithm, hence used for target

class estimation. The unknown parameter θ of the model needs to be estimated in a

Artery wall

Area inside artery

Background tissues



73

way that |p X becomes maximize. Log likelihood function is used to estimate the

parameter θ as described in Expression (5.1).

ln |L p X (5-1)

At given data X, the parameter can be considered as a function of log

likelihood. The logarithmic function is strictly increasing and so that value of

maximizes |p X . Resultantly, the L will become maximum. EM algorithm

iteratively maximizes L . Further, it is assumed that after last iteration, the current

estimate of becomes n .

nL L (5-2)

Mathematically, the difference should have to be maximized using equation (5-3),

ln | ln |n nL L p p X X (5-3)

The sole objective is to select the values of θ parameter, where L get maximized.

EM algorithm is used to select the θ at maximum | nL . The updated value is

represented by n and finally got,

1arg max |

arg max | , ln | , | | | ,

n n

n n

z

n n

l

L p z p z p z p p z

X X X X

The constant terms w.r.t. θ needs to be dropped.

arg max | , ln | , |

arg max | , ln

arg max | , ln , |

, , , ,

n

z

n

z

n

z

p z p z p z

p z

p z p z

p z p z p z p

X X

X

X X

X

| ,arg max ln , |

n

E p z

Z X

X (5-4)



74

The EM algorithm has two steps, one is expectation and other is maximization. These

steps are represented in Equation (5-4). EM algorithm thus consists of the following

steps:

1. Expectation-step, which is used to find the conditional expectation

| , ln , |n

E p z Z X X .

2. Maximization-step, which maximizes the step 1, w.r.t.θ.

Further detail and derivation of EM algorithm can be found in [89, 90]. Mean

and standard deviation for GMM model has been estimated using EM algorithm.

These estimated parameters are then used in the following expressions to find more

refined labels for the input data.

2

1

for 1, 2, ...,

n

ik

j

ij k kz k lx

(5-5)

where ijx is the jth element of the ith input data point and k is the number of class,

and are the mean and standard deviation respectively estimated by the EM

algorithm. The predicted labels are calculated by using the following expression:

1

1 ek i f

ik d

yz f

X (5-6)

where yk represents the membership of each pixel to a certain class k and zik values are

obtained from equation (5-5). The variables fe and fd are used for controlling the

fuzziness amount and these values are set empirically [91]. It is obvious that the

output of yk will be in the interval of 0 and 1. Using the maximum value of index yk,

we get the target label associated with each pixel. The graphical representation GMM

estimated by EM algorithm is presented in Figure 5-3, where, horizontal axis ( x )

shows the number of data points in association with its probability density function at

vertical axis, and the dotted curves represents data clusters.

5.1.3 Feature Extraction

The objective of this phase is to formulate a feature vector for every pixel and thus

used for segmentation. In this study, feature extraction strategy based on MGH has



75

been used. These extracted features are advantageous to represent the medical images

[12].

Figure 5-3 GMM parameters estimation using EM algorithm

Moments of Gray Level Histogram

Statistical image descriptors are being used in image analysis [12]. These features are

based on the intensity histogram and are extracted for every pixel. The nth order

statistical moment about mean can be computed by using the following equation;

1

0

( ) . ( )

Ln

n i i

i

u z m p z

(5-7)

where ip z is the histogram of image intensity level in a certain region, and z shows

the intensity. L represents the possible intensity levels and m is an average intensity

value.

A total of nine MGH features are extracted from the input image.

Mathematical description of these features is given in Section 3.1.2. A fixed window

of size (5x5) is used for feature extraction and important features are selected using

GA and thus used for segmentation of carotid artery ultrasound images.

Feature Selection

Feature selection has been carried out to reduce the feature vector dimensionality

while preserving the accuracy. In order to improve computational cost and save

-50 0 50 100 150 200 250 3000

0.002

0.004

0.006

0.008

0.01

0.012

0.014

0.016

0.018

0.02

X

Pro

ba

bility d

en

sity

Gaussian Mixture estimated by EM



76

valuable resources, irrelevant and redundant features need to be omitted using some

searching algorithm. Many researchers have explored various feature selection

strategies such as sequential backward selection, greedy stepwise, forward selection,

GA, and swarm optimization etc. Whereby, each searching technique has its own

advantages and disadvantages [13].

In this study, GA has been used to select discriminating image descriptors

from the already extracted feature set, hence used for segmentation. WEKA 3.6, a

machine learning software has been employed for this purpose [72]. Same GA

parameters have been used as mentioned in Section 3.1.3. Only 4 out of 9 features

have been selected by GA hence was used for segmentation, which obviously saves

time and resources. The following features namely; FM2, FM4, FM5 and FM6 have

been selected by GA and thus used for segmentation.

5.1.4 Neuro Fuzzy Classifier

In image segmentation, the effectiveness of the NFC approach has been analyzed,

hence was used for segmentation of carotid artery ultrasound images. The basic

objective of fuzzy classification is to segment out the feature space into specified

fuzzy classes. In medical images, as the regions are overlapped with each other, it is

quite possible that a pixel may belong to more than one class with different degree of

membership. In this scenario, fuzzy approaches might be suitable approaches [92,

93]. Like other fuzzy techniques, in NFC, fuzzy if then else rules are applied to

develop a classifier. NFC based segmentation needs a FIS which is actually a

collection of fuzzy rules.

NFC is a supervised learning approach, its structure (i.e. input, hidden and

output layers) is similar to other neural networks. Input to the classifier is optimized

features selected by GA and a hidden layer of fifty neurons. Output layer consists of

three neurons, which represents three different classes of carotid artery ultrasound

image. Neural networks have capability to learn the system behavior and can

successfully generate the fuzzy “if then else rules” and fuzzy membership functions.

The general structure of NFC is as below:

Input layer: Four neurons

Hidden layer: Fifty neurons



77

Output layer: Three neurons

The aforementioned neutral network architecture is used for generation of FIS and the

generated FIS is thus used for segmentation. The hidden layer neurons are set

empirically and at fifty hidden layer neurons the network offers superior performance.

Training and Testing of NFC

After the selection of important image descriptors using GA, NFC training and

validation procedure starts. Class labels have been assigned to every pixel by EM

algorithm. For training and validation of NFC, 18 different carotid artery images have

been utilized from obtained dataset. Training and validation process was performed at

the ratios of 67 and 33 percent. The FIS was generated by NFC and hence used for

segmentation of carotid artery ultrasound images. Root mean square error (RMSE)

has measured by the difference between actual and predicted values Figure 5-4 shows

the RMSE curve plotted against epochs of NFC training. From Figure 5-4, it can be

observed that NFC converges when error gradient becomes very small. RMSE shows

monotonically decreasing behavior, therefore error rate decreases with increasing

number of epochs. The NFC was trained at 400 epochs so that it can learn the patterns

effectively.

Figure 5-4 Behavior of RMSE curve of NFC training at different epochs

FIS has been generated and is used for segmentation of carotid artery

ultrasound images. The proposed RSC-US approach performance is evaluated by a

significant amount of real carotid artery ultrasound images, dataset details is given in

Chapter 2. The proposed RSC-US generated FIS is shown in Figure 5-5. The

horizontal axis shows the number of selected image descriptors, whereas the vertical



78

axis shows the number of fuzzy rules used for classification of the input data. The

input feature values were scaled into the range of 0 and 1. The index line in Figure 5-5

indicated that which rule might be activated for input data and it may change for

every input instances. The shaded region of the membership function shows visually

apparent fuzzy membership value. The fuzzy rule view is used to interpret the overall

fuzzy inference process at once. There are three fuzzy rules incorporated in FIS,

which are further used for classification given input feature vector. The graphical

representation of fuzzy membership functions for the selected feature 1 and feature 2

are shown in Figure 5-6. Similarly, visual representation of fuzzy rules for feature 3

and 4 can also be generated.

Figure 5-5 NFC generated FIS used for segmentation of carotid artery ultrasound

images.

5.1 Decision Making System for Carotid Artery Ultrasound Images

Carotid artery ultrasound images segmented by the proposed RSC-US approach need

to be classified into normal and abnormal subjects. For this purpose, IMT values have

been measured from segmented images. Based on measured IMT values, a feature

vector is formed and given as an input to SVM classifier. Since, SVM has advantages

over other classifiers because of its enhanced learning capability. Hence, SVM is used

for classification of segmented images.

Five different commonly used features namely Mean, Variance, Standard

Deviation, Max., and Min. are extracted from measured IMT values and are feed as



79

input to the classifier. The extracted feature values are normalized into the range of 0

to 1. By utilizing these features, SVM offers superior classification accuracy.

a) b)

Figure 5-6 Graphical representation of fuzzy membership functions for (a) selected

feature 1 and (b) selected feature 2

5.2.1 Classification Performance Measures

To evaluate the performance of the classifier, accuracy, MCC, sensitivity, specificity,

F-score, ROC and AUC measures have been computed, and details of these measures

is described in Section 2.6.


Real carotid artery ultrasound images have been segmented by the proposed RSC-US

approach. Details of the dataset can be found in Chapter 2. The obtained dataset is

segmented by the proposed RSC-US technique. To confirm robustness of the

proposed approach, it has been evaluated by considering various intensities of

Gaussian noise. The results of the proposed technique have been compared with few

state of the art approaches. The proposed RSC-US technique outperforms both at

noisy and noise free images.

The Figure 5-7(a) shows one of the original longitudinal carotid artery

ultrasound images with marked plaque. Selected ROI is shown in Figure 5-7(b).

Segmented carotid artery ultrasound image by the proposed technique is presented in

Figure 5-7(c). The proposed technique successfully separated plaque in carotid artery

with minimum mislabeled patterns as compared with other techniques. Results

obtained through the proposed technique are also compared with few exiting state of

the art techniques. Figure 5-7(d), 5-7 (e) and Figure 5-7(f) show the images

segmented by the FCM, K-means, and sFCMLSM methods, respectively. From visual



80

inspection, superiority of the proposed RSC-US approach is clear. On the other side,

segmented images by FCM, K-means and sFCMLSM segmented images contain

much misclassified patterns especially in marked plaque area. Further, IMT has been

measured from segmented images to form a feature vector and have been fed as input

to the SVM classifier. Therefore, image post-processing techniques are highly

dependent on segmentation quality. If carotid artery segmented images contain

misclassifications, IMT value may not be accurately measured. Consequently, the

effect of the inaccurate IMT measure produces misleading results. Particularly when

dealing with medical images a great care is needed because wrong predictions may

not be affordable in this field.

(a) (b)

(c) (d)

(e) (f)

Figure 5-7 a) One of the longitudinal original carotid artery ultrasound images with

marked plaque, b) selected ROI c) segmented by the proposed RSC-US scheme, d)

FCM segmentation, e) K-means and f) Magenta represents initial and green represents

final segmentation of sFCMLSM technique.

Magenta: Initial; Green: Final

Plaque

Plaque



81

The proposed technique’s performance robustness has also been assessed. The

carotid artery ultrasound images have been degraded by white Gaussian noise of

various intensities. Keeping in mind, medical images may not contain high intensity

of noise, beside this; we have evaluated the proposed technique at Gaussian noises of

0.01, 0.015, 0.05 and 0.1. Some sample noisy carotid artery images segmented by the

proposed RSC-US approach with Gaussian noise of variance 0.05 are shown in Figure

5-8. From Figure 5-8, it can be observed that the proposed approach has successfully

segmented images even in the presence of noise.

a) Original carotid artery image with selected ROI

b) Noisy carotid artery image with Gaussian noise of 0.05 variance

c) The proposed approach segmented image d) The proposed approach segmented image

e) Original carotid artery image with selected

ROI f) Noisy carotid artery image with Gaussian

noise of 0.05 variance



82

g) The proposed approach segmented image h) The proposed approach segmented image

Figure 5-8 Sample noise free and noisy longitudinal carotid artery ultrasound images

segmented by the proposed RSC-US approach.

(a) (b)

(c) (d)

(e) (f)

Figure 5-9 (a) One of the original longitudinal carotid artery ultrasound images, b)

image corrupted by Gaussian noise of variance 0.10 with marked ROI c) the proposed

RSC-US technique segmented image; d) FCM segmentation e) K-means and f)

sFCMLSM segmentation.

Figure 5-9(a) shows one of the original longitudinal carotid artery ultrasound

images. The image has been corrupted with Gaussian white noise of variance 0.10 and

selected ROI is shown in Figure 5-9(b). Images segmented using the proposed RSC-

US, FCM, K-means, and sFCMLSM approaches are shown in Figure 5-9(c), 5-9(d),

Plaque Plaque

Plaque



83

5-9(e), and Figure 5-9 (f), respectively. Apparently, it can be observed that the image

segmented by the proposed approach outperforms all mentioned techniques, with

minimum misclassifications even in the presence of noise. In Figure 5-9(d), the FCM

segmented image has a lot of misclassified regions, hence IMT values measurement

might be incorrect. In other segmentation approaches, such as K-means, Figure 5-9(e)

and sFCMLSM Figure 5-9(f), we can also observe the existence of misclassified

pixels. These noisy patterns may mislead IMT measurements and consequently, may

increase the false detection of plaque in carotid artery. Segmentation performance,

even in the presence of noise, shows the robustness and effectiveness of the RSC-US

approach.

5.3.1 Performance Analysis of the Proposed Segmentation Technique

Apparently, we have observed that the proposed RSC-US technique outperforms

other segmentation techniques at various noise levels. To verify the superiority of the

proposed approach qualitatively, Davies Bouldin Index (DBI) of segmented images

has been computed. Smaller the value of DBI is, the betters the clustering quality is

[58]. Using DBI clustering quality measure, Table 5.1 shows the average

performance of different techniques for obtained dataset of real carotid artery

ultrasound images on various noise levels. From Table 5.1, it can be observed that the

proposed RSC-US outperforms the mentioned techniques at all given noise levels.

From the results presented in Table 5.1, a significant fact established is that the noise

has minimum effect on the proposed approach compared with the other considered

techniques. Hence, RSC-US segmentation shows the superiority in terms of

quantitative quality measures. The graphical representation of segmentation quality

measure based on DBI at various noise levels is shown in Figure 5-10. It is evident

from Figure 5-10 that the proposed segmentation approach works efficiently also in

the presence of noise.

Segmentation time (in sec.) of the proposed RSC-US technique across the

increasing number of features is shown in Figure 5-11. Optimal performance has been

obtained by utilizing only four important features. Less segmentation time and

resources are required at selected feature set. It is evident from Table 5.1, and Table

5.2 that the proposed approach at reduced features maintains accuracy; hence offers



84

superior segmentation and classification results. Thus, the proposed RSC-US

outperforms the other segmentation techniques and shows its effectiveness.

5.3.2 SVM based Decision Making System

Accurate segmentation would surely be effective for post-processing techniques. To

investigate this, we have evaluated classification performance using different

segmentation algorithms. To classify the segmented carotid artery image at hand as

normal or abnormal, IMT values have been measured from segmented carotid artery

images. Successful segregation of ROI contributes a lot to IMT measurements with

higher level of confidence. Consequently, its affect will be in the form of better

classification. IMT mean and standard deviation of the normal carotid artery

ultrasound images dataset used in current research are 0.392 mm and 0.109 mm

respectively. Whereas, mean and standard deviation of IMT values for abnormal

carotid artery ultrasound images are 0.745 mm and 0.173 mm, respectively. Figure

5-12(a) and (b) show the graphical representation of the IMT measurements for the

normal and abnormal carotid artery ultrasound images, respectively.

Table 5.1 Average performance comparison of the proposed RSC-US and other

techniques over 300 carotid artery ultrasound images

Segmentation Techniques

FCM sFCM K-means sFCMLSM RSC-

US

Gaussian noise levels Davies Bouldin Index (DBI)

Original image 0.4653 0.5660 0.4640 0.4710 0.4379

0.01 0.4772 0.5921 0.4739 0.4835 0.4417

0.02 0.4922 0.6341 0.4911 0.5002 0.4527

0.05 0.5213 0.7971 0.5171 0.5514 0.4783

0.10 0.5926 0.9582 0.6041 0.6127 0.4826



85

Figure 5-10 Segmentation quality comparison of different techniques at various

noise levels

Figure 5-11 Computational time (in Sec.) of the proposed scheme across different

number of extracted features.

In this chapter, KNN, MLBPNN, and SVM classifiers have been used for

separation of normal and abnormal subjects. To choose an optimum classifier for

carotid artery ultrasound images, performance of each classifier in terms of various

quality measures has been analyzed on the proposed approach segmentation results.

Table 5.2, shows the performance comparison of SVM, MLBPNN, and KNN

approach at various classification quality parameters. We have achieved 98.84%

classification accuracy by SVM which shows the effectiveness of the proposed

segmentation technique. Classification accuracy based on the proposed RSC-US

segmentation approach has also been compared with recent published techniques.

Santhiyakumari et al. [32] have used MLBPNN for classification and reported 96%

classification accuracy at their dataset. Hassan et al. [36] have employed PNN to

classify the normal and abnormal subjects and have reported 98.4% classification



86

accuracy. Hassan et al. [59] have employed MLBPNN for classification and achieved

98.4% classification accuracy at their dataset. Whereas, using the current study’s

segmentation approach, classification accuracy of 98.84% is obtained by SVM

classifier. The results indicate that classification using the proposed technique

segmentation outperform the other techniques and hence shows the usefulness of the

proposed RSC-US segmentation approach. Graphically, a comparison of various

classifiers at different validity measures has also been shown in Figure 5-13. It can be

perceived from Figure 5-13 that SVM outperforms the other classification techniques.

Further, a comparison of different classifiers in term of ROC and AUC has

also been made. Figure 5-14 shows ROC curves for SVM, MLBPNN and KNN

classifiers. From Figure 5-14, it can be seen that SVM’s ROC is close to vertical axis,

which depicts high classification accuracy. Obtained AUC value 0.98 shows the

overall better diagnostic test and confirmed the advantage of the proposed scheme.

(a) (b)

Figure 5-12 (a) IMT measurement of a normal and (b) abnormal carotid artery

Table 5.2 Classification performance comparisons of SVM, KNN and MLBPNN

using the proposed RSC-US segmentation technique

Techniques Accuracy (%) F-Score MCC Sensitivity Specificity

KNN 98.30 0.9700 0.9483 0.9697 0.9486

MLBPNN 98.40 0.9551 0.9127 0.9750 0.9385

SVM 98.84 0.9880 0.9767 1.0000 0.9773



87

Figure 5-13 Performance comparisons of KNN, MLBPNN and SVM classifiers using

RSC-US segmentation at various classification quality measures

Effects of Segmentation on Classification

The effects of segmentation on classification phase have been explored and it has

been observed that the proposed segmentation technique based classification offers

superior performance. As described earlier, the post-processing techniques are highly

dependent upon quality segmentation. In this research work, images have been

segmented by different techniques namely; FCM, K-means, sFCMLSM and the

proposed RSC-US algorithms. Statistical analysis of classification results reveals that

performance accuracy has been improved by a margin of 2.47-4.00 % (Figure 5-15

and Table 5.3) based on the proposed technique based segmentation. This significant

improvement in classification accuracy is largely due to the correct segmentation of

carotid artery ultrasound images.

Keeping in view of the segmentation and classification accuracy, the proposed

RSC-US technique can be used as secondary observer to identify the plaque in carotid

artery. Further, in countries like Pakistan, it can also be used in remote areas with less

experienced technicians for initial screening of plaque buildup in carotid artery based

on the ultrasound imaging.



88

Figure 5-14 ROC curves of SVM, MLBPNN and KNN.

Table 5.3 Effects of segmentation on object classification

Classification Techniques

SVM KNN MLBPNN

Segmentation

Techniques Accuracy (%)

FCM 96.40 95.50 95.91

K-Means 95.20 94.37 94.63

sFCMLSM 96.80 95.78 96.00

Proposed RSC-US 98.84 98.30 98.40

Figure 5-15 Comparison of classification accuracy for various segmentation

techniques

Summary

In this chapter, a robust segmentation and classification technique has been proposed.

The proposed technique employed neuro fuzzy classifier, EM, and GA. FIS has been

generated which utilizes optimal features. Obtained segmentation results of the



89

proposed technique have been compared visually and quantitatively on noisy and

original images with FCM, K-means, and sFCMLSM approaches. The segmented

carotid artery ultrasound images are classified into normal or abnormal subjects by

SVM classifier. SVM offers high classification accuracy 98.84%. In addition to this,

effect of segmentation on classification has also been investigated and classification

results have been improved by a margin of 2.47-4.00 % (see Figure 5-15 and Table

5.3)

The proposed RSC-US technique successfully segmented the carotid artery

ultrasound images. In the development of proposed RSC-US technique, spatial

information of a pixel under consideration has not been exploited. Spatial information

plays a vital role in segmentation of images particularly when image are polluted with

noisy contents. Therefore, the investigations are extended through the development of

a Robust Fuzzy Radial Basis Function Neural Networks (RFRBFN) in order to offer

better segmentation for medical images.


1. Robust Segmentation of Carotid Artery Ultrasound Images using Neuro Fuzzy

and Expectation Maximization: Employing Intima-Media Thickness and SVM

for Disease Prediction, Submitted in Information Sciences, 2015 Journal.

90

Chapter 6 : Robust Fuzzy RBF Network Based

Segmentation and Decision Making System for


In this chapter, a robust clustering method for accurate segmentation of medical

images and its subsequent intelligent classification has been proposed. Medical

images may have various types of degradations and it is challenging to segment the

degraded images. In order to segment the degraded images, a new approach RFRBFN

has been proposed. For this purpose, the fuzzy RBF algorithm has been modified by

incorporating spatial information and a smoothing parameter into its objective

function, and therefore it intelligently copes with noise related variation. To validate

its effectiveness, the proposed RFRBFN is applied on the segmentation of different

imaging modalities as well as tested against impulse and Gaussian noise of various

intensities. Multi-layer backpropagation neural network is employed to classify the

segmented images into normal or abnormal subjects. The proposed intelligent

decision making system can thus be used as a secondary observer for identification of

plaque in the carotid artery.

6.1 The Proposed Robust Fuzzy RBF Network Approach

As described in Section 2.2.4, clusters centroids of radial basis function network

(RBFN) hidden units are optimized using FCM clustering algorithm. For noise free

images, using the FRBFN approach may offer better segmentation results.

Considering high susceptibility of FCM algorithm to noise, it is expected that FRBF

network may not offer better segmentation results for noisy images.

The main reason for decreased performance of FRBF network especially in

case of noise is that it has to update weights between input and hidden layer without

utilizing the spatial information shown in Equation 2-1. Thus, we incorporate the

spatial information into Equation 2-1 and the modified objective function becomes:

2

1 12p k

C Cq q q

RFCM pk j k pk lm

p k p k l N m M

J u z v u u

(6-1)

Chapter 6: Robust Fuzzy RBF Network Based Segmentation and Decision Making System for Carotid


91

where 1,2,3,...,kM C and pN is the neighborhood of pixel p . controls the

smoothness of membership function pku . Higher values of smooth the membership

functions for a cluster by suppressing noise signal. For 0 , Equation 6-1 becomes a

standard FCM cost function and for the large values of , clustering depends on the

neighborhood rather than the pixel itself. Therefore, an optimal value of is desired

so that the noise is suppressed and consequently better segmentation is performed.

The cost function used in Equation 6-1 was optimized for optimal value of kv

and pku using the method described in [34]. For optimization, we have used Lagrange

multipliers for the enforcement of constraint (1

1C

pk

k

u

) and taking the partial

derivative of Equation 6-1 w.r.t. pku we obtained:

2

1 1 1

21

12

p k

p k

C C Cq q q

pk p k pk lm p pk

p k p k l N m M p kpk

q q

pk p k lm p

l N m M

u z v u u uu

qu z v u

(6-2)

where, p is Lagrange multiplier and the factor 1

2 of vanishes due to the derivative

operator results in a term corresponding to the product of pku and its neighbors plus

its term corresponding to the reverse product of the neighbors and pku . After

obtaining partial derivative and equating it to zero, we get:

1 1

2

p k

q

q

p k lm

l N m M

pk

p

q z v u

u

(6-3)

To solve Equation 6-2 for p , after employing the constraint factor1

1C

pk

k

u

, we

obtain:



92

1 1

2

1

1p k

qC

q

p k lm p

k l N m M

q z v u

(6-4)

p does not depend on k, therefore, it can be factored out of the summation:

1 1121

1 p k

qC

q q

p p k lm

k l N m M

q z v u

(6-5)

Now a combination of Equations 6-3 and 6-5 results in the following necessary

condition for pku to be at locally minimum of RFCMJ :

1 1

2

1 1

2

1

p k

p k

q

q

p k lm

l N m M

pk qC

q

p i lm

i l N m M

z v u

u

z v u

(6-6)

The cost function used in Equation 6-1 was optimized. The optimal value of kv , using

Equation 2-3 and pku,has been obtained by utilizing the method described in [34].

The closed form of equation for membership functions pku , after optimization of

Equation 6-1, becomes:

1

11

12 2

121

2

11

q

p k p k

p k

p k

qq q

p k lm p k lm

l N m M l N m M

pkC

qC qq p r lm

p r lm r l N m Mr l N m M

z v u z v u

u

z v uz v u

(6-7)

The equation used to find cluster centroids in the hidden layer is same as used

in Equation 2-2. If 0 , Equation 6-7 will be reduced to the standard FCM as

mentioned in Equation 2-1.

The cost function used in Equation 6-1 was optimized for optimal value of kv and pku

using the method described in. The closed form of equation for membership functions

pku , after optimization of Equation 6-1 becomes:



93

2

1

1

1

1

p k

p k

qq

pk lm

l N m M

pk

qCq

pr lm

r l N m M

d u

u

d u

(6-8)

The equation used to find cluster centroids in the hidden layer is the same as

used in Equation 2-2. Setting 0 in Equation 6-8 it will reduces into Equation 2-3,

which is the similar used for calculating pku in standard FCM.

The basic block diagram of the proposed method is shown in Figure 6-1. The

description of each section is given in the following subsections.

Figure 6-1 Schematic diagram of the proposed RFRBFN technique

6.1.1 Targets Outputs

Instead of using hard labels as used in RBF, we incorporate the fuzzy set concepts at

the output layer. Consider a k class problem, let each cluster has a mean k and

standard deviation k . Then the target outputs pT z can be calculated using the

following expression:



94

1

1

ep q

pk

d

T zy

q

(6-9)

where eq and dq control fuzziness of the output layer and pky is calculated as:

2

1

npl kl

pk

l kl

zy

(6-10)

where, plz is the value of feature p along dimension l . Similarly, kl is the mean value

of cluster k along dimension l . Equation 6-12 is used when clusters have different

means. If two or more clusters have same (or nearly same) mean then pky should be

computed as proposed by [38]. Following steps are employed to calculate the mean

and the standard deviation of each cluster.

Partition the data into k classes using unsupervised algorithm.

Select the pattern vectors pz for each cluster.

Use expectation maximization (EM) algorithm to find the mean and the

standard deviation of each cluster using pz as an input vector.

Here, we assume that each cluster is normally distributed with mean k and

standard deviation k , respectively. Wavelet based FCM, as proposed by [97], is

employed to perform Step 1. For training of the network, three hundred pixels were

selected from each cluster and were feed as input to the network.

6.1.2 Input Features

In the proposed technique, fuzzy concepts were implemented at input layer

1 2, ,...,pz z z z as symbolized 1 1 1 2 2 2, , , , , ,..., , ,p L M H L M H L M Hz z z z z z z z z z

in Figure 6-1. Where pLz , pMz and pHz represent low, medium and high linguistic

property sets, respectively and were modeled using membership function

introduced in [91]. The membership function is the combination of S and 1 S

shaped membership functions. As modeled in [97], the equation of membership

function becomes:



95

, , , 2

, ,

1 , , , 2

aS z v v a v if z c

z v aa

S z v v v a if z c

(6-11)

where 0a is the radius of function, v is the central point of the function and c

is the number of clusters. For all three types of membership functions, values of these

parameters ,a c should be chosen such that at least one of pLz , pMz and pHz is

greater than 0.5 [91].

6.1.3 Training of RFRBFN Clustering Approach

Training algorithm consists of the following three steps:

1. Find target outputs pT z as described in Section 6.1.1

2. Compute input membership functions pz as illustrated in Section 6.1.1

3. Train the proposed RFRBFN using data obtained from step 1 and step 2

The detailed description of the last step of the training algorithm is provided as

follows. When pz is presented as an input to the RFRBF network, hidden layer

partitioned it into N classes. The output layer of the proposed scheme will perform a

linear combination of the hidden layer responses and will cluster the data into k

classes by employing Equation 2-10. In order to express Equation 2-10 in terms of

hidden layer responses in the combination of Equation 6-8, the output node response

can be written as:

1

1

1

1

1

p k

p k

qq

pk lm

l N m M

c ck

k N qCq

pr lm

r l N m M

d u

O W

d u

(6-12)

where ckW corresponds to the weight between output layer node c and hidden layer

node k . Let p

k h be the thk hidden layer node response, which can be written as:

1

1

p k

qp q

k h pk lm

l N m M

d u

(6-13)

Using this notation Equation 6-12 is simplified to.



96

1

p

k hc ck C

pk Nr h

r

O W

(6-14)

1 p

c ck k hpk Nr H

O W

(6-15)

where

1

Cp p

r H r h

r

(6-16)

If we consider p

r H as normalization term, Equation 6-13 and Equation 6-15

imply that the output node creates a linear combination of the hidden node responses.

Therefore, RFRBF network architecture is same as FRBF network proposed in [97]

and equation for updating the weights between hidden layer and output layers are as

follows:

p p

ck p c k hp

r H

W T Z O

(6-17)

Here, pT Z is the target output obtained using Equation 6-9. Thus, step 3 of

training algorithm is summarized in the following steps.

Each hidden node response p

k h is calculated by using Equation 6-13

Calculate the normalization term p

r H by using Equation 6-16

Compute the output layer response p

cO by using Equation 6-15

The hidden and output layers weights are updated by using Equation 6-17

Update the weights of input and hidden layers by using Equations 2-2 and 6-8

Repeat these steps until MSE is minimized or becomes below the specified threshold


The proposed RFRBFN method has been applied on different imaging modalities

such as the synthetic image, brain MRI and CCA (common carotid artery) ultrasound

images with different types (Gaussian, Rician and impulse) and intensities of noises.

The proposed technique has also been compared with state of the art clustering

approaches i.e. FCM, RBF and fuzzy RBF.



97

6.2.1 Performance Comparison on Synthetic Image

In order to examine the robustness of proposed RFRBFN method, we have generated

a gray scale synthetic image having three clusters. The size of image is 200x200 with

gray levels 0.00, 0.7 and 1.00. Gaussian noise of varied intensity (range from 0.001 to

0.3) was added to original image and the performance of the proposed RFRBFN was

analyzed. For performance evaluation, error rate rE has been computed and described

as the percentage of misclassified pixels. It is computed by using the following

expression:

100mcp

r

NE

R C

(6-18)

where mcpN is the total number of misclassified pixels, R and C correspond to number

of rows and the number of columns of the image, respectively. Figure 6-2(b) shows

the synthetic image corrupted with Gaussian noise of intensity 0.008. Different

algorithms have been applied to segment the noisy image. Figure 6-2(c) shows the

FCM segmented image, where a lot of misclassified patterns can be seen. It is due to

the fact that FCM is sensitive to noise RBF segmented image also contains

misclassifications at a specific cluster as shown in Figure 6-2(d). Figure 6-2(e) shows

the image segmented by fuzzy RBF network. Like FCM and RBF network segmented

images, fuzzy RBF network segmented image also contains mislabeled patterns.

The proposed RFRBFN technique segmented image is shown in Figure 6-2(f).

We can see that there are fewer mislabeled patterns as compared with FCM, RBF, and

FRBF network segmentation. From visual inspection, the proposed RFRBFN

approach offers better results compared with other techniques mentioned earlier. The

percentage of mislabeled pixels on various Gaussian noise levels is presented in

Figure 6-3.

The error rate of mislabeled pixels is shown in Figure 6-3. From the error rate

curves, we can see that the proposed RFRBF network has lower error rates at all noise

levels compared with other techniques. As shown in Figure 6-3, the proposed

technique outperforms other state of the art techniques showing robustness at all

mentioned noise levels. This is because of the smoothing parameter ( ) and hidden

layer outputs, which is not only depending on the current pixel but also utilizes the



98

spatial information. For robust clustering, the parameter should be selected very

carefully in order to minimize the effect of noise. Equation 6-13 may be rewritten as

Equation 6-15 for a very small value of parameter , which will act like fuzzy RBF.

Large values of will make the classification dependent only on pixel’s

neighborhood rather than the pixel itself. Therefore, for an accurate classification,

selection of appropriate parameter is critical.

(a) (b) (c)

(d) (e) (f)

Figure 6-2 (a) original synthetic image b) Noisy image (Gaussian noise intensity of

0.008) c) Image segmented by FCM d) RBF segmented image e) Fuzzy RBF

segmentation image and f) The proposed RFRBFN segmented image.

Cross validation technique reported in [34] was utilized in the proposed

technique to find optimum value of . In the cross validation technique, dataset was

divided into two sets and RFRBFN was trained. The parameter was initialized

randomly at one part of the dataset and the cluster centers kv as well as membership

functions pku were calculated. Using the computed kv and pku , error rate was

calculated using Expression 2-4 and same procedure was repeated for a number of

steps to obtain value. Minimum value has been selected as shown in Figure 6-4.



99

Figure 6-3 Misclassification error rE at various Gaussian noise levels

Figure 6-4 Cross validation error (in term of mean square error) of FCMJ vs

6.2.2 The proposed RFRBFN Segmentation Performance on Brain MRI

During image acquisition process, quality of brain MR images is usually

degraded. Segmentation of such degraded images is a challenging task. The proposed

RFRBFN segmentation technique is applied at T1-weighted brain MR images. The

brain MR images have been acquired from Brain web simulated brain database which

is publically available online (http://www.bic.mni.mcgill.ca/brainweb/). These images

were generated by a brain MRI simulator reported in [98].

http://www.bic.mni.mcgill.ca/brainweb/



100

The RFRBFN was trained using the steps discussed in Section 6.1.1. Once the

network has been trained for a specific imaging modality, it can be successfully

applied to segment the images of that modality. The proposed RFRBFN has been

trained in such a way that the root mean squared error (RMSE) between the target

outputs pT Z and the actual calculated outputsp

cO is below some specific threshold.

The root means squared error RMSEE can be computed by the following equations:

2

1

Kp

p c

p c

RMSE

T Z O

E

(6-19)

where K is the total number of clusters at the output layer and is the total number

of training patterns. The training parameters used for brain MR image segmentation

are summarized in Table 6.1.The obtained RMSE values for brain MR image are

illustrated in Figure 6-5.

In order to evaluate the performance of the proposed RFRBFN technique at

brain MR images, obtained segmentation results are compared with FCM, RBF

network and fuzzy RBF network techniques. Visual analysis shows that the proposed

technique offers better segmentation results. For robustness, two scenarios of noise

have been considered.

Table 6.1 Parameters used for training of network for brain MR images

Network architecture detail

2.83

0.01 for 100 iterations,

0.001 for remaining iterations

No. of input neurons 4

No. of hidden neurons 8

No. of output neurons 4

No. of iterations 1000

RMSE 0.0632

No. of training patterns 1200

300 for each cluster



101

Figure 6-5 RMSE of the proposed RFRBF network training for MR T1-weighted

image

(a) (b) (c)

(d) (e) (f)

Figure 6-6 (a) Original brain MR image (b) image with 0.01 Gaussian noise (c) FCM

segmentation (d) Image segmented by RBF technique (e) Fuzzy RBF segmented

image(f) The proposed RFRBFN segmented image.



102

(a) (d) (g) (j)

(b) (e) (h) (k)

(c) (f) (i) (l)

Figure 6-7 (a-c) Original brain MR images, (d-f) respective ground truths, (g-i)

RFRBFN based segmentation and (j-l) difference between ground truth and the image

by the proposed RFRBFN approach.

In first scenario, Gaussian noise of variance 0.01 was added to the original

brain MR image. Noisy brain MR images segmented by proposed, FCM, RBF

network and fuzzy RBF techniques as well as the proposed one, are shown in Figure

6-6. The proposed RFRBFN offers superior segmentation results even in the presence

of noise, as compared with the other techniques and it is shown in Figure 6-7 (f).

Further, Figure 6-7 shows a few sample brain MR images, their ground truths,

segmented images by the proposed technique and the difference between the ground

truth and proposed RFRBFN segmented images. Visual analysis of the segmented and

ground truth image reveals fewer misclassifications using RFRBFN. This is evident

from the difference images shown in last column of Figure 6-7 where only a few



103

misclassified pixels have been perceived. The accuracy and Rand Index are computed

from the segmented images [99]. We have obtained an accuracy of Figure 6-7(g-i)

99.68%, 99.39%, and 99.36%, respectively. While, the Rand Index for these images is

0.9924, 0.9918 and 0.9782, respectively. High value of the segmentation accuracy

and Rand Index show the usefulness of the proposed approach.

In second scenario, brain MR images are degraded by using Rician noise of

different intensities. The performance of the proposed RFRBFN segmentation

approach has been evaluated on the images containing the Rician noise. The images

are corrupted using Rician noise of intensity 4%-10% [100]. We have observed that

the proposed approach successfully segmented the brain MR images at given intensity

of Rician noise. Figure 6-8 column (a) shows the original brain MR images and

column (b) shows the brain MR images corrupted by the Rician noise of various

intensities. Figure 6-8 column(c) shows the ground truth of the column (a) images.

The proposed approach segmented images are shown in Figure 6-8 column (d). From

Figure 6-8 column (d), we can observe that the proposed RFRBFN approach

successfully segment the noisy brain MR images and showed the robustness of the

proposed approach.

Rician noise 4%

Rician noise 4%



104

Rician noise 5%

Rician noise 6%

Rician noise 8%

Rician noise 10%

Column (a) Column (b) Column (c) Column (d)

Figure 6-8, Column (a) Original brain MR images, column (b) the noisy version of

the original brain MR images, column (c) the proposed RFRBFN approach segmented

images, column (d) the ground truth of the respective brain MR images.



105

(a) (b)

(c) (d)

(e) (f)

Figure 6-9 (a) Original carotid artery ultrasound image selected ROI (b) Noisy

carotid artery image of intensity 0.05 (salt & pepper) (c) FCM segmented image (d)

image segmented by RBF (e) FRBF network segmentation and (f) the proposed

RFRBFN segmented image.

6.2.3 Segmentation of Carotid Artery Ultrasound Images

The proposed RFRBFN approach has been applied to segment on a significant real

carotid artery ultrasound images. The detail of the dataset used for segmentation is

described in Chapter 2.

Taking the advantage of the proposed segmentation approach, it has also been

applied to the common carotid artery ultrasound images. Table 6.2 shows the

architecture used for the training of RFRBFN on carotid artery ultrasound images.

Ultrasound images are affected by different types of degradations. A potential

solution is to apply some noise reduction filter to images as a preprocessing step

before segmentation. On the contrary, it will result in loss of fine details or important

edge information of the image. It is due to the fact that smoothing corresponds to a



106

low pass filtering and the fine detail lies in high frequencies. These fine details will be

lost by applying some denoising filter on image before segmentation. Therefore, a

lossless algorithm is required to suppress the noisy signal and also preserve image

details. The proposed method can be successfully used in this situation because it

smoothes membership functions within a cluster rather than between clusters, thus

suppresses noises within the cluster and preserve details taking advantage of

segmentation.

Carotid artery ultrasound images were segmented by the proposed and other

techniques. The proposed technique successfully segmented ultrasound images even

in the presence of noise. Figure 6-9(a) shows ROI of one of the original carotid artery

ultrasound images. As marked, the lower arterial wall needs to be segmented for IMT

measurements. Impulse noise of intensity 0.05 has been added to the image and the

segmentation techniques are applied. Figure 6-9(c) shows the FCM segmented image

and due to the sensitivity of FCM to noise, it contains a significant amount of

misclassified patterns. RBF segmented image is shown in Figure 6-9(d), which

contain apparently misclassified patterns. Figure 6-9(e) shows an image segmented by

the FRBF network technique. Like FCM and RBF segmented images, FRBF network

segmented images also contain many misclassified pixels. On the other hand, Figure

6-9(f) shows the image segmented by the proposed RFRBFN approach. The image is

segmented with minimum misclassified patterns. From segmented results, one can

observe the superiority of the proposed approach. The sample carotid artery

ultrasound images with marked plaque, segmented by the proposed RFRBF network

approach, are presented in Figure 6-10. The proposed approach segmented the marked

plaque with high accuracy. Figure 6-10 (a, c) show the sample carotid artery

ultrasound images with marked plaque, whereas the Figure 6-10(b, d) show the

segmented images, where the plaque has been successfully separated from

background.

Similarly, Figure 6-10 (e, g) show the carotid artery ultrasound images and

Figure 6-10 (f, h) represent the segmented images by the proposed RFRBF network

approach. The post-processing techniques are highly affected by the segmentation

quality. The proposed RFRBFN technique segments the ultrasound images with a

high precision, which has a direct impact on the classification.



107

Figure 6-10 Left column: The original carotid artery ultrasound images with marked

plaque and right column: the proposed RFRBFN segmented images.

6.2.4 Decision System for Carotid Artery Ultrasound Images

It is highly required to identify the presence of plaque in the carotid artery. For this

purpose, IMT values have been measured from the segmented images using the

(a) (b)

(c) (d)

(e) (f)

(g) (h)

IMT



108

proposed RFRBF network approach. IMT is the one of the effective indicators to

measure the thickness of arterial wall. A feature vector including mean, standard

deviation, skewness, min, and max is created so that it comprises measured IMT,

given as input to the MLBPNN classifier. The sample IMT measurements of a normal

(Figure 6-10(b)) and an abnormal (Figure 6-10(d)) carotid artery are presented in

Figure 6-11. The network structure of the MLBPNN is as follows:

Input neurons: 05

Hidden layer neurons: 50

Output layer neurons: 02

Activation function: Binary sigmoidal function.

Training epochs: 500

Employing the given structure, MLBPNN offers better classification results on the

given dataset.

Table 6.2 Parameters used for training of the network for carotid artery ultrasound

images

Network architecture detail

3.02 0.01 for 100 iterations,

0.001 for remaining iterations

No. of input neurons 3

No. of hidden neurons 6

No. of output neurons 3

No. of iterations 1000

RMSE 0.07532

No. of training patterns 900

300 for each cluster

Figure 6-11 IMT measurements of normal and abnormal carotid arteries

Different classification quality indices i.e. accuracy, sensitivity, specificity, F-score

and Mathew correlation coefficient (MCC) were computed to evaluate the



109

classification performance based on proposed segmentation technique. Furthermore,

the receiver operating characteristic curve (ROC) and area under the curve (AUC)

have also been computed. Details of the above mentioned classification performance

measures can be found in Section 2.6. Table 6.3 shows the MLBPNN decision

performance at different quality measures. For the classification of segmented carotid

artery ultrasound images, MLBPNN offers 98.20% decision accuracy. Such a

significant performance shows the effectiveness of the proposed approach. The

classification results were also compared with the recently published techniques.

Santhiyakumari et al. [32] have used MLBPNN classification for carotid artery

ultrasound images and reported 96% classification accuracy using their dataset. ROC

curves plotted for MLBPNN classification on different hidden layer neurons is shown

in Figure 6-12. ROC curve of MLBPNN at 50 hidden layer neurons is closer to the

vertical axis, which is evident of high classification accuracy. From the ROC curve

the area under the curve is computed as 0.98 which shows better diagnostic test.

Classification results show that the proposed method can successfully be applied for

the identification of a plaque in the carotid artery.

Table 6.3 The MLBPNN classification performance measures on 200 carotid artery

ultrasound images segmented by the proposed RFRBFN technique

Classification Quality Measures Performance

Accuracy (%) 98.20%

F-Score 0.9551

MCC 0.9127

Sensitivity 0.9750

Specificity 0.9385



110

Figure 6-12 The ROC curves for different MLBPNN classifier (with different

number of hidden layer neuron)

Table 6.4 Effects of hidden layer neurons on classification accuracy of carotid artery

ultrasound images

Hidden Layer Neurons Classification Accuracy (%)

30 95.20

40 96.00

50 98.20

55 97.37

60 96.74

The effect of varying hidden layer neurons on classification accuracy of

carotid artery ultrasound images is presented in Table 6.4. The hidden layer neurons

are selected empirically. It is evident from Table 6.4 that MLBPNN using the above

mentioned structure offers a superior performance. Keeping in view of segmentation

and classification performance of the proposed RFRBFN approach, it can be used as

secondary observer to identify the presence of plaque in carotid artery.

Summary

In this chapter, a new robust segmentation technique which integrates fuzzy RBF

network and robust FCM segmentation techniques has been proposed. In this

approach, a new smoothing parameter has been introduced, which minimizes the



111

effect of noise. The proposed RFRBFN approach is applied to different imaging

modalities and different noise types, and its results are compared with those of FCM,

RBF and fuzzy RBF network approaches. For robustness assessment of the proposed

approach, images were degraded by various intensities of Gaussian, Rician, and

impulse noises. The proposed technique outperformed other techniques in

segmentation of noisy or noise-free images. The proposed segmentation technique has

also applied successfully at various imaging modalities. RFRBFN approach was

applied to segment a significant amount of real carotid artery ultrasound images. A

decision system based on the proposed scheme segmentation results has been

designed by employing MLBPNN classifier. High classification accuracy 98.20% has

been achieved by using MLBPNN which shows the usefulness of the proposed

approach. The proposed approach can be applied as a secondary observer to the

experienced radiologists.

The proposed technique, however, introduces some computationally overhead

due to the incorporation of Lagrange function and spatial information. Therefore, in

order to compensate this limitation, another technique based on deformable model is

proposed, which automatically segments the carotid artery ultrasound images with

high precision. The deformable method is useful when analyzing natural scenes as

well as medical images.


1. Robust fuzzy RBF Network Based Segmentation and Intelligent Decision

Making System for Carotid Artery Ultrasound Images, Neurocomputing,

Journal (Impact Factor 2.005).

112

Chapter 7 : Automatic Active Contour Based

Segmentation and Classification of Carotid


In this chapter, a new automatic active contour based segmentation and classification

of carotid artery ultrasound images has been proposed. Early detection of the plaque

in carotid artery can avoid serious brain strokes. The active contour based techniques

have been applied successfully to segment out the carotid artery ultrasound images.

Support Vector Machine (SVM) classifier has employed to segregate the segmented

images into normal or diseased subjects. The proposed approach needs minimum

interaction for plaque detection in carotid artery.

7.1 The Proposed Approach

The proposed scheme of automated carotid artery image segmentation and

classification consists of a) preprocessing b) image alignments c) carotid artery image

segmentation d) IMT measurement e) classification of segmented carotid artery

ultrasound images into normal and abnormal subjects. Graphical representation of the

proposed approach is shown in Figure 7-1. Each step of the proposed technique is

described below in detail.

7.1.1 Alignment of Carotid Artery Ultrasound Images

During our experiments, we have found that all images are not aligned. It is due to the

fact that during the imaging process, ultrasound transducer has continuously been

moved toward around the carotid artery. Patient movement is another factor for non-

uniform imaging. Thus the movement of ultrasound transducer and the patient will

bring out images taken in out of ROI. It is easy for a human expert to locate the region

of interest (ROI) and obtain IMT measurements.

However, automatic detection of ROI by computer is not an easy task. It

requires aligned images for an accurate ROI identification. To handle rotation,

translation, and shearing transformation in carotid artery ultrasound images, we have

incorporated the image alignment as a pre-processing step in the proposed approach.

Chapter 7: Automatic Active Contour Based Segmentation and Classification of Carotid Artery

Ultrasound Images

113

Figure 7-1 Flow chart of the proposed approach

The main objective of image alignment is to make two or more images in the

same orientation. Usually, one of the undistorted images is taken as a base (reference)

image. The image which needs transformation due to any reason is considered as an

input. This input image is required to be aligned with reference image before further

computation. One of the possible approaches to align the images is the use of control

points. These control points are also called corresponding points. Location of the

control points should be known in input and reference images. In our approach, we

have applied an iterative process to select the control points. For this purpose, four

control points are selected in both the input and reference images for estimating the

transformation function. The transformation function estimation is one of the

modeling problems [28, 63]. The bilinear based approximation model is given by

1 2 3 4x c u c w c uw c (7-1)

and


Ultrasound Images

114

5 6 7 8y c u c w c uw c

(7-2)

where ,x y and ,u w are the control point coordinates of reference and input

images, respectively. In both images, if we have four corresponding pair points, eight

equations have to be defined based on Equations 7-1 and 7-2. These equations are

used to solve eight unknown coefficients 1 2 3 8, , ,...,c c c c .

Accurate alignment of images is greatly influenced by the selected control

points. Inappropriate selection of control points may lead to misleading results. For

accurate control point’s selection, the concept of spatial information has been

incorporated. In this regard, cross correlation is calculated between the selected input

and reference control points. The advantage of incorporating spatial information is an

accurate input control-point selection and results in a better aligned image. This

accurate image alignment plays a significant role in the automated carotid artery

image segmentation.

7.1.2 Snake Initialization

The major shortcoming of active contour model is snakes initialization [101]. The

initial selected window greatly affects the segmentation results. The proposed

technique automatically selects the snake initialization window which is used to

segment carotid artery ultrasound image. Snakes initialization in an automatic process

and is one of the main contributions of this research work. The input to this step is a

pre-processed image. Automatic window selection methodology is described in detail

in the following subsections.

7.1.3 Separation of Objects from Background

First of all, we have to find the objects of interest in the carotid artery ultrasound

images. For this purpose, Otsu’s method [102] has been used to separate the objects

from the background. The Otsu’s method is an automatic method of histogram based

thresholding. It is a non-parametric and unsupervised threshold selection method. The

goal is to separate out the objects from background by considering it as a two class

problem. The algorithm calculates optimal threshold to separate the objects from

background so that the intra-class variance becomes minimal by using the following

equations.


Ultrasound Images

115

2 2 2

1 1 2 2t t t t t (7-3)

where, 2

i is the class variance, i denotes the probabilities of the classes separated by

threshold t . Equation 7-4 is used to minimize the inter-class variance.

22 2 2

1 2 1 2b t t t t t t (7-4)

where 1 2 and are the means of class 1 and class 2, respectively. The probability of

class 1 t is computed from the histogram at t using Equation 7-5:

1

0

t

t p i (7-5)

where p i is the probability of a pixel belonging to respective class, whereas mean

of the class is computed by using Equation 8-6:

1

0

.t

t p i x i (7-6)

where p i is probability of the variable x i in the respective class.

The objects and background are separated from the carotid artery ultrasound

images. Some noisy and isolated patterns may remain there. These isolated patterns

have to be removed by morphological opening operation using Equation 7-7. The

morphological opening operation is usually used for smoothing object contours and

for elimination of thin protrusions [28, 63]. The opening of image f by the

structuring element b is denoted by f b , and defined as:

f b f b b (7-7)

where the symbols and denote the erosion and dilation operations, respectively.

After the morphological opening operation, area inside the artery walls has to be

intelligently identified. For this purpose, we take negative of the segmented image

using the following expression.

1s r (7-8)

where s is the resultant negative binary image and r is the segmented image using

Otsu’s approach.


Ultrasound Images

116

The advantage of an image negative is that it can be used effectively in finding

the area inside the arterial walls. After finding area inside the artery, the algorithm

intelligently decides the location and size of the window for snakes’ initialization. The

height of the initial window will be the distance between the upper and lower arterial

walls. Location and size of the window vary from image to image depending upon the

objects in the carotid artery ultrasound images. Once the window is determined, active

contour method can be successfully applied to segment the carotid artery ultrasound

images [28].

7.1.4 Segmentation of Carotid Artery Ultrasound Images

In this study, we have employed an active contour method to segment the carotid

artery ultrasound images in an automatic way. Active contour model is proposed by

Kass et al. [103] and is widely used in computer vision applications. It is based on the

energy minimization function. The advantage of an active contour model for

segmentation is that the snakes are autonomous and self-adapting in search of a

minimal energy state and can easily be manipulated. The active contour model can be

used to track objects in both the spatial and temporal domains.

The governing equation of the snake is based on the internal and external

energy components. Liang et al. [104] proposed that snake is a time variant

parametric curve T

, , , ,s t x s t y s tv located on image surface ,x y ,

where x and y are coordinate functions depending on the parameter s and t. The

Equation 7-9 is used as the governing equation of active contour model [28].

22 2

2

0

L

S

E s s ds Ps s

v

v vv v

(7-9)

The parameters and are used to control the internal energy, S v shows the

contour stiffness and elasticity. The gradient of image is calculated to find the external

energy P v . Minimum energy is obtained when contour approaches at edge of

interest.

Practically, N piecewise polynomials are used to assemble the contour. Space

independent shape function like B-Spline has been used to build each polynomial


Ultrasound Images

117

[105] weighted by the node parameters. The energy function i.e. Equation 7-9 is

minimized corresponding to Euler Lagrange partial differential equations (PDE).

2

2t

d t d tt

dt dt

u uM C Ku q (7-10)

where u(t) is the vector of N terminals, M, C and K are mass, damping and stiffness

matrices respectively. The external forces are represented by q [104].

The motion Equation 7-10 are solved with the replacement of time gradients

and its discrete approximations. By discretizing time variable t, Equation 7-10 results

in the following difference matrix equation.

1 1 2 2 1 Fu A u A u q (7-11)

where discrete time is represented by and,

2

1 2 / / ,t t A M C (7-.12)

2

2 / ,t A M (7-13)

1 2 F A A K (7-14)

The computational cost of Equation 7-11 is very high because of F inverse

calculation. An alternative formulation has been proposed by Weruaga et al. [106]

for translation of the energy function i.e. expression 7.9 in frequency domain. In this

procedure, matrix inversion is avoided because in frequency domain it becomes a

point-wise inversion.

To get the final solution smoother, B-splines are utilized as a shape function

[104], and at each update of the contour some additional calculations are performed.

To get better execution time ratio, cubic B-splines have been used in this study [107].

Cubic B-splines interpolation technique has been applied at the node parameters u to

achieve the active contour v.


The carotid artery disease diagnosis heavily depends upon an accurate segmentation

and classification of the segmented images. Segmented images need to be classified

into normal or abnormal subjects. In this study, we have employed SVM classifier for

classification of the carotid artery images. Four different features are extracted from


Ultrasound Images

118

measured IMT values obtained from the carotid artery segmented images for training

and testing of SVM classifier.

7.2.1 IMT Feature Extraction

The following four commonly used statistical features are extracted from IMT

measurements of the segmented carotid artery image. These features are normalized

in the range of [0 1] and then used as input to the SVM classifier. Classification

performance has been obtained to be the best using these features for the given

dataset. Mathematical description of each feature is as under:

1

1Average

n

i

i

xN

(7-15)

where, 1 2, ,...., nx x x are the observed values and n is the total number of

observations.

21

Var(x) iE xN

(7-16)

where, 1 2, ,....,

Nx x x are the observed values, µ is the mean of the observed values

and N is the total number of observations.

21

s.d iE xN

(7-17)

Standard deviation is the square root of the variance defined in Equation 8-16. 3

Skewness ixE

(7-18)

where µ shows the mean and expected value of quantity is represented by E.

7.2.2 Classification Performance Measurements

For classification of segmented carotid artery ultrasound images, we have used SVM

classifier and the detail of SVM classifier can be found in Section Error! Reference

ource not found.. For classification evaluation the following performance measures

accuracy, sensitivity, specificity, MCC, F-score, ROC, AUC and Negative predictive

value (NPV) have been computed. These measures are described and defined in

Section 2.6, respectively.


Ultrasound Images

119


The real carotid artery ultrasound images dataset, described in Chapter 2, are

segmented and then classified using SVM classifier. Figure 7-2(a) shows one of the

original carotid artery ultrasound images. The region of interest (ROI) is selected

from the original image. Figure 7-2(b) shows the selected ROI. To reduce the speckle

noise and wave interferences, the median filter is applied on the original image. As

described in Section 7.1, a window for snake initialization has been determined

automatically using the mentioned procedure. The selected window for snake

initialization is shown in Figure 7-2(c). Carotid artery ultrasound image is segmented

using active contour approach as shown in Figure 7-2(d). From Figure 7-2(d), it can

be observed that the arterial walls are segmented accurately by the active contour

method. Accurate segmentation of carotid artery ultrasound images using active

contour approach highly depends upon the window initialization. If initialization is

done well, one can get significant results. In the proposed approach, we have

developed a technique for automatic initialization of snakes for carotid artery

ultrasound image segmentation. As described earlier, manual initialization of snakes

by inexperience users may lead toward false segmentation results. Hence, there is

need of such a mechanism in which minimum interaction from user is required. The

proposed approach requires minimum user interaction for segmentation of carotid

artery ultrasound images

A comparison of manual and an automatic snake initialization have been made

for segmentation of carotid artery ultrasound images. The proposed approach shows

the promising segmentation results for carotid artery ultrasound images. Figure 7-3

column (a) shows the carotid artery images of different patients segmented by our

proposed approach using automatic initialization of snakes and Figure 7-3 column (b)

shows results through manual snakes initialization. The last row of Figure 7-3 shows

the normal carotid artery ultrasound images segmented by the proposed and manual

initialization techniques. From Figure 7-3, it can be observed that images segmented

by the proposed and manual initialization approach do not have any difference, which

shows the effectiveness of our proposed approach.

The problem in automatic initialization of snakes arises when images are not

aligned. The images might be rotated, sheared and translated during image acquisition


Ultrasound Images

120

process. Such images are required to be aligned before snake initialization and

segmentation. To cope with this problem, we have incorporated the concept of image

registration into the proposed methodology. The advantage of image registration is

that it aligns two or more images. It requires the input image to be aligned with the

base image. Accurate registration needs a precise for selection of control points. To

deal with this problem, we have incorporated spatial information for selection of

control points. The spatial information helps to correlate with base and input image

control points. It increases the corresponding points matching and results in the form

of a better registered image.

Table 7.1, shows the IMT measurement (mm) of maximum and minimum of

one of the normal carotid artery ultrasound images. The graphical representation of

upper and lower arterial walls IMT measurement (mm) are shown in Figure 7-4. IMT

is one of the effective measures by which seriousness of a disease can be identified.

Figure 7-2 a) The original carotid artery ultrasound image, b) the cropped and median

filtered ROI, c) automatic snake initialization window, d) segmented carotid artery

ultrasound image using active contour method.

IMT values are measured from segmented images and four features, namely,

average, variance, standard deviation and skewness are extracted. These feature


Ultrasound Images

121

values are given as input to SVM classification for training and testing. Various

classification performance measures are shown in Table 7.2. It is edident from Table

7.2 SVM offers high classification accuracy based on proposed technique

segmentation.

Table 7.1 Various IMT measurements of upper and lower wall of carotid artery in

term of mm

Figure 8.4 ( Segmented carotid artery) IMT (mm)

Upper Artery Wall (min) 0.149

Upper Artery Wall (max) 0.298

Lower Artery Wall (min) 0.149

Lower Artery Wall (max) 0.447

Figure 7-3 Column a) segmentation results using our proposed automatic snake

initialization approach and column b) images segmented by manual snake

initialization.

Column(a) Column(b)


Ultrasound Images

122

Figure 7-4 IMT measurement of a normal carotid artery ultrasound image

Table 7.2 Classification performance comparison of the various classifiers based on

proposed technique segmentation.

Techniques Accuracy

(%)

F-

Score

MCC Sensitivity Specificity NPV

(%)

KNN 98.40 0.9799 0.9600 0.9839 0.9762 98.4

PNN 98.40 0.9799 0.9600 0.9839 0.9762 98.4

SVM 98.80 0.9879 0.9763 1.0000 0.9766 100

Furthermore, we compare the performance classification techniques in terms

of ROC and AUC. Figure 7-5 shows ROC curves obtained for the true positive (TP)

and false positive (FP) rates using SVM, KNN and PNN classifiers, respectively.

From Figure 7-5, it can be observed that SVM ROC is close to vertical axis which

shows a high classification accuracy. Small deviation in ROC can be observed in both

KNN and PNN as compared to SVM. Table 7.2 shows that KNN and PNN offers

equal classification performance; hence ROC of PNN and KNN overlaps in the graph.

We have obtained 98.80% classification accuracy for the proposed approach. ROC

curve shows some variation at y-axis and this variation represents the

misclassifications when the threshold for classification is very strict. To check the

overall performance of the classifier, we have also computed area under the curve

(AUC). SVM offers high AUC value 0.98 which shows high classification rate. SVM

value of AUC is close to 1.0 which shows the overall better classification

performance (better diagnostic test). It confirms the superiority of the proposed


Ultrasound Images

123

approach. High NPV value 100% is obtained by the proposed approach based

classification. Resulted high value of NPV means that the probability of a patient has

become risk free for which the test was conducted. Therefore, it can be concluded that

our proposed approach can be used effectively for diagnostics of plaque into carotid

artery ultrasound images.

In addition, performance comparison of the proposed scheme in terms of a

size of the feature vector for classification has been carried out with other techniques.

Minghao et al. [95] have used SVM classification with 22 different features obtained

from IMT measurements. They have reported F-measure 0.98 using their own dataset.

While a large feature vector dimension may affect the classification accuracy because

of the redundant features. In the proposed scheme, high F-measure ~ 0.99 value has

been achieved with only four features as compared to 22 features dataset. The reduced

features may save classification time and resources. Moreover, the proposed scheme

outperforms using given features.

In term of overall accuracy, Santhiyakumari et al. [32] have employed

Multilayer Backpropagation Neural Networks (MLBPNN) for classification and have

reported a maximum of 96% classification accuracy. In their approach, IMT values

are used to train the neural networks, but they did not mention feature extraction

strategy and it might be difficult to reproduce the results. On the other hand, in the

proposed scheme, a straight forward approach has been utilized to classify the

segmented images and have obtained 98.8% classification accuracy. Similarly,

Hassan et al. [59] have obtained 98.4% classification accuracy. Statistical results

indicate the usefulness of the proposed approach.

Furthermore, to check the performance of our proposed approach, we have

applied KNN (k-nearest neighborhood) and PNN (probabilistic neural networks) on

the given dataset. The detail of KNN and PNN can be found in [108] and [109],

respectively. Table 7.2 shows comparison of different methods at different quality

measures. From Table 7.2, it is clear that SVM offers superior results as compared to

the other techniques. Statistical analysis of results represents the usefulness of the

proposed approach. Figure 7-6 shows the graphical performance comparison of PNN,

KNN and SVM at various classification validity measures. Classification using SVM


Ultrasound Images

124

offers superior performance as compared with the other techniques at all validity

measures.

Figure 7-5 ROC curves of SVM, KNN and PNN classifiers showing true positive

(TP) and false positive (FP) rates at different thresholds

Figure 7-6 Performance comparison of KNN, PNN and SVM classifiers at various

classification performance measures

The proposed automatic active contour and manual initialization of snakes

offers similar performance. But the advantage of our proposed approach for carotid

artery ultrasound image segmentation is that it requires minimal user intervention.

Image classification is one of the most important steps to know about the presence of

a plaque in carotid artery. In the proposed scheme, SVM classifier has been used for


Ultrasound Images

125

the classification of normal and abnormal subjects. Statistical analysis shows that the

proposed approach outperforms other techniques at hand.

The countries like Pakistan where there is a lack of health facilities to the

people, the proposed approach might be useful to check the possible presence of a

plaque in the carotid artery at early stage. Hence, the proposed approach can be

successfully used as an initial diagnosing tool. It can also be used by minimal

experienced users and early detection of plaque in artery may prevent serious brain

strokes.

Summary

In this chapter, a new approach using automatic initialization of snakes for carotid

artery image segmentation has been proposed. For manual initialization it is required

to place snake initialization window at appropriate location. However, inexperienced

user initialization may lead toward misleading results. The proposed automatic active

contour segmentation approach performance is compared with the manual

initialization of snakes. It has been observed that the proposed approach successfully

initialized the snakes to segment the carotid artery ultrasound images. Furthermore,

we have employed SVM classifier and have achieved 98.80% classification accuracy.

Classification results of the proposed technique segmentation are compared with PNN

and KNN classifiers as well. It has been observed that SVM classifier outperforms

using the proposed technique segmented images.


Journal Publications:

1. Automatic Active Contour-Based Segmentation and Classification of Carotid

Artery Ultrasound Images, Published, Journal of Digital Imaging, IF=1.26.

2. Automatic Carotid Artery Image Segmentation using Snake Based Model,

Published, Journal of Korean Navigation Institute, Volume 17(1), pp. 115-

122, 2013”. Impact Factor 0.278.

Book Chapter:

1. Automatic Segmentation and Decision Making of Carotid Artery Ultrasound

Images. Advances in Intelligent Systems and Computing, 2012 (ISI

Indexed).

126

Chapter 8 : Conclusions and Future Directions

Carotid arteries supply oxygenated blood to the brain. The carotid artery stenosis is

the process of narrowing the carotid artery due to the presence of atherosclerosis. The

atherosclerosis is caused by the cholesterol and other fatty materials, which introduce

plaque in carotid artery. The plaque may partially or fully block the blood flow to the

brain and the probability of cerebrovascular accident (stroke) becomes high. Many

techniques (invasive and non-invasive) are currently in practice to identify the plaque

in carotid artery. Among those, we have selected carotid artery ultrasound imaging

modality for plaque detection, which is a safe and non-invasive diagnosis approach.

The sole objective of this research was to identify the presence of plaque in

carotid artery by utilizing the ultrasound imaging modality. Several techniques have

been proposed to achieve the objective. First and the most important phase of a

computer aided diagnostic system is image segmentation. Accurate segmentation of

the ultrasound imaging is challenging, because of its low quality, presence of noise

and wave interferences. We have proposed five techniques which have successfully

segmented the carotid artery ultrasound images. The major features and results of

each technique are briefly described below.

8.1 Modified Spatial Fuzzy C-means and Ensemble Clustering

A new clustering algorithm named sFCMM has been proposed for segmentation of

carotid artery ultrasound images. The proposed sFCMM overcomes the equal

weightage problem of spatial sFCM algorithm. In addition to this, ensemble clustering

of majority voting has also investigated. The proposed scheme utilized hybrid feature

including GLCM, 2D-CWT and MGH. Most important features are selected by GA.

The main objective of features optimization is to reduce feature vector dimensionality

which results in low computational cost and higher accuracy. Visual and quantitative

results of the proposed clustering algorithm have been compared with the state of the

art clustering techniques such as FCM, sFCM, K-means, SOM, and sFCMLSM. High

quality clustering has been achieved using ensemble clustering of majority voting

scheme. It can be observed that the ROI is efficiently separated from background

Chapter 8: Conclusions and Future Directions

127

tissues by the proposed segmentation technique. IMT values are measured from

segmented carotid artery ultrasound images and given as input to MLBPNN for

classification. MLBPNN offers an enhanced classification accuracy of 98.40% based

on the proposed segmentation technique. However, the proposed sFCMM approach

utilizes median filter as a pre-processing step and it may smoothes the important

image details.

8.2 Robust Information Gain Based FCM Clustering

A new robust segmentation technique called IGFCM approach has been proposed.

The proposed algorithm incorporates the concept of information gain into

conventional fuzzy framework in order to overcome the shortcomings of conventional

FCM algorithm. The proposed IGFCM algorithm has been applied to segment

different modality images like synthetic, CT liver, and daylight, results are compared

with the conventional FCM, sFCM, sFCMM, and FLICM techniques. Gaussian noise

of various intensities has been added to the images in order to check the robustness of

the proposed IGFCM algorithm. Visually and quantitatively the superiority of the

proposed IGFCM segmentation algorithm has been verified. PNN classifier is

employed to separate the normal and abnormal subjects. More than 98% classification

accuracy has been achieved by PNN using the proposed scheme segmented images.

Furthermore, effects of segmentation at classification have also been investigated. It

has been observed that the classification accuracy has been increased by 4.31% using

the proposed IGFCM segmentation. However, the proposed IGFCM did not utilizes

the spatial information of a pixel thus at very high noise level it may produces

misleading results.

8.3 Robust Segmentation of Carotid Artery Ultrasound Images using

Neuro Fuzzy and Expectation Maximization

A new robust segmentation approach namely RSC-US has been proposed. The

proposed approach employed neuro fuzzy classifier, EM, and GA for segmentation of

the carotid artery ultrasound images. By utilizing the GA based selected features;

NFC has been employed to generate FIS, hence used for segmentation. The results of

the proposed technique have been compared visually and quantitatively with FCM, K-

means, and sFCMLSM approaches on noisy and noise free images. The proposed

RSC-US shows outstanding segmentation results compared to other state of the art


128

techniques. SVM classifier is employed to distinguish normal and abnormal subjects.

High classification accuracy of 98.84% is achieved by SVM classification using the

proposed RSC-US segmentation approach. The RSC-US is semi supervised approach

and did not utilizes the spatial information of pixel which plays a vital role in

segmentation.

8.4 Robust Fuzzy RBF Network Based Segmentation

A new robust segmentation technique, which integrates fuzzy RBF network and the

robust FCM segmentation techniques have been proposed. In this approach, a new

smoothing which reduces the effect of noise without applying any noise removal

technique has been introduced. The proposed RFRBFN approach is applied to the

different imaging modalities and different noise types, and results are compared with

FCM, RBF and fuzzy RBFN approaches. In order to check the robustness, the images

were degraded by Gaussian, Rician as well as the impulse noises. The proposed

technique outperforms the other state of the art techniques for segmentation of noisy

images as well. Medical decision making system based on MLBPNN has been

designed for plaque identification in carotid artery; hence classification accuracy more

than 98% has been achieved. However, the RFRBFN approach is some

computationally overhead due to the incorporation of Lagrange function and spatial

information.

8.5 Automatic Active Contour Based Segmentation of Carotid Artery

Ultrasound Images

A new automatic active contour based approach for carotid artery ultrasound image

segmentation has been proposed. Using manual initialization, a window for snake

initialization has been set empirically and required to be placed at right location.

However, improper snake initialization may produce misleading results. Performance

comparison of the proposed approach has been carried out with the manual

initialization of snakes. The proposed approach of an automatic active contour based

segmentation shows the effectiveness of the proposed scheme. Furthermore, the

segmented images are classified into normal and abnormal subjects. High

classification accuracy has been achieved by SVM classifier using the proposed

segmentation scheme. However to obtained accurate segmentation of carotid artery


129

ultrasound images, selection of control points is very sensitive and it expects an

expertise from the user to alignment images.

8.6 IMT Measurements and Medical Decision Systems

IMT values have been measured from carotid artery ultrasound images segmented by

our proposed segmentation techniques. Based on IMT values, it can decide whether

the carotid artery is normal or abnormal. IMT measurements are highly depending on

the segmentation quality. The results of post-processing steps are dependent on

segmentation. It is mandatory that the images are segmented with high accuracy so

that medical decision system may be proposed for carotid artery plaque identification.

Classification of the segmented carotid artery is vital because based on its results

patient’s rehabilitation process may start. Different features from measured IMT

values were extracted and given as input to the decision making system for

identification of normal and abnormal subjects. In this thesis, various decision system

such as, SVM, MLBPNN, KNN, and PNN to segregate normal and abnormal

subjects. Among those, an enhanced classification accuracy 98.84% with SVM

classifier using RBF kernel function. It has been observed that classification results

based on the proposed segmentation technique are superior as compared to other state

of the art techniques. Moreover, the effects of segmentation at classification have also

been investigated. The proposed segmentation techniques improve the classification

accuracy by 4%. Keeping in view the segmentation quality and decision system

accuracy, the decision system can be used in health care centers to diagnose the

presence of a plaque in carotid artery. In Pakistan, where there is lack of medical

facility, especially in rural areas, the proposed system can be implemented

successfully for initial screening of carotid artery plaque patients. The system can also

be used as a secondary observer to the medical experts for plaque diagnosis in carotid

artery.

8.7 Future Recommendations

In order to obtain robust and accurate segmentation, the proposed segmentation

techniques can also be extended. We have worked on ultrasound imaging modality

and the techniques can be extended for other imaging modalities such as CT scan and

MRI. Quality of segmentation and consequently high classification results are

expected by employing computational intelligence techniques. A medical decision


130

system can be designed for diagnosis based on other imaging modalities. Such

systems might be useful for initial screening especially for those patients who are

living in remote areas of Pakistan.

131

References

[1] NHLBI, U.S. Department of Health & Human Services, "Explore Carotid

Artery Disease ", Internet: http://www.nhlbi.nih.gov/health/health-

topics/topics/catd/., May. 15, 2013,[Feb. 14, 2014].

[2] J. Stein "Stroke: Essentials of Physical Medicine and Rehabilitation: National

Institute of Health", Internet:

http://www.nlm.nih.gov/medlineplus/ency/article/000726.htm , Oct, 20. 2012,

[Feb. 14, 2014].

[3] S. L. Murphy, J. Xu, and K. D. Kochanek, "National Vital Statistics Report,"

National Center for Health Statistics, Hyattsville 2013.

[4] National Stroke Association, "Types of Stroke", Internet:

http://www.stroke.org/site/PageServer?pagename=type, Oct 12, 2013, [Feb.

14, 2014].

[5] Mayo Clinic, "Stroke: Symptoms", Internet:

http://www.mayoclinic.com/health/stroke/DS00150/DSECTION=symptoms.

Sep 21, 2013, [Feb 14, 2014].

[6] J. Tan, D. Lim Joon, G. Fitt, M. Wada, M. Lim Joon, A. Mercuri, M. Marr, M.

Chao, and V. Khoo, "The utility of multimodality imaging with CT and MRI

in defining rectal tumour volumes for radiotherapy treatment planning: a pilot

study," Journal of Medical Imaging and Radiation Oncology, vol. 54, pp. 562-

568, 2010.

[7] S. M. Smith, P. T. Fox, K. L. Miller, D. C. Glahn, P. M. Fox, C. E. Mackay,

N. Filippini, K. E. Watkins, R. Toro, A. R. Laird, and C. F. Beckmann,

"Correspondence of the brain's functional architecture during activation and

rest," Proceedings of the National Academy of Sciences, vol. 106, pp. 13040-

45, 2009.

[8] N. Ayache, P. Cinquin, I. Cohen, L. Cohen, F. Leitner, and O. Monga.,

"Segmentation of complex three dimensional medical objects: a challenge and

a requirement for computer-assisted surgery planning and performance,"

Computer integrated surgery: technology and clinical applications, pp. 59-74,

1996.

[9] N. Sharma and L. M. Aggarwal, "Automated medical image segmentation

techniques," Journal of Medical Physics / Association of Medical Physicists of

India, vol. 35, pp. 3-14, 2010.

http://www.nhlbi.nih.gov/health/health-topics/topics/catd/

http://www.nhlbi.nih.gov/health/health-topics/topics/catd/

http://www.nlm.nih.gov/medlineplus/ency/article/000726.htm

http://www.stroke.org/site/PageServer?pagename=type

http://www.mayoclinic.com/health/stroke/DS00150/DSECTION=symptoms

132

[10] K.-S. Chuang, H.-L. Tzeng, S. Chen, J. Wu, and T.-J. Chen, "Fuzzy c-means

clustering with spatial information for image segmentation," Computerized

Medical Imaging and Graphics, vol. 30, pp. 9-15, 2006.

[11] A. Chaudhry, M. Hassan, A. Khan, J. Y. Kim, and T. A. Tuan, "Image

clustering using improved spatial fuzzy C-means," in Proceedings of the 6th

International Conference on Ubiquitous Information Management and

Communication. Kuala Lumpur, Malaysia: ACM, 2012.

[12] Z. Iscan, A. Yüksel, Z. Dokur, M. Korürek, and T. Ölmez, "Medical image

segmentation with transform and moment based features and incremental

supervised neural network," Digital Signal Processing, vol. 19, pp. 890-901,

2009.

[13] M. Vasantha, Dr. V. Sunniah Bharathi, and T. Dhamodharan, "Medical Image

Features, Extraction, Selection and Classification," International Journal of

Engineering Sciences and Technology, vol. 2, pp. 2071-2076, 2010.

[14] R. Rocha, A. l. Campilho, J. Silva, E. Azevedo, and R. Santos, "Segmentation

of ultrasound images of the carotid using RANSAC and cubic splines,"

Computer Methods and Programs in Biomedicine, vol. 101, pp. 94-106, 2011.

[15] B. N. Li, C. K. Chui, S. Chang, and S. H. Ong, "Integrating spatial fuzzy

clustering with level set methods for automated medical image segmentation,"

Computers in Biology and Medicine, vol. 41, pp. 1-10, 2011.

[16] J. Yu and J. Tan, "Object density-based image segmentation and its

applications in biomedical image analysis," Computer Methods and Programs

in Biomedicine, vol. 96, pp. 193-204, 2009.

[17] F. Mao, J. Gill, D. Downey, and A. Fenster, "Segmentation of carotid artery in

ultrasound images," in Proceedings of the Proceedings of the 22nd Annual

International Conference of the IEEE Engineering in Medicine and Biology

Society, 2000, pp. 1734-1737

[18] P. Abolmaesumi, M. R. Sirouspour, and S. E. Salcudean, "Real-time

extraction of carotid artery contours from ultrasound images," in Proceedings

of the 13th IEEE Symposium on Computer-Based Medical Systems, CBMS. ,

2000, pp. 181-186.

[19] D.-C. Cheng, A. Schmidt-Trucksass, K.-s. Cheng, and H. Burkhardt, "Using

snakes to detect the intimal and adventitial layers of the common carotid artery

wall in sonographic images," Computer Methods and Programs in

Biomedicine, vol. 67, pp. 27-37, 2002.

133

[20] S. Hovda, H. Rue, and B. Olstad, "New Doppler-based imaging method in

echocardiography with applications in blood/tissue segmentation," Computer

Methods and Programs in Biomedicine, vol. 96, pp. 12-24, 2009.

[21] A. K. Hamou and M. R. El-Sakka, "A novel segmentation technique for

carotid ultrasound images," in Proceedings of the IEEE International

Conference on Acoustics, Speech, and Signal Processing, (ICASSP '04). 2004,

pp. 521-524.

[22] A. R. Abdel-Dayem and M. R. Ei-Sakka, "A novel morphological-based

carotid artery contour extraction," in Proceedings of the Canadian Conference

on Electrical and Computer Engineering 2004, pp. 1873-1876.

[23] L. Vincent and P. Soille, "Watersheds in digital spaces: an efficient algorithm

based on immersion simulations," IEEE Transactions on Pattern Analysis and

Machine Intelligence, vol. 13, pp. 583-598, 1991.

[24] A. R. Abdel-Dayem, M. R. El-Sakka, and A. Fenster, "Watershed

segmentation for carotid artery ultrasound images," in Proceedings of the The

3rd ACS/IEEE International Conference on Computer Systems and

Applications, 2005, pp. 131.

[25] M. Kamel, A. Campilho, A. R. Abdel-Dayem, and M. R. El-Sakka Senior

Member Ieee, "Carotid Artery Ultrasound Image Segmentation Using Fuzzy

Region Growing," in Image Analysis and Recognition, vol. 3656, Lecture

Notes in Computer Science: Springer Berlin / Heidelberg, 2005, pp. 869-878.

[26] C. P. Loizou, C. S. Pattichis, M. Pantziaris, and A. Nicolaides, "An Integrated

System for the Segmentation of Atherosclerotic Carotid Plaque," IEEE

Transactions on Information Technology in Biomedicine, vol. 11, pp. 661-

667, 2007.

[27] M. Bastida-Jumilla, R. Menchon-Lara, J. Morales-Sanchez, R. Verdu-

Monedero, J. Larrey-Ruiz, and J. L. Sancho-Gomez, "Segmentation of the

Common Carotid Artery Walls Based on a Frequency Implementation of

Active Contours," Journal of Digital Imaging, vol. 26, pp. 129-139, 2013.

[28] A. Chaudhry, M. Hassan, A. Khan, and J. Kim, "Automatic Active Contour-

Based Segmentation and Classification of Carotid Artery Ultrasound Images,"

Journal of Digital Imaging, vol. 26, pp. 1071-81, 2013.

[29] S. Golemati, J. Stoitsis, E. Sifakis, T. Balkizas, and K. S. Nikita, "Using the

Hough transform to segment ultrasound images of longitudinal and transverse

sections of the carotid artery," Ultrasound in Medicine & Biology, vol. 33, pp.

1918-1932, 2007.

134

[30] A. Abdel Dayem and M. El Sakka, "Carotid Artery Contour Extraction from

Ultrasound Images Using Multi-Resolution-Analysis and Watershed

Segmentation Scheme," International journal on Graphics Vision and Image

processing, vol. 5, pp. 1-10, 2010.

[31] L. Quan, I. Wendelhag, J. Wikstrand, and T. Gustavsson, "A multiscale

dynamic programming procedure for boundary detection in ultrasonic artery

images," Medical Imaging, IEEE Transactions on, vol. 19, pp. 127-142, 2000.

[32] N. Santhiyakumari, P. Rajendran, and M. Madheswaran, "Medical Decision-

Making System of Ultrasound Carotid Artery Intima–Media Thickness Using

Neural Networks," Journal of Digital Imaging, vol. 24, pp. 1112-1125, 2011.

[33] J. Bezdek, L. Hall, and L. Clarke, "Review of MR image segmentation using

pattern recognition," Med. Phys., vol. 20, pp. 1033-48, 1993.

[34] D. L. Pham, "Spatial Models for Fuzzy Clustering," Computer Vision and

Image Understanding, vol. 84, pp. 285-297, 2001.

[35] N. S. Iyer, A. Kandel, and M. Schneider, "Feature-based fuzzy classification

for interpretation of mammograms," Fuzzy Sets and Systems, vol. 114, pp.

271-280, 2000.

[36] M. Hassan, A. Chaudhry, A. Khan, and M. A. Iftikhar, "Robust information

gain based fuzzy c-means clustering and classification of carotid artery

ultrasound images," Computer Methods and Programs in Biomedicine, vol.

113, pp. 593-609, 2013.

[37] P. E. Andries, Computational Intelligence: An Introduction, 2nd ed: Wiley

Publishing, 2007.

[38] S. K. Pal and S. Mitra, "Multilayer perceptron, fuzzy sets, and classification,"

IEEE Transactions on Neural Networks, vol. 3, pp. 683-697, 1992.

[39] L. G. Shaprio and G. C. Stockman, Computer Vision. Upper Saddle River, NJ:

Printice Hall, 2001.

[40] Kohonen T, Self Organizing and Associative Memory. New York, USA:

Springer Verlag 1989.

[41] M. Lorenz, H. Markus, M. Bots, M. Rosvall, and S. M., "Prediction of clinical

cardiovascular events with carotid intima-media thickness: a systematic

review and metaanalysis," Circulation, vol. 115, pp. 459-67, 2007.

[42] Y.-F. Juan, T. Chen, T.-H. Chao, and C.-C. Huang, "The investigation of the

relationship between carotid intima-media thickness and vascular compliance

135

in patients with coronary artery disease," Biomedical Engineering:

Applications, Basis and Communications, vol. 16, pp. 37-42, 2004.

[43] M. A. Mansour and N. Labropoulos, Vascular Diagnosis. Philadelphia,

Pennsylvania: Elsevier, 2005.

[44] C. Sass, B. Herbeth, O. Chapet, G. Siest, S. Visvikis, and F. Zannad, "Intima-

media thickness and diameter of carotid and femoral arteries in children,

adolescents and adults from the Stanislas cohort: effect of age, sex,

anthropometry and blood pressure," J Hypertens., vol. 16, pp. 1593-602, 1998.

[45] R. Salonen and J. Salonen, "Progression of carotid atherosclerosis and its

determinants: A population-based ultrasonography study," Atherosclerosis,

vol. 18, pp. 33-40, 1990.

[46] G. Howard, A. Sharrett, G. Heiss, G. Evans, L. Chambless, Riley WA, and B.

GL, "Carotid artery intimalmedial thickness distribution in general populations

as evaluated by B-mode ultrasound," Storke, vol. 24, pp. 1297-304, 1993.

[47] J. Stevens, H. Tyroler, J. Cai, C. Paton, A. Folsom, G. Tell, P. Schreiner, and

L. Chambless, "Body weight changes and carotid artery wall thickness. The

Atherosclerosis Risk in Communities Study," Am. J. Epidemiol, vol. 147, pp.

563-73, 1998.

[48] M. Muiesan, G. Pasini, M. Salvetti, S. Calebich, R. Zulli, M. Castellano, D.

Rizzoni, G. Bettoni, A. Cinelli, E. Porteri, V. Corsetti, and E. Agabiti-Rosei,

"Cardiac and vascular structural changes. Prevalence and relation to

ambulatory blood pressure in a middle-aged general population in northern

Italy: the Vobarno Study.," Hypertesion, vol. 27, pp. 1046-52, 1996.

[49] A. Visona, L. Lusiani, A. Bonanome, G. Beltramello, L. Confortin, B.

Papesso, F. Costa, and A. Pagnan, "Wall thickening of common carotid

arteries in patients affected by noninsulin-dependent diabetes mellitus:

relationship to microvascular complications," Angiology, vol. 46, pp. 793-9,

1995.

[50] L. Wagenknecht, R. J. D'Agostino, P. Savage, D. O'Leary, M. Saad, and S.

Haffner, "Duration of diabetes and carotid wall thickness. The Insulin

Resistance Atherosclerosis Study (IRAS)." Storke, vol. 28, pp. 999-1005,

1997.

[51] C. I. Christodoulou, C. S. Pattichis, M. Pantziaris, and A. Nicolaides,

"Texture-based classification of atherosclerotic carotid plaques," IEEE

Transactions on Medical Imaging, vol. 22, pp. 902-912, 2003.

136

[52] C. I. Christodoulou, C. S. Pattichis, E. Kyriacou, and A. Nicolaides, "Image

Retrieval and Classification of Carotid Plaque Ultrasound Images," Journal of

Cardiovascular Imaging, vol. 2, pp. 18-28, 2010.

[53] F. Latifoglu, K. Polat, S. Kara, and S. Gunes, "Medical diagnosis of

atherosclerosis from Carotid Artery Doppler Signals using principal

component analysis (PCA), k-NN based weighting pre-processing and

Artificial Immune Recognition System (AIRS)," Journal of Biomedical

Informatics, vol. 41, pp. 15-23, 2008.

[54] E. C. Kyriacou, S. Petroudi, C. S. Pattichis, M. S. Pattichis, M. Griffin, S.

Kakkos, and A. Nicolaides, "Prediction of High-Risk Asymptomatic Carotid

Plaques Based on Ultrasonic Image Features," IEEE Transactions on

Information Technology in Biomedicine, vol. 16, pp. 966-973, 2012.

[55] N. Tsiaparas, S. Golemati, I. Andreadis, J. S. Stoitsis, I. Valavanis, and K. S.

Nikita, "Comparison of Multiresolution Features for Texture Classification of

Carotid Atherosclerosis From B-Mode Ultrasound," Trans. Info. Tech.

Biomed., vol. 15, pp. 130-137, 2011.

[56] S. Petrović, "A Comparison Between the Silhouette Index and the Davies-

Bouldin Index in Labelling IDS Clusters," in Proceedings of the 11th Nordic

Workshop on Secure IT-systems, NORDSEC Linkoping, Sweden, 2006, pp.

53-56.

[57] M. Wigness, "Evaluating Cluster Quality For Visual Data," in Department of

Computer Science, vol. MSCS. Fort Collins, Colorado: Colorado State

University, 2013.

[58] D. L. Davies and D. W. Bouldin, "A Cluster Separation Measure," IEEE

Transaction on Pattern Analysis and Machine Intelligence, vol. 1, pp. 224-227,

1979.

[59] M. Hassan, A. Chaudhry, A. Khan, and J. Y. Kim, "Carotid artery image

segmentation using modified spatial fuzzy c-means and ensemble clustering,"

Computer Methods and Programs in Biomedicine, vol. 108, pp. 1261-76,

2012.

[60] J. C. Bezdek "Cluster Validity with Fuzzy Sets," Journal of Cybernetics, vol.

3, pp. 58 - 73, 1973.

[61] Bazdek JC., "Mathematical models for systematic and taxonomy," in

Proceedings of the International Conference on Numerical Taxonomy, San

Frabcisco, 1975, pp. 143-66.

[62] D. G. Altman, "Statistics Notes: Diagnostic tests 2: predictive values," BMJ,

vol. 309, pp. 102, 1994.

137

[63] G. Gonzalez and E. Woods, Digital Image Processing, Third Edition ed:

Prentice Hall, 2008.

[64] T. Huang, G. Yang, and G. Tang, "A fast two-dimensional median filtering

algorithm," IEEE Transactions on Acoustics, Speech and Signal Processing,

vol. 27, pp. 13-18, 1979.

[65] T. Loupas, W. N. McDicken, and P. L. Allan, "An adaptive weighted median

filter for speckle suppression in medical ultrasonic images," IEEE

Transactions on Circuits and Systems, vol. 36, pp. 129-135, 1989.

[66] A. Toprak and İ. Güler, "Suppression of Impulse Noise in Medical Images

with the Use of Fuzzy Adaptive Median Filter," Journal of Medical Systems,

vol. 30, pp. 465-471, 2006.

[67] C. Tomasi and R. Manduchi, "Bilateral filtering for gray and color images," in

Proceedings of the Sixth International Conference on Computer Vision,

Bombay, India, 1998, pp. 839-846.

[68] H. S. Sheshadri and A. Kandaswamy, "Experimental investigation on breast

tissue classification based on statistical feature extraction of mammograms,"

Computerized Medical Imaging and Graphics, vol. 31, pp. 46-48, 2007.

[69] S. Ao, B. B. Rieger, and M. Amouzegar, Machine Learning and Systems

Engineering, vol. 68. California, USA: Springer, 2010.

[70] R. Polikar, "Fundamental concepts & an overview of the wavelet theory:

Tutorial, http://users.rowan.edu/~polikar/WAVELETS/WTpart1.html," 1996.

[71] J. Ma, Z. Wang, M. Vo, and L. Luu, "Parameter discretization in two-

dimensional continuous wavelet transform for fast fringe pattern analysis,"

Optics Info Base, pp. 6399-63408, 2011.

[72] G. Holmes, A. Donkin, and I. H. Witten, "WEKA: a machine learning

workbench," in Proceedings of the Proceedings of the 1994 Second Australian

and New Zealand Conference on Intelligent Information Systems. , 1994, pp.

357-361.

[73] M. Hassan, A. Chaudhry, A. Khan, and S. Riaz, "An Optimized Fuzzy C-

Means Clustering with Spatial Information for Carotid Artery Image

Segmentation,," in Proceedings of the IBCAST, Islamabad, 2011.

[74] R. Polikar, "Ensemble based systems in decision making," IEEE Circuits and

Systems Magazine, vol. 6, pp. 21-45, 2006.

http://users.rowan.edu/~polikar/WAVELETS/WTpart1.html,

138

[75] M. S. El Sherif and M. S. Abdel Samee, "Pattern recognition using neural

networks that learn from fuzzy rules," in Proceedings of the Proceedings of the

37th Midwest Symposium on Circuits and Systems., 1994, pp. 599-602.

[76] L. Rudasi and S. A. Zahorian, "Pattern recognition using neural networks with

a binary partitioning approach," in Proceedings of the IEEE Proceedings of

Southeastcon, Williamsburg, VA, 1991, pp. 726-730.

[77] H. Takahashi, N. Pecharanin, Y. Akima, and M. Sone, "The reliability of

neural networks on pattern recognition," in Proceedings of the IEEE World

Congress on Computational Intelligence.,, 1994, pp. 3067-3070.

[78] D. Greenhil and E. R. Dvvies, "Relative effectiveness of neural networks for

image noise suppression," in Proceedings of the Pattern Recognition in

Practice, 1994, pp. 367-378.

[79] T. M. Jochem, D. A. Pomerleau, and C. E. Thorpe, "Vision-based neural

network road and intersection detection and traversal," in Proceedings of the

IEEE/RSJ International Conference on Intelligent Robots and Systems 95

Human Robot Interaction and Cooperative Robots, , Pittsburgh, PA, 1995, pp.

344-349.

[80] D. Svozil, V. r. Kvasnicka, and J. Ã. Pospichal, "Introduction to multi-layer

feed-forward neural networks," Chemometrics and Intelligent Laboratory

Systems, vol. 39, pp. 43-62, 1997.

[81] U. Anders and O. Korn, "Model selection in neural networks," Neural

Networks, vol. 12, pp. 309-323, 1999.

[82] C. E. Shanoon, "A Mathematical Theory of Communication," Bell System

Technical Journal vol. 27, pp. 379-423, 1948.

[83] M. Filippo, Z. Guang, and S. S. Jasjit, "Review: A state of the art review on

intima-media thickness (IMT) measurement and wall segmentation techniques

for carotid ultrasound," Comput. Methods Prog. Biomed., vol. 100, pp. 201-

221, 2012.

[84] M. Filippo, M. M. Kristen, S. Luca, U. R. Acharya, L. Giuseppe, Z. Guang, H.

Sin Yee Stella, T. A. Anil, C. H. Suzanne, N. Andrew, and S. S. Jasjit,

"Ultrasound IMT measurement on a multi-ethnic and multi-institutional

database: Our review and experience using four fully automated and one semi-

automated methods," Comput. Methods Prog. Biomed., vol. 108, pp. 946-960,

2012.

[85] A. Chaudhry, M. Hassan, A. Khan, J. Kim, and T. Tuan, "Automatic

Segmentation and Decision Making of Carotid Artery Ultrasound Images," in

139

Intelligent Autonomous Systems, vol. 194, Advances in Intelligent Systems

and Computing: Springer Berlin Heidelberg, 2012, pp. 185-196.

[86] F. S. Donald, "Probabilistic neural networks," Neural Netw., vol. 3, pp. 109-

118, 1990.

[87] S. H. Park, J. M. Goo, and C.-H. Jo, "Receiver Operating Characteristic

(ROC) Curve: Practical Review for Radiologists," Korean Journal of

Radiology, vol. 5, pp. 11-18, 2004.

[88] R. Gallea, E. Ardizzone, R. Pirrone, and O. Gambino, "Three-dimensional

Fuzzy Kernel Regression framework for registration of medical volume data,"

Pattern Recognition, vol. 46, pp. 3000-3016, 2013.

[89] S. Borman, "The Expectation Maximization Algorithm – A Short Tutorial,":

Internet: http://www.seanborman.com/publications/EM_algorithm.pdf., Jan. 9,

2009, [Dec 20, 2013].

[90] M. S. Mahmoud and H. M. Khalid, "Expectation maximization approach to

data-based fault diagnostics," Information Sciences, vol. 235, pp. 80-96, 2013.

[91] S. Mitra and J. Basak, "FRBF: A Fuzzy Radial Basis Function Network,"

Neural Computing & Applications, vol. 10, pp. 244-252, 2001.

[92] S. Ovchinnikov, "Similarity relations, fuzzy partitions and fuzzy ordering,"

Fuzzy Sets and Systems, vol. 40, pp. 107-126, 1991.

[93] L. A. Zadeh, "Fuzzy sets and their application to pattern classification and

clustering analysis," in Fuzzy sets, fuzzy logic, and fuzzy systems: World

Scientific Publishing Co., Inc., 1996, pp. 355-393.

[94] N. Cristianini and J. S. Taylor, An Introduction to Support Vector Machines

and other kernel-based learning methods: Cambridge University Press, 2000.

[95] P. Minghao, L. Heon Gyu, P. Couchol, and R. Keun Ho, "A data mining

approach for dyslipidemia disease prediction using carotid arterial feature

vectors," in Proceedings of the IEEE, 2nd International Conference on

Computer Engineering and Technology (ICCET), Chengdu, 2010, pp. 171-

175.

[96] C. Chih-Chung and L. Chih-Jen, "LIBSVM: A library for support vector

machines," ACM Trans. Intell. Syst. Technol., vol. 2, pp. 1-27, 2011.

[97] K. B. Zayanthi and R. S. D. W. Banu, "Carotid Artery Boundary Extraction

Using Segmentation Techniques: A Comparative Study," in Proceedings of

http://www.seanborman.com/publications/EM_algorithm.pdf.

140

the Sixth International Conference on Information Technology: New

Generations Las Vegas, NV, 2009, pp. 1290-1295.

[98] R. K. S. Kwan, A. C. Evans, and G. B. Pike, "An Extensible MRI Simulator

for Post-Processing Evaluation," in Proceedings of the Visualization in

Biomedical Computing, Lecture Notes in Computer Science, Hamburg,

Germany, 1996, pp. 135-140.

[99] W. M. Rand, "Objective criteria for the evaluation of clustering methods,"

Journal of the American Statistical association, vol. 66, pp. 846-850, 1971.

[100] M. Aksam Iftikhar, A. Jalil, S. Rathore, and M. Hussain, "Robust brain MRI

denoising and segmentation using enhanced non-local means algorithm,"

International Journal of Imaging Systems and Technology, vol. 24, pp. 52-66,

2014.

[101] D. L. Pham, C. Xu, and J. L. Prince, "A Survey of Current Methods in

Medical Image Segmentation," in Annual Review of Biomedical Engineering,

vol. 2, 2000, pp. 315-338.

[102] N. Otsu, "A Threshold Selection Method from Gray-Level Histograms," IEEE

Transactions on Systems, Man., and Cybernetics., vol. 9, pp. 62-66, 1979.

[103] M. Kass, A. Witkin, and D. Terzopoulos, "Snakes: Active contour models,"

International Journal of Computer Vision, vol. 1, pp. 321-331, 1988.

[104] J. Liang, T. McInerny, and D. Terzopoulos, "United Snakes," in Proceedings

of the IEEE, Int. Conf. Computer Vision, 1999, pp. 933-940.

[105] P. Brigger, J. Hoeg, and M. Unser, "B-spline snakes: a flexible tool for

parametric contour detection," IEEE Transaction on Image Processing, vol. 9,

pp. 1484-1496, 2000.

[106] L. Weruaga, R. Verdu, and J. Morales, "Frequency domain formulation of

active parametric deformable models," IEEE Transaction on Pattern Analysis

and Machine Intelligence, vol. 26, pp. 1568-1578, 2004.

[107] M. Unser, Splines: "A perfect fit for medical imaging, Progress in Biomedical

Optics and Imaging", 2002.

[108] S. K. Pal and P. Mitra, Pattern Recognition Algorithms for Data Mining:

Scalability, Knowledge Discovery, and Soft Granular Computing: Chapma,

Hall, Ltd., 2004.

141

[109] D. F. Specht, "Probabilistic neural networks and the polynomial Adaline as

complementary techniques for classification," IEEE Transaction on Neural

Networks, vol. 1, pp. 111-121, 1990.

robust medical image segmentation for accurate computer...

Documents