correlating carotid imaging and phylogenetic trees...
TRANSCRIPT
CORRELATING CAROTID IMAGING AND
PHYLOGENETIC TREES FOR THE PRE AND POST
ANALYSIS OF GENETIC ISCHEMIC STROKES _____________________________________________________________________
A THESIS SUBMITTED TO LAHORE COLLEGE FOR WOMEN UNIVERSITY IN
PARTIAL FULFILMENT OF THE REQUIREMENTS FOR THE DEGREE OF
DOCTOR OF PHILOSOPHY IN COMPUTER SCIENCE
By
HUMA IFTIKHAR
Registration No.: 07-M/LCWU-12446
________________________________________________________________________
DEPARTMENT OF COMPUTER SCIENCE
LAHORE COLLEGE FOR WOMEN UNIVERSITY, LAHORE
PAKISTAN
2015
CERTIFICATE
This is to certify that the research work described in this thesis submitted by Ms. Huma
Iftikhar to Department of Computer Science, Lahore College for Women University has
been carried out under my direct supervision. I have personally gone through the raw data
and certify the correctness and authenticity of all results reported herein. I further certify that
thesis data have not been used in part or full, in a manuscript already submitted or in the
process of submission in Partial/complete fulfillment of the award of any other degree from
any other institution or home or abroad. We also certified that the enclosed manuscript, has
been to paid under my supervision and I endorse its evaluation for the award of PhD degree
through the official procedure of University.
____________________________
Dr. Muhammad Abuzar Fahiem
Supervisor
Date:
Verified By
________________
Dr. Muhammad Abuzar Fahiem
Chairperson
Department of Computer Science
Stamp
_________________
Controller of Examination
Stamp
Date: ___________
Dedicated to
My Parents
Muhammad Iftikhar Ali & Farhana Iftikhar For their unconditional love and support
My loving Husband Tauseef Ali Sulehri
For giving me my identity. Who wipes out the sense of time, memory of a difficult beginning and fear of an end in me.
My kids
Areesha, Marva, Shafay, Saad and Eshaal Who are indeed treasures from Allah. They have made me
complete.
ACKNOWLEDGMENTS
I’m grateful to Allah for giving me courage and to enlighten my path. I am being blessed by
Him to have caring and loving people around me.
My supervisor, Dr. Muhammad Abuzar Fahiem has been a great source of guidance. He has
been providing constant support and guidance through each step of this research work. This
work could never have been completed without his earnest supervision.
I would like to pay thanks to my parent university, Lahore College for Women University,
Lahore, Pakistan, that played a very important role throughout my studies as well as in the
completion of my PhD.
I would like to thank HapMap consortium (www.hapmap.org/) for providing data for this
research.
I would like to thank my friends for bearing with me in the hard times especially Saima for
spending her precious time to help me and support me.
Most of all, I owe my degree to my husband, kids, family and in-laws, without their support
and patience, achieving my goal was impossible. I want to express a special gratitude to my
mother-in-law and my parents for their moral support and encouragement.
Huma Iftikhar.
CONTENTS
Title Page No.
List of Tables i
List of Figures ii
List of Abbreviations iii
Abstract vi
Chapter 1: Introduction 1
1.1 Types of Strokes 1
1.2 Ischemic Strokes 2
1.2.1 Subtypes of Ischemic Strokes 2
1.2.2 Causes of Ischemic Strokes 3
1.2.3 Prevention and Diagnostics 3
Chapter 2: Review of Literature 6
2.1 Ischemic Stroke Risk Estimation Approaches 6
2.1.1 Medical Image Analysis based Ischemic Stroke
Risk Estimation
7
2.1.1.1 Image Acquisition 7
2.1.1.1.1 Carotid Duplex Ultrasound
(CDU)
7
2.1.1.1.2 Computed Tomography
Angiography (CTA)
7
2.1.1.1.3 Magnetic Resonance
Angiography (MRA)
8
2.1.1.1.4 Cerebral Angiography (CAG) 8
2.1.1.1.5 Digital Subtraction
Angiography (DSA)
8
2.1.1.2 CA and IMT Segmentation Techniques 9
2.1.1.2.1 Edge Tracking and Gradient 9
2.1.1.2.2 Dynamic Programming 10
2.1.1.2.3 Active Contours 10
2.1.1.2.4 Nakagami Mixture Modeling 11
2.1.1.2.5 Hough Transform 11
2.1.1.2.6 Integrated Approaches 11
2.1.1.3 Plaque Segmentation Techniques 12
2.1.1.3.1 Discrete Dynamic Contour 12
2.1.1.3.2 Kalman Filters 12
2.1.1.3.3 Balloon 12
2.1.1.3.4 Canny Edge Detection 13
2.1.1.3.5 Morphological Operations 13
2.1.1.4 Features 13
2.1.2 Genetic Data Analysis based Ischemic Stroke
Risk Estimation
15
2.1.2.1 Phylogenetic Trees 16
2.1.2.1.1 Phylogenetic Data 16
2.1.2.1.2 Phylogenetic Tree
Construction Methods
17
2.1.3 Classifiers 18
2.2 Comparison 20
2.2.1 Comparison of Different Imaging Techniques 20
2.2.2 Comparison of Different Approaches for Ischemic
Stroke Risk Estimation
27
Chapter 3: Proposed Approach- Ischemic Stroke Risk Estimation
Using Carotid Imaging
38
3.1 Materials and Methods 39
3.1.1 Phase-I: Preprocessing 44
3.1.2 Phase-II: Intima-Media Segmentation 47
3.1.3 Phase -III: Estimation 53
3.2 Results 58
3.3 Discussion 63
3.4 Conclusion 65
Chapter 4: Proposed Approach- Genetic Data Based Ischemic Stroke
Classification
71
4.1 Materials and Methods 84
4.1.1 Data 84
4.1.2 Method 84
4.1.3 Classification 88
4.1.3.1 Bayes Net 88
4.1.3.2 Naïve Bayes 88
4.1.3.3 IBk 88
4.1.3.4 AdaBoostM1 89
4.1.3.5 Classification via Regression 89
4.1.3.6 J48 90
4.1.3.7 Random Forest 90
4.1.3.8 Bagging 90
4.1.3.9 Multilayer Perceptron 91
4.2 Results and Discussion 91
4.3 Conclusion 99
Chapter 5: Analysis & Discussion- Correlating Phylogenetic Trees 100
5.1 Phylogenetic Tree Construction Methods 101
5.1.1 Unweighted Pair Group Method using Arithmetic
Averages (UPGMA)
101
5.1.2 Neighbor Joining (NJ) 102
5.1.3 Maximum Parsimony (MP) 102
5.1.4 Maximum Likelihood (ML) 102
5.2 Softwares and Tools Available for Phylogenetic Tree 103
Construction
5.3 Bioinformatics Databanks 104
5.4 Materials and Methods 106
5.4.1 Data 106
5.4.2 Method 107
5.5 Results and Discussion 121
5.6 Conclusion 125
Chapter 6: Conclusion & Future Recommendations 126
6.1 Conclusion 126
6.2 Future Recommendations 126
References 127
Plagiarism Report viii
List of Publications and Reprints ix
i
List of Tables
Table No. Title Page No.
2.1 Comparison of different medical imaging techniques for the
carotid artery
21
2.2 Review of different stenosis and IMT estimation approaches 28
3.1 Dataset details 40
3.2 Comparison of manual and proposed approach measurements
for 100 carotid artery ultrasound images
59
3.3 Wilcoxon ranksum test computed for Expert 1, Expert 2 and
Automatic measurements for 100 carotid artery ultrasound
images
60
3.4 Difference computed for classification results of Expert 1,
Expert 2 and Automatic measurements for 100 carotid artery
ultrasound images
63
3.5 Comparison of proposed approach with existing approaches 67
4.1 Stroke associated genes/ locus, SNP id/ haplotype and
corresponding SNPs
77
4.2 Details of dataset 84
4.3 Sample genetic data for randomly chosen subjects 85
4.4 Classification results for genetic data using different classifiers 93
5.1 Phylogenetic tools and their implementation methods 103
5.2 Details of selected HapMap population data 106
5.3 Stroke associated genes/ locus, SNP id/ haplotype and
corresponding allele risk
108
5.4 Allele frequencies for the SNPs from sample population data 110
5.5 Percent allele frequencies for the SNPs from sample population
data
114
5.6 Distance matrix for all populations 119
ii
List of Figures
Figure No. Title Page No.
3.1 Layers of carotid arterial wall 38
3.2 Block diagram of proposed approach 42
3.3 Detailed working of proposed approach 43
3.4 Image processing steps on carotid ultrasound images 52
3.5 Decision tree for classification of stenosis and ischemic stroke
risk
57
3.6 Bland-Altman plots of (a) Expert1 NM1 versus Automatic
measurements by proposed approach NA (b) Expert2 NM2
versus Automatic measurements by proposed approach NA
62
4.1 Identified SNPs causing stroke risk 76
4.2 Comparison of % accuracy of classifiers using genetic data 96
4.3 Comparison of % specificity of classifiers using genetic data 97
4.4 Comparison of % sensitivity of classifiers using genetic data 98
5.1 Phylogenetic tree using our distance matrix 120
5.2 Comparison of % allele frequency of all sample populations 122
5.3 Combined allele frequencies of all sample populations 123
5.4 Constructed phylogenetic tree using FST matrix as calculated
by Altshuler et. al.[201]
124
iii
List of Abbreviations
ANN Artificial Neural Networks
ASM Angular Second Moment
BMI Body Mass Index
CA Carotid Artery
CAG Cerebral Angiography
CDU Carotid Duplex Ultrasound
CIMT Carotid Intima Media Thickness
CT Computed Tomography
CTA Computed Tomography Angiography
CV Coefficient of Variation
DDP Dual Dynamic Programming
DGV Database of Genomic Variants
DNA DeoxyriboNucleic Acid
DP Dynamic Programming
DPAD Detail Preserving Anisotropic Diffusion
DSA Digital Subtraction Angiography
EGA European Genome Phenome Archive
EMBL European Molecular Biology Laboratories
FDTA Fractal Dimension Texture Analysis
FOAM First Order Absolute Moment
FOM Figure of Merit
FPS Fourier Power Spectrum
iv
GLDS Gray Level Difference Statistics
GVF Gradient Vector Flow
GWAS Genome-Wide Association Studies
HGMD Human Gene Mutation Database
HMM Hidden Markov Model
HT Hough Transform
IDM Inverse Difference Moment
IM Intima Media
IMT Intima Media Thickness
IOE Intra-Observer Error
ISGS Ischaemic Stroke Genetics Study
LI Lumen-Intima
MA Media Adventitia
MBPN Multilayer Back Propagation Network
ML Maximum Likelihood
MLE Maximum Likelihood Estimation
MLP Multilayer Perceptron
MP Maximum Parsimony
MRA Magnetic Resonance Angiography
MRI Magnetic Resonance Imaging
mRNA Messenger RNA
NGTDM Neighborhood Gray Tone Difference Matrix
NJ Neighbor Joining
v
OMIM Online Mendelian Inheritance in Man
RF Radio Frequency
RNA Ribonucleic Acid
ROI Region of Interest
rRNA Ribosomal RNA
SF Statistical Features
SGLDM Spatial Gray Level Dependence Matrices
SNP Single Nucleotide Polymorphism
SVM Support Vector Machine
TEM Texture Energy Measures
TIA Transient Ischemic Attack
UPGMA Unweighted Pair Group Method using arithmetic Averages
WHO World Health Organization
vi
Abstract
Ischemic stroke is the most commonly occurring type of stroke and one of the most communal
causes for disability and death in the world as per World Health Organization. Multiple factors
such as hypertension, diabetes, arterial fibrillation, heart diseases, transient ischemic strokes, etc.
contribute to ischemic stroke susceptibility. There is a compelling need for follow up checkups
and post analysis to prevent further strokes. Apart from clinical tests, a lot of research is being
carried out on computer based automated techniques and mechanisms for estimation of ischemic
stroke risk. Ultrasound images of the carotid artery are used for development of noninvasive
image based methods for stroke risk estimation however; carotid artery morphology, noise and
artifacts in the ultrasound images can lead to false classification.
Carotid intima media thickness is an indicator of future ischemic stroke. In this research, we
have proposed an automatic ischemic stroke risk estimation approach using carotid intima media
thickness from longitudinal carotid B-mode ultrasound images. Based on carotid intima media
thickness, a classification scheme is proposed to associate the carotid artery stenosis with
ischemic stroke risk. The proposed approach is tested and clinically validated on a data set of
100 longitudinal ultrasound images of the carotid artery. There is no significant difference
between intima media thickness measurements obtained using our approach and the manual
measurements by experts. The intra-observer error of 0.088, a Coefficient of Variation of
12.99%, Bland-Altman plots with small differences between experts (0.01 and 0.03 for Expert 1
and Expert 2, respectively) and Figure of Merit of 98.5% are obtained. The proposed approach
makes the risk estimation process automatic and yet reduces the risk of subjectivity and operator
variability for intima media thickness measurement.
Additionally, some of stroke cases are suspected to be genetic as the patients do not suffer from
vii
the conventional risk factors. Extensive research has been conducted to investigate the unknown
factors other than the conventional ones and their relationship with genetics. We have analyzed
genotype data for stroke risk estimation. Nine classification models are used on the SNPs data to
analyze and classify individuals. An accuracy of 88.16% is achieved by the proposed approach.
Ischemic stroke risk has been correlated with genetic distances. For this purpose phylogenetic
trees have been used. Analysis suggests that given two populations might be genetically close
but they might be far with respect to ischemic stroke risk.
Proposed research has addressed both the medical image analysis and genetic data analysis for
stroke risk estimation. The proposed approach has achieved higher accuracy, specificity and
sensitivity values when compared to existing approaches.
CHAPTER NO. 1
INTRODUCTION
1
Introduction
Stroke is caused due to deprivation of oxygen to the brain that causes brain cells death. Poor
blood flow causes oxygen deprivation in the brain. According to the World Health
Organization (WHO) [1], 15 million people suffer from stroke worldwide each year. Of these, 5
million die and another 5 million are permanently disabled. In 2013 stroke was the second major
cause of death [2]. Stroke is responsible for 6.4 million of the total deaths in year 2013 i.e. 12%
of the total deaths. Almost 67% of the strokes patients are 65 years or above. 50% patients who
suffer from stroke live for a maximum period of one year after the stroke occurrence.
1.1 Types of Strokes
There are three major types of strokes:
1. Ischemic Stroke
Blood clot causes ischemic stroke. This type of stroke accounts for 87% of all of the
stroke cases[3]. Fatty deposits lining the blood vessels are the major source of
obstruction causing ischemic strokes.
2. Hemorrhagic Stroke
The second type of strokes i.e. hemorrhagic strokes are caused by bleeding in brain due
to rupture of blood vessels. It accounts for 13% of the stroke cases. Weakened blood
vessels rupture and bleed in the brain and cause compression in the brain tissues.
Uncontrolled hypertension is the major cause of hemorrhagic strokes.
3. Transient Ischemic Attack (TIA)
TIAs are also known as mini strokes as they are mini episodes of stroke. These are
2
caused by temporary blockage of a vessel. They are taken as warning of an upcoming
stroke. TIA occurrence is rapid and the effects last for a short period of time. They
usually do not cause permanent injury or loss.
Our research work is focused on ischemic strokes and its risk estimation. Ischemic strokes are
discussed in detail in the next section.
1.2 Ischemic Strokes
Ischemic strokes [4] occur due to thrombosis[5], embolism [6] or atherosclerosis [7, 8]. An
individual affected with ischemic stroke may lose the ability to move one side of the body,
speak, see, eat and drink. It also increases the individual’s chances for heart attack and heart
failure [6]. The damage caused by ischemic stroke may be temporary or permanent. It is the
leading cause of long-lasting disability, long lasting injury and death. Studies confirm that an
individual who has suffered from ischemic stroke is at high risk of having more strokes [9-11].
It is a common perception that stroke usually occurs in old people but statistics show that 28%
of the strokes are in the individuals younger than 65 years of age. Studies have proven that
stroke is preventable and treatable provided needful is done in time [12]. Preventive measures
can preclude 80% of the strokes.
1.2.1 Subtypes of Ischemic Strokes
There are two major subtypes of ischemic strokes, namely:
1. Embolic Strokes
2. Thrombotic Strokes
In embolic stroke an emboli or a clot is formed in some part of body that eventually travels to
the brain and forms blockage. This blockage is then responsible for stroke. However, thrombotic
3
strokes are caused by thrombus or ruptured plaque that causes a blockage in the artery that
supplies blood to the brain.
1.2.2 Causes of Ischemic Strokes
Stroke is a complex disease as there are many factors that play role in the stroke susceptibility.
The risk factors are either modifiable or unmodifiable. Some of the major modifiable risk factors
for stroke are:
1. Hypertension
2. Diabetes
3. Arterial fibrillation
4. Smoking
5. Heart diseases like coronary artery disease, valve defects, enlargement of one of the
heart’s chambers.
6. TIAs
7. Cholesterol imbalance
8. Physical inactivity
9. Obesity
Above listed factors can be modified with medical treatment or lifestyle changes. Some of the
unmodifiable or non-amendable risk factors [7] are:
1. Age
2. Gender
3. Race
1.2.3 Prevention and Diagnostics
Two things that can help prevent stroke and the risk of death or disability from it are to control
4
the factors that cause stroke and to lookout for the stroke warning signs. Medical imaging
techniques are useful in detection of any changes in the blood flow to the brain. The invisible or
inheritable factors can be diagnosed using genetic data. The methods used for diagnostics and
risk estimation include:
1. Medical Image Analysis
2. Gene Data Analysis
Once stroke occurs less can be done. After the occurrence of stroke the nature and severity of
damage to the brain and its functions is accessed. Stroke can result in minor impairments, severe
impairments or in severe cases death. Rehabilitation assistance starts shortly once the patient is
stable (usually in 24 to 48 hours). In last decade intense research has been done to identify the
apparent causes of strokes and their prevention.
The need is felt to devise automated algorithms that are helpful in stroke risk assessment.
Medical images are used to assess the stroke risk as visible changes in the arteries can be
analyzed. Genetic data can be used for risk identification of an individual by analyzing the risk
alleles. Genetic data may aid to pre-diagnose individuals at high risk of ischemic stroke.
Genotypes as well as phenotypes contribute highly to the risk of stroke. Further, the individuals
who already have suffered from ischemic stroke are at very high risk of having stroke again. So
there is a compelling need of post analysis and follow up checkups to prevent further strokes.
Intima media thickness (IMT), the atherosclerotic carotid plaque, severity of stenosis due to
atherosclerotic plaques and plaque characterization can be taken under consideration using
carotid imaging. Texture patterns extracted from the carotid ultrasound images can help in
analysis of the plaques and their characterizations for pre diagnosis of the ischemic strokes.
Plaques have different compositions and appearances, so texture patterns are helpful in
5
characterizing different types of plaques. Similarly carotid artery (CA) morphology, noise and
artifacts in the carotid images can lead to false classification. Phylogenetic trees can help to take
into account the genetic risk factors and gene mutations. The risk patterns can be extracted using
these trees from the sampled group of data. These patterns can facilitate improvement and
accuracy of risk estimation.
Our research mainly takes into account carotid images and genetic data for the risk estimation of
ischemic stroke for an individual. The primary focus is on the prediction accuracy. The objective
of our research work is to carry out a comprehensive analysis of ultrasound images of the CA
and genes data to pre diagnose the individuals at high risk of ischemic strokes. For effective
prevention and prognostic implications of ischemic strokes, real-time analysis of genetic data
and ultrasound imaging is used. The research work improves the process of correct identification
of individuals at high risk of ischemic stroke by contributing in the visual assessment procedure
conducted by the medical personals. It also facilitates early diagnosis and the assessment of the
stroke risk.
The thesis is organized as follows: chapter 2 comprises of comprehensive literature review,
comparative analysis of different carotid imaging techniques, various approaches in the field of
carotid image analysis for stroke risk identification and assessment and advancements done in
the risk assessment of an individual for genetic ischemic stroke. Proposed approach for ischemic
stroke risk identification using carotid imaging, classification and results are presented in
chapter 3. The proposed risk assessment of ischemic stroke of an individual using genetic data,
classification and results comprises chapter 4. Phylogenetic tree analysis for the risk estimation
of different populations and their comparison with traditional phylogenetic tree are given in
chapter 5. Chapter 6 concludes the thesis and provides recommendations for future work.
CHAPTER NO. 2
REVIEW OF LITERATURE
6
Review of Literature
Stroke and the risk of death or disability from it can be prevented by taking some precautionary
measures. The measures that can prevent stroke are to control the factors that are its root cause
and to lookout for the warning signs. However in many stroke cases the phenotypes are not
present. Researchers are of the opinion that strokes are strongly related to genes and are passed
on in generations. Even the risk factors that contribute to strokes are passed on genetically [7].
Genetics influence the inherited predisposition for certain diseases.
The real-time analysis of the genetic data and ultrasound imaging provide quick means to
qualitatively analyze the input data and draw meaningful interpretation. It subsequently helps in
analyzing the risk factors and takes into account the preventive measures for stroke.
Phylogenetic trees can be used for the representation of highly diverse, multidimensional data
sets. They are used to evaluate the findings of the genetic data analysis among various
populations.
Analysis of the images of the CA and of genetic data plays a key role when assessing the
ischemic stroke risk. This chapter gives a comprehensive survey on medical imaging techniques,
various medical image analysis based ischemic stroke risk estimation techniques, phylogenetic
trees and classifiers generally used for classification purposes. Comparisons of different CA
imaging techniques and ischemic stroke risk estimation approaches are also given in this
chapter. Our improved approach for ischemic stroke risk estimation is proposed keeping in view
the shortcomings and limitations of the existing techniques.
2.1 Ischemic Stroke Risk Estimation Approaches
There are two types of risk estimation approaches for ischemic strokes; medical image analysis
7
based approaches and genetic data analysis based approaches. Both of these approaches and the
steps involved are discussed in the following section.
2.1.1 Medical Image Analysis based Ischemic Stroke Risk
Estimation
Medical images are used in image based techniques for analysis. Analysis includes image
segmentation for CA and IMT followed by segmentation for plaque inside the CA. A feature set
is then extracted which is used for classification purpose. Tests that provide information about
the CA structure and the blood flow information are useful in estimating the stroke risk.
2.1.1.1 Image Acquisition
There are different medical imaging techniques which are being practiced. The most common
ones are Carotid Duplex Ultrasound (CDU), Computed Tomography Angiography (CTA),
Magnetic Resonance Angiography (MRA), Cerebral Angiography (CAG) and Digital
Subtraction Angiography (DSA).
2.1.1.1.1 Carotid Duplex Ultrasound (CDU)
CDU is used to observe the blood flow in the CA. It uses sound waves to produce images of the
CA. It combines the blood flow information with the traditional imaging of the carotid vessels.
The term duplex means that two modes of ultrasound are used. One is doppler and the other is
B-mode. Doppler evaluates the velocity and direction of the blood flowing inside the artery. The
B-mode obtains the image of the artery. This technique is the most frequently used technique for
the estimation of stenosis.
2.1.1.1.2 Computed Tomography Angiography (CTA)
CTA is used to see the blood flow in the blood vessels throughout the body. It makes use of the
8
Computed Tomography (CT) that uses x-rays and a computer system to generate the images of
the blood vessels. CT produces detailed images of blood vessels and soft tissues. Sometimes a
dye is injected using a catheter. As the images are captured by a rotating device to capture
images at different angles, hence projection images are obtained. These images can be seen in
different planes and projections.
2.1.1.1.3 Magnetic Resonance Angiography (MRA)
MRA is based on Magnetic Resonance Imaging (MRI). It is the MRI of the blood vessels. The
images are formed by a scan that uses magnetic field and pulses of radio wave energy. It makes
use of magnetic field to capture images of the blood vessels generally in the head and neck
region. Contrast material / dye may be used for clarity of blood vessels.
2.1.1.1.4 Cerebral Angiography (CAG)
CAG is also known as intra-arterial digital subtraction angiography. In CAG a contrast based
dye is injected through a catheter in the blood vessels. The catheter is moved all the way up to
the heart. X rays are used to get images of the blood vessels. This technique is invasive and
usually done after some other noninvasive method confirms the stenosis. People having diabetes
or kidney disease are at risk of having complications during the test. The dye injected can cause
a temporary damage to the kidneys.
2.1.1.1.5 Digital Subtraction Angiography (DSA)
DSA is the process in which an image is acquired before injecting the contrast dye in the blood
vessels and an image is taken after injecting the dye. The pre contrast image is subtracted from
the contrast image to remove the overlying structures other than the blood vessels. This method
uses an image intensifier. It is a type of fluoroscopy testing technique. Area of interest is exposed
to time-controlled x-rays to capture the images. DSA is replaced by CTA and rarely used in the
9
hospitals and imaging departments.
2.1.1.2 CA and IMT Segmentation Techniques
In this section some of the well-known and commonly practiced segmentation techniques for
CA and IMT are discussed in detail. IMT serves as an indicator of stroke risk and cardiovascular
diseases. IMT is directly associated with increased risk of stroke especially in elderly population
without any history of cardiovascular diseases. IMT is measured as the distance between the
Lumen-Intima (LI) and Media-Adventitia (MA) interfaces of the CA. IMT measurement
methods can be manual or computer assisted. Usually IMT is calculated manually by an
operator that introduces operator variability and requires clinical experience. Manual method is
time consuming and varies according to training and subjective judgment of the operator.
Computer aided methods are used to calculate IMT thickness to overcome these problems.
These methods mainly rely on the computerized segmentation of the LI and MA interfaces in
CA images. Different image segmentation algorithms have been used to segment CA, leading to
more accurate results. Computer aided methods are either semi-automatic or completely
automatic. Most of the computer aided methods proposed are semi-automatic [13-20] and may
require operator or user to manually provide the Region of Interest (ROI) or to manually
perform the initial segmentation or to provide the segmentation seed points or do manual
corrections of the system segmented images. On the other hand the fully automatic techniques
do not require user interaction [21-25].
Computer aided methods can be categorized according to the type of technique used for
segmentation:
2.1.1.2.1 Edge Tracking and Gradient
Edge tracking and gradient based methods generally make use of the intensity profile graphs
10
[16, 20] or image gradients [15, 19] for segmentation of LI and MA interfaces. Faita et al. [13]
proposed an improved gradient based approach using First Order Absolute Moment (FOAM)
edge operator. This approach was robust to noise but could not process curved and non-
horizontal CA. Another variation is multistep gradient based algorithm [26] that makes use of
intensity, intensity gradient and interface continuity of pixels to segment CA.
2.1.1.2.2 Dynamic Programming (DP)
These algorithms segment IMT from the image of CA and outperformed maximum gradient,
mathematical models and matched filter approaches in speed and continuity of boundary [27].
DP based techniques [28] generally make use of echo intensity, intensity gradient and boundary
continuity as weighted terms of a cost function that is to be minimized. These techniques
generally require training and in case of change of scanner, retraining is required for the system
to work properly. A variation is multiscale based dynamic programming technique [29] for CA
analysis that iteratively calculates exact location of the CA wall by coarse to fine location. Dual
Dynamic Programming (DDP) [30] is also used for CA segmentation.
2.1.1.2.3 Active Contours
Active contours also known as snakes are deformable models that adapt themselves using a
dynamic process that minimizes a global energy function. Snake is a deformable spline.
Constraint image forces attract the snake or the spline towards the object and the internal forces
resist deformation. Conventional snakes need some seed points or initial contour points from the
user. They are used for object tracking, shape recognition, segmentation and edge detection.
Snake model is also being used to detect the CA from B-Mode and sonography images [31-35].
Local image gradients are mostly used to model snake based algorithms [25, 36, 37]. One of the
approaches is to combine local statistics with snakes to segment CA [14, 38, 39]. Some
11
approaches use local statistics combined with snakes and fuzzy k-means classifier to segment
CA [40, 41].
2.1.1.2.4 Nakagami Mixture Modeling
A Mixture of Nakagami distributions can be used to model the brightness of Radio Frequency
(RF) envelop. Nakagami distribution is effective for modeling the Radio Frequency (RF)
ultrasound signal scattered by the artery wall layers. The parameters of the model are estimated
using EM algorithm. Nakagami modeling and stochastic optimization is used for CA
segmentation [42] and plaque segmentation from B-mode ultrasound images [43].
2.1.1.2.5 Hough Transform (HT)
HT is a technique used for feature extraction with its applications in image analysis, computer
vision and image processing. It works by finding imperfect instances of objects within a certain
class of shapes by a voting procedure. Earlier application of the HT included identification of
lines in an image. But it has been extended to other shapes including circles and ellipses.
Segmentation algorithm based on HT [22, 44-46] is also used to segment CA boundaries. Xu et
al. [35] proposed HT and dual snake model for CA segmentation.
2.1.1.2.6 Integrated Approaches
Integrated approaches proposed for CA segmentation use a combination of more than one
technique for CA segmentation. Molinari et al. [47] developed an approach using local intensity
maxima and fuzzy k-means classifier for CA segmentation. DelSanto et al. [48] proposed an
approach based on signal analysis, image gradients and active contours for CA segmentation
(CULEXsa). Molinari et al. [24] combined signal processing, snakes, fuzzy clustering,
probability based connectivity and morphological approaches to segment CA. Molinari et. al.
proposed an automated multiresolution edge snapper based segmentation technique using scale
12
space and statistical classification in multiresolution framework (CAMES) [49]. CALEXia [50],
CARES 3.0 [51] and CAMES [49] are all integrated technique based fully automated
approaches for CA segmentation.
2.1.1.3 Plaque Segmentation Techniques
The segmented CA and IMT images are further segmented to find the plaque inside the arteries.
Several different algorithms are generally used for this purpose.
2.1.1.3.1 Discrete Dynamic Contour
Contours in 2D images can be defined using discrete dynamic model. This model is a set of
connected vertices which is automatically modified by an energy minimizing process. Local
contour curvature determines the internal energy and image features determine the external
energy of the contour. Energy Entropy map generated from a large database of artery images
and initial seed point provided by an expert is used to generate the contour of inner CA [52].
2.1.1.3.2 Kalman Filters
Kalman filters are also known as linear quadratic estimators. It observes estimates for data
variables over time or space to estimate unknown variables. The data might contain noise and
other inaccuracies. It recursively analyzes real-time noisy data to produce optimal values for the
unknown variables. They are used in navigation and control of vehicles, space crafts and
aircrafts, robotic motion planning and control. Both temporal and spatial Kalman filters are used
in this technique to extract the boundaries of the CA and center of its walls to measure the
diameter of the artery [53].
2.1.1.3.3 Balloon
Balloon model has an extra force as of working for snakes model to extract the contour of an
object. This additional force is inflation that expands the snakes contour into the minima instead
13
of shrinking into it. This eliminates the problem of a snake shrinking inwards. Gill et al. [54]
proposed an algorithm to detect atherosclerotic plaque in CA using the triangular mesh based
balloon model proposed by Cohen [55].
2.1.1.3.4 Canny Edge Detection
Canny is a famous edge detector with applications in various fields. Canny edge detection
works by first smoothing the image with Gaussian and then computing the gradient magnitudes.
Finally edges are detected by double thresholding. Hamou & El-Sakka [56] proposed an
algorithm that used Canny edge detector with threshold parameters to segment CA plaque.
2.1.1.3.5 Morphological Operations
Morphological operations can be used to analyze and process shape based data. Edges of an
object in an image are produced using morphological gradient operation. An appropriate
structuring element is selected by the morphological edge detection algorithm for the processed
image. The structuring element makes use of the basic morphological theory including erosion,
dilation, opening and closing and their combined operations to extract edges from the image. A
multistage method was proposed by Abdel-Dayen & Sakka [57]. The proposed work generates
carotid boundaries in the form of small contours from ultrasound images. Different stages
include filtering, quantization, edge detection and edge enhancement.
2.1.1.4 Features
Segmented images are used to extract feature sets. The features commonly used for texture
analysis in medical images are statistical features (SF), spatial gray level dependence matrices
(SGLDM), gray level difference statistics (GLDS), neighborhood gray tone difference matrix
(NGTDM), texture energy measures (TEM), fractal dimension texture analysis (FDTA) and
fourier power spectrum (FPS).
14
1. Statistical Features (SF)
SF have many applications in image processing. These can be used for probabilistic
description and classification of images as well as for the quality estimation of the
images. These features describe the gray level histogram distribution without
considering the spatial dependence of the pixels. Different SF include mean, median,
variance etc.
2. Spatial Gray Level Dependence Matrices (SGLDM)
SGLDM are the most commonly used features [57] that are computed on the basis of
probability density functions. These density functions are second-order joint conditional
probability density functions. An intermediate matrix of measures is computed from
image. Features are defined as functions of the calculated intermediate matrix. The
commonly used features include angular second moment (ASM), contrast, correlation,
inverse difference moment (IDM), sum average, variance and entropy of the pixel values.
3. Gray Level Difference Statistics (GLDS)
GLDS are calculated using first order statistics of an image. Probability density of image
pixel pairs is estimated at a given distance having a certain absolute gray level difference
value. These features are calculated from difference between pairs of gray levels [58].
Commonly used features are contrast, ASM, entropy and mean.
4. Neighborhood Gray Tone Difference Matrix (NGTDM)
NGTDM is calculated using pixels in the neighborhood of the pixel under consideration
but excluding the pixel under consideration. NGTDM features comprise of coarseness,
contrast, busyness, complexity and strength as texture features[59].
15
5. Texture Energy Measures (TEM)
TEM are also known as Laws TEM. These features are calculated using local masks to
detect various types of textures in images. Energy of texture is computed using
convolution masks. The energy of texture for each pixel is represented by vectors. The
statistical features mean, standard deviation and entropy of these vectors are used in
place of images [60].
6. Fractal Dimension Texture Analysis (FDTA)
Fractal values at different scales are used in FDTA approaches to get image features.
Features are computed from the parameters of affine relations among different regions of
an image. One of the methods used to calculate FDTA features that are used to measure
the roughness of a surface by fractional Brownian motion model of an image. The
extracted features include roughness or smoothness of a surface calculated using Hurst
coefficients [61, 62].
7. Fourier Power Spectrum (FPS)
Fourier analysis can be used to study texture properties of images. The FPS reveals
coarseness, fineness and directionality of a texture. FPS calculates the radial and angular
sum to find out the nature of the surface i.e. whether the surface is coarse or fine [58].
2.1.2 Genetic Data Analysis based Ischemic Stroke Risk Estimation
Genetics play an important role in the predisposition of many diseases. Genes data analysis is
conducted using different techniques. One method is to use Phylogenetic data and Phylogenetic
trees.
16
2.1.2.1 Phylogenetic Trees
Phylogenetic trees [63] represent the evolutionary descent of different species or genes from a
common ancestor. They represent diversity among same species or genes of a common ancestor.
They are helpful for structuring classification and identifying the changes that took place in
genes over the course of time. Recently genetic [64, 65] and genomic [66] studies are greatly
contributing to the study of brain and genetically transmitted diseases. Advanced studies have
shown that certain genes can be mutated or deactivated [67] to reduce the risks of these diseases.
2.1.2.1.1 Phylogenetic Data
Different types of data are used to construct phylogenetic trees. The data selection depends on
the purpose for which the tree is to be constructed [68]. Data can be phenotypes, genome gene
ordered data, and nucleotide or protein sequenced data.
1. Phenotype Data
Phenotypes are the data which can be easily obtained by appearance. There are certain
heritable factors that clearly indicate the risk for ischemic stroke [69].
2. Genome Gene Ordered Data
Genome rearrangements in gene order data are used to construct phylogenetic trees
[70].
3. Nucleotide Sequenced Data
Deoxyribonucleic Acid (DNA) or Ribonucleic Acid (RNA) data are coded using an
alphabet for the four nucleotides. This data represents the genetic characters. The
phylogenetic trees are constructed using these DNA or RNA sequences as these provide
immense phylogenetic information[71, 72].
17
4. Protein Sequenced Data
Proteins are encoded into sequences based on amino acids. This data is also rich in
phylogenetic information and thus is very commonly used nowadays for phylogenetic
tree construction [73].
2.1.2.1.2 Phylogenetic Tree Construction Methods
Phylogenetic data which is in any of the above mentioned forms is used to construct
phylogenetic tree. Trees are most logical way to represent data for evolution representation.
There are many algorithms that are used for phylogenetic tree construction [74, 75]. All these
fall into three major classes:
1. Maximum Parsimony Methods (MP)
MP methods are also known as minimum evolution methods. MP method tries to
estimate a tree with most mutations during the evolution period using minimum
numbers of evolutionary steps. The trees predicted are the one that requires minimum
number of steps to generate the observed variation in sequences from the sequences of
the common ancestors. It uses the simplest and the most parsimonious explanation of an
observed variation.
2. Evolutionary Distances based Methods
These methods infer evolutionary relationships from similarity among organisms.
Organisms sharing an ancestor in recent past are believed to be more similar than the
organisms that have a common ancestor that is more ancient. These methods calculate
genetic distances between sequence alignments in the form of distance matrix. A
distance matrix is also known as a table of evolutionary distances. This matrix is used
18
to construct phylogenetic trees.
3. Maximum Likelihood Estimation (MLE) Principle
MLE is a method for estimation of the parameters of a statistical model. MLE principle
tries to estimate the tree that is optimal on some function. It requires sample data on
which the function is constructed. For a set of data and statistical model, MLE selects
the values of the model’s parameters such that when used the maximum likelihood
function is maximized. Probabilities are assigned to possible phylogenetic trees. A
substitution model is used to calculate the probability of a mutation.
2.1.3 Classifiers
Assigning a well-known and defined class or category to a new observation is known as
classification. Various classification algorithms have been implemented. Different types of
classifiers are used to classify the data based upon the application. In this section some of the
well-known classifiers are discussed in detail.
Linear Classifiers
Linear classifiers [76] make a classification decision based on the characteristics of the
objects. These characteristics are known as feature values and are fed to the classifier. A
linear predictor score function is used to calculate score for each possible category based
on feature vector and weights vector. The assigned class is the one with the highest score.
Support Vector Machines (SVM)
SVM [77] are supervised algorithms which are used for classification purpose and
regression analysis of data. SVM is a non-probabilistic binary linear classifier which tries
to predict the input data to be one of the two classes. SVM model is mapped so that
19
separate classes are divided by a clear gap when plotted i.e. they are easily separable.
Quadratic Classifiers
Quadratic classifiers [76] use quadratic discriminant analysis for classification. It
assumes measurements from each class to be normally distributed. The classification
result is assumed to be quadratic in nature. It is assumed that two classes are separable by
a quadric surface.
Kernel Estimation
Kernel estimation also known as k-nearest neighbors [78], is the most simple
classification algorithm. The classification decision is based on majority voting of
neighbors. The classification function is approximated locally resulting in sensitivity to
the local structure of the data.
Decision Trees Learning
Decision trees learning [77] is a vastly used machine learning and classification
algorithm. It uses a decision tree in which the classes are represented as leaves and the
feature conjunctions are labeled on the branches. A model is created that is used to
predict the class of an observation based on several input variables. The predicted
outcome is the class of the new observation.
Artificial Neural Networks (ANN)
ANN [77] also known as neural networks, works on the principle of human brain. It is
continuously changing its design as it adapts to variations during the learning process.
These are well suited for the complex datasets and for the situations where complex
relationships exist between inputs and outputs, or where one has to find patterns and
20
explore statistical structures with unknown joint probability distribution between
observed variables.
Bayesian Networks
Bayesian network [77] also known as belief network or directed acyclic graphical model.
It is a probabilistic graphical model. It uses directed acyclic graph to show the
conditional dependencies of the variables based on probabilities. The classification
decision is made on the basis of probabilities.
Hidden Markov Models (HMM)
HMM [76] assumes the problem to be classified to have unobserved or hidden states and
is closely related to optimal nonlinear filtering problem. In these models the states are
hidden by the output which is dependent on these states is visible. The sequence of
output tokens produced by the HMM can be used to get knowledge about the sequence
of states.
2.2 Comparison
A detailed comparison of existing medical imaging techniques for CA conducted is presented in
this section. Moreover a comprehensive review of research work conducted for ischemic stroke
risk estimation is given. Summarized tables of both comparisons are given as Table 2.1 and 2.2.
2.2.1 Comparison of Different Imaging Techniques
A comparison of commonly practiced medical imaging techniques for the CA analysis, their
advantages and limitations are given in Table 2.1
21
Table 2.1: Comparison of different medical imaging techniques for the carotid artery
Imaging CDU CTA MRA CAG DSA
Invasive × × × √ √
Radiation Exposure × √ × √ √
Expensive × × √ √ ×
Safety Risk None Minimal Minimal Significant Significant
Repetition Frequency
over Short Duration of
Time
Often Often Rare Rare Rare
Motion Sensitive × √ √ √ √
Sedative × × √ Sometime Sometime
22
Imaging CDU CTA MRA CAG DSA
Stenosis
50-
69%
≥ 70% ≥ 70% ≥ 70% - ≥ 70%
Sensitivity 93% 99% 95% 93% - 92.9%
Specificity 68% 86% 98% 97% - 81.9%
Accuracy 85% 95% 97% 95% - -
Overestimation of
Degree of Stenosis
√ √ √ × ×
Hair line Lumen
Detection
× √ × √ √
Risk of Allergic
Reaction
× √ √ √ √
23
Imaging CDU CTA MRA CAG DSA
Contrast Agents × √ × √ √
Contrast Dosage NA High NA High Low
Image Quality High High High
Artifacts
Image Speckle
Shadowing
Poor definition
of artery
boundaries
Motion artifacts
Turbulent flow
Phase wrapping
Maxwell terms
Laminar flow
Venetian blinds
Motion artifacts
Motion artifacts
Larynx artifact
Technology Sound Waves X-Rays
Magnetic Field and
Radio Frequency
Waves
X-Rays X-Rays
24
Imaging CDU CTA MRA CAG DSA
Usage
Blood flow
velocity
Severity of
Stenosis
Carotid index
Arterial &
Venous blood
flow
Anatomic image
of CA Lumen
Image of soft
tissue & bony
structure
Evaluation of
Extra cranial CA
Image carotid
arteries
Information about
disease process
Assess collaterals
Arterial and
Venous
occlusions
Arterial Stenosis
Cerebral
aneurysms
Effected by Metallic
Implants
× √ √ √ √
Performed in
Individuals with Renal
Insufficiency
√ × × × ×
25
Imaging CDU CTA MRA CAG DSA
Procedure Duration Short Short Long Long Short
Views of Carotid
Bifurcation
Limited Seen Seen Limited Limited
Operator Dependent √ × × √ √
Image Usually 2D 3D/4D 3D/4D 3D 3D
Blood Flow Information √ × √ √ √
SNR Low High High High High
Accuracy effected by
Carotid Calcification
√ √ × ×
26
Imaging CDU CTA MRA CAG DSA
Limitations Cannot assess
intracranial CA.
Superimposed
jugular veins and
arteries may hide
stenosis.
Evaluation of
small vessels is
difficult.
Risk of stroke.
Bones and muscle
tissue are present
in images.
Larynx artifact.
External carotid
or vertebral artery
overlying the
internal CA.
Poor arterial
contrast density.
27
. It is evident from the comparison that Ultrasound has many advantages; a noninvasive
technique has low cost and includes no radiation exposure. Moreover it can be performed in
patients with renal insufficiency and individuals with metallic implants. It does not need any
sedative to be given to the individual undergoing the test and has no safety risk. So ultrasound
proves to be a safe method for analysis of stenosis.
2.2.2 Comparison of Different Approaches for Ischemic Stroke Risk
Estimation
Intensive research has been done for the detection of stenosis and plaque in the CA in past
decade. A comprehensive comparison is given in Table 2.2. Different factors which are being
considered include input, genetics, segmentation technique, features, classifier and results. Input
type, it’s source, modality and sample size are being considered while reviewing the literature.
Input can be in the form of images or signals. Image can be whole or they can be of plaque only.
Input source can be from some database or from a laboratory or collected by the researcher for
experimentation. Modality may include ultrasound, MRI, CT etc. Features which are being
extracted for classification, their extraction techniques and methods used to reduce features are
also being reviewed. Classifier and its type are also being considered. Different results achieved
by different researchers are also mentioned.
28
Table 2.2: Review of Different Stenosis and IMT Estimation Approaches
Ref
eren
ce
Inp
ut
Typ
e
Inp
ut
Sou
rce
Mod
ali
ty
Sam
ple
Siz
e
Gen
etic
s
Au
tom
ate
d/
Sem
i
Au
tom
ate
d
Syst
em
Seg
men
tati
on
Tec
hn
iqu
e
Fea
ture
s
Extr
act
ed
Features
Mach
ine
Lea
rnin
g
Alg
ori
thm
/
Cla
ssif
ier
Cla
ssif
ier
Typ
e
Acc
ura
cy
Posi
tive
Pre
dic
t
Valu
e
Sen
siti
vit
y
Fals
e P
osi
tive
Fals
e N
egati
ve
Sp
ecif
icit
y
Oth
er
Fea
ture
Sel
ecti
on
Tec
hn
iqu
e
Fea
ture
Red
uct
ion
Tec
hn
iqu
e
Tsi
apar
as e
t al.
(20
12
) [7
9]
Imag
e
-
B m
ode
ult
raso
und
of
pla
qu
e
20
× - - Tex
ture
Fea
ture
s
DT
CW
T:
36
Z s
core
SVM Supervised 67.6 - 68.9 - - 72.1 -
FR
IT:
6
Z s
core
SVM Supervised 71.5 - 75.2 - - 72.6 -
FD
CT
: 44
Z s
core
SVM Supervised 79.3 - 78.2 - - 84.3 - F
req
uen
cy
Fea
ture
s
WP
-
SVM - 70.9 - 71.9 - - 75.2 -
Kem
ény
et
al.
(199
9)
[82
]
Colo
r-co
ded
po
wer
spec
tra
of
the
audib
le
Do
pp
ler
shif
t
Sel
f
Tra
nsc
ran
ial
Do
pp
ler
Ult
raso
und
282
√
Au
tom
atic
-
FF
T
- -
ANN Supervised - 56.7 73.4 - - - -
29
Ref
eren
ce
Inp
ut
Typ
e
Inp
ut
Sou
rce
Mod
ali
ty
Sam
ple
Siz
e
Gen
etic
s
Au
tom
ate
d/
Sem
i
Au
tom
ate
d
Syst
em
Seg
men
tati
on
Tec
hn
iqu
e
Fea
ture
s
Extr
act
ed
Features
Mach
ine
Lea
rnin
g
Alg
ori
thm
/
Cla
ssif
ier
Cla
ssif
ier
Typ
e
Acc
ura
cy
Posi
tive
Pre
dic
t
Valu
e
Sen
siti
vit
y
Fals
e P
osi
tive
Fals
e N
egati
ve
Sp
ecif
icit
y
Oth
er
Fea
ture
Sel
ecti
on
Tec
hn
iqu
e
Fea
ture
Red
uct
ion
Tec
hn
iqu
e
Ky
riac
ou
et
al.
(2
009
) [8
3]
Car
oti
d p
laq
ue
imag
es
Irv
ine
Lab
Ult
raso
und
im
ages
274
im
ages
×
Sem
i au
tom
atic
Man
ual
Mo
rph
olo
gic
al F
eatu
res
Mult
ilev
el b
inar
y
mo
rpho
logic
al
mo
del
PC
A
SVM Supervised 73.72
- 83.94
36.5 16.06
63.5 Acoustic shadows
are
excluded
Dir
ect
Gra
y s
cale
mo
rpho
logy
mo
del
-
SVM Supervised 66.79
- 54.01
20.44
45.99
79.56
Sto
itsi
s et
al.
(200
4)
[80
]
Sta
tic
imag
es, im
age
sequ
ence
s
Irv
ine
Lab
B-
mode
ult
raso
un
d
19 p
atie
nts
×
Au
tom
atic
AN
AL
YS
IS
(s/w
fo
r in
terp
reta
tion
of
med
ical
imag
es)
Tex
ture
fea
ture
s: 9
9
FO
S,
SO
S,
law
’s
TE
M,
FD
TA
AN
OV
A
Fuzzy c-
means
Unsupervis
ed
74 - - - - - Atherom
atous
plaques were
predicte
d.
Moti
on
feat
ure
: 2
MS
V,
MR
SV
AN
OV
A
Fuzzy c-
means
Unsupervis
ed
79 - - - - -
30
Ref
eren
ce
Inp
ut
Typ
e
Inp
ut
Sou
rce
Mod
ali
ty
Sam
ple
Siz
e
Gen
etic
s
Au
tom
ate
d/
Sem
i
Au
tom
ate
d
Syst
em
Seg
men
tati
on
Tec
hn
iqu
e
Fea
ture
s
Extr
act
ed
Features
Mach
ine
Lea
rnin
g
Alg
ori
thm
/
Cla
ssif
ier
Cla
ssif
ier
Typ
e
Acc
ura
cy
Posi
tive
Pre
dic
t
Valu
e
Sen
siti
vit
y
Fals
e P
osi
tive
Fals
e N
egati
ve
Sp
ecif
icit
y
Oth
er
Fea
ture
Sel
ecti
on
Tec
hn
iqu
e
Fea
ture
Red
uct
ion
Tec
hn
iqu
e
Tex
ture
+ M
oti
on
feat
ure
FO
S,
SO
S,
Law
’s
TE
M,
FD
TA
, M
SV
,
MR
SV
AN
OV
A
Fuzzy c-means
Unsupervised
84 - - - - -
Ch
rist
odou
lou e
t a
l. (
200
3)
[84
]
caro
tid
pla
que
u/s
im
age
Irv
ine
Lab
B m
ode
Ult
raso
und
230
im
ages
Sem
i au
tom
ated
Man
ual
Tex
ture
+ S
hap
e :
61
FO
S,
SG
LD
M (
mea
n),
SG
LD
M (
ran
ge)
, G
LD
S,
NG
TD
M, S
FM
, T
EM
,
FD
TA
, F
PS
, S
hap
e P
aram
eter
s
Maj
ori
ty V
ote
SOM Un supervised
66 - - - - - -
KNN Supervised 65.1 - - - - - -
Av
erag
ing
con
fid
ence
Mea
sure
s
SOM Un
supervised
73.1 - - - - - -
KNN Supervised 68.8 - - - - - -
All
SOM Un
supervised
68.8 - - - - - -
KNN Supervised 69.7 - - - - - -
31
Ref
eren
ce
Inp
ut
Typ
e
Inp
ut
Sou
rce
Mod
ali
ty
Sam
ple
Siz
e
Gen
etic
s
Au
tom
ate
d/
Sem
i
Au
tom
ate
d
Syst
em
Seg
men
tati
on
Tec
hn
iqu
e
Fea
ture
s
Extr
act
ed
Features
Mach
ine
Lea
rnin
g
Alg
ori
thm
/
Cla
ssif
ier
Cla
ssif
ier
Typ
e
Acc
ura
cy
Posi
tive
Pre
dic
t
Valu
e
Sen
siti
vit
y
Fals
e P
osi
tive
Fals
e N
egati
ve
Sp
ecif
icit
y
Oth
er
Fea
ture
Sel
ecti
on
Tec
hn
iqu
e
Fea
ture
Red
uct
ion
Tec
hn
iqu
e
Ky
riac
ou
et
al.
(2
006
) [8
5]
Gra
y s
cale
car
oti
d p
laq
ue
imag
es
Irv
ine
lab
Ult
raso
und
274
×
Sem
i au
tom
ated
Man
ual
Mo
rph
olo
gic
al f
eatu
re
Gra
y S
cale
Mo
rph
olo
gic
al A
nal
ysi
s
-
PNN Supervised 62.04
- 59.85
35.77
40.15
64.23
Acoustic shadows
are
excluded
SVM Supervised 63.1
4
- 62.7
7
36.5 37.2
3
63.5
PC
A
PNN Supervised 60.58
- 57.66
36.5 42.34
63.5
SVM Supervised 66.7
9
- 54.0
1
20.4
4
45.9
9
79.5
6
Ky
riac
ou
et
al.
(2
005
) [8
6]
Cli
nic
al d
ata
+ u
ltra
sound
im
ages
EU
BIO
ME
D I
I A
CS
RS
+ I
rvin
e
Lab
ora
tory
Lon
git
ud
inal
sca
ns
usi
ng d
up
lex
scan
nin
g a
nd
colo
r fl
ow
im
agin
g
Cli
nic
al d
ata:
1298
cas
es
+ U
/S:
274
im
ages
×
Sem
i au
tom
ated
Man
ual
340
cli
nic
al f
acto
rs +
54 T
extu
re
Fea
ture
s
SF
-
PNN Supervised 65.3 - - - - - Acoustic
shadows
are
excluded SVM Supervised 69.3 - - - - -
PC
A
PNN Supervised 65.3 - - - - -
SVM Supervised 70.1 - - - - -
32
Ref
eren
ce
Inp
ut
Typ
e
Inp
ut
Sou
rce
Mod
ali
ty
Sam
ple
Siz
e
Gen
etic
s
Au
tom
ate
d/
Sem
i
Au
tom
ate
d
Syst
em
Seg
men
tati
on
Tec
hn
iqu
e
Fea
ture
s
Extr
act
ed
Features
Mach
ine
Lea
rnin
g
Alg
ori
thm
/
Cla
ssif
ier
Cla
ssif
ier
Typ
e
Acc
ura
cy
Posi
tive
Pre
dic
t
Valu
e
Sen
siti
vit
y
Fals
e P
osi
tive
Fals
e N
egati
ve
Sp
ecif
icit
y
Oth
er
Fea
ture
Sel
ecti
on
Tec
hn
iqu
e
Fea
ture
Red
uct
ion
Tec
hn
iqu
e
SG
LD
M
-
PNN Supervised 71.2 - - - - -
SVM Supervised 69.7 - - - - -
PC
A
PNN Supervised 70.8 - - - - -
SVM Supervised 68.6 - - - - -
SF
+ N
GT
DM
-
PNN Supervised 62.8 - - - - -
SVM Supervised 71.2 - - - - -
PC
A
PNN Supervised 62.4 - - - - -
SVM Supervised 69.3 - - - - -
SF
+ S
GL
DM
+
TE
M +
NG
TD
M
-
PNN Supervised 65.3 - - - - -
SVM Supervised 70.8 - - - - -
33
Ref
eren
ce
Inp
ut
Typ
e
Inp
ut
Sou
rce
Mod
ali
ty
Sam
ple
Siz
e
Gen
etic
s
Au
tom
ate
d/
Sem
i
Au
tom
ate
d
Syst
em
Seg
men
tati
on
Tec
hn
iqu
e
Fea
ture
s
Extr
act
ed
Features
Mach
ine
Lea
rnin
g
Alg
ori
thm
/
Cla
ssif
ier
Cla
ssif
ier
Typ
e
Acc
ura
cy
Posi
tive
Pre
dic
t
Valu
e
Sen
siti
vit
y
Fals
e P
osi
tive
Fals
e N
egati
ve
Sp
ecif
icit
y
Oth
er
Fea
ture
Sel
ecti
on
Tec
hn
iqu
e
Fea
ture
Red
uct
ion
Tec
hn
iqu
e
PC
A
PNN Supervised 64.2 - - - - -
SVM Supervised 70.1 - - - - -
Ky
riac
ou
et
al.
(2
007
) [8
7]
Cli
nic
al r
isk f
acto
rs +
ult
raso
und
im
ages
AC
SR
S,
Irv
ine
Lab
Lon
git
ud
inal
sca
ns
usi
ng d
up
lex s
cann
ing a
nd
co
lor
flo
w i
mag
ing
18 c
lin
ical
fac
tors
, 2
74
im
ages
Sem
i au
tom
ated
IMT
seg
men
tati
on
: W
illi
ams
and
Sh
ah. A
ther
osc
lero
tic
pla
que
seg
men
tati
on
: L
ai
and
Chin
snak
e.
Tex
ture
Fea
ture
s
SF
+ S
GL
DM
+ G
LD
S+
NG
TD
M+
SF
M+
Law
’s T
EM
+ F
ract
als
PC
A
PNN Supervised 72.3 - 75.9 31.4 24.1 68.6 -
SVM Supervised 73 - 82.5 36.5 17.5 63.5
PNN Supervised 69.7 - 74.4 35 25.5 65
SVM Supervised 73.4 - 81 34.3 19 65.7
34
Ref
eren
ce
Inp
ut
Typ
e
Inp
ut
Sou
rce
Mod
ali
ty
Sam
ple
Siz
e
Gen
etic
s
Au
tom
ate
d/
Sem
i
Au
tom
ate
d
Syst
em
Seg
men
tati
on
Tec
hn
iqu
e
Fea
ture
s
Extr
act
ed
Features
Mach
ine
Lea
rnin
g
Alg
ori
thm
/
Cla
ssif
ier
Cla
ssif
ier
Typ
e
Acc
ura
cy
Posi
tive
Pre
dic
t
Valu
e
Sen
siti
vit
y
Fals
e P
osi
tive
Fals
e N
egati
ve
Sp
ecif
icit
y
Oth
er
Fea
ture
Sel
ecti
on
Tec
hn
iqu
e
Fea
ture
Red
uct
ion
Tec
hn
iqu
e
San
thiy
aku
mar
i et
al.
(2
011
) [8
1]
Ult
raso
und
car
oti
d a
rter
y i
mag
es
-
Ult
raso
und
100
im
ages
×
Au
tom
ated
Con
tou
r ex
trac
tio
n u
sing
ener
gy m
inim
izat
ion
pro
cess
.
Con
tou
rs
- -
ANN, MBPN
Supervised Normal :
96%,
Cardiovas
cular
Dise
ase:
90%,
Cerebrov
ascul
ar Dise
ase:
92%
- - - - - -
Ab
do
lmal
eki
et a
l. (
2005
) [8
8]
Lon
git
ud
inal
vie
w
-
Hig
h r
eso
luti
on B
-mo
de
ult
raso
und
, co
lor
Dop
ple
r im
ages
128
×
Sem
i A
uto
mat
ed
-
Qu
anti
tati
ve
feat
ure
s: 1
0
Ult
raso
nic
mea
sure
men
t
-
Logistic
Regressi
on
Model
Supervised - .94 - - - - -
35
Ref
eren
ce
Inp
ut
Typ
e
Inp
ut
Sou
rce
Mod
ali
ty
Sam
ple
Siz
e
Gen
etic
s
Au
tom
ate
d/
Sem
i
Au
tom
ate
d
Syst
em
Seg
men
tati
on
Tec
hn
iqu
e
Fea
ture
s
Extr
act
ed
Features
Mach
ine
Lea
rnin
g
Alg
ori
thm
/
Cla
ssif
ier
Cla
ssif
ier
Typ
e
Acc
ura
cy
Posi
tive
Pre
dic
t
Valu
e
Sen
siti
vit
y
Fals
e P
osi
tive
Fals
e N
egati
ve
Sp
ecif
icit
y
Oth
er
Fea
ture
Sel
ecti
on
Tec
hn
iqu
e
Fea
ture
Red
uct
ion
Tec
hn
iqu
e
Lam
bro
u e
t a
l. (
201
2)
[89
]
Ult
raso
und
im
ages
of
car
oti
d p
laq
ue
Irv
ine
Lab
-
274
×
Sem
i au
tom
ated
Man
ual
Tex
ture
: 7
fea
ture
set
s
SF,
SG
LD
M,
GL
DS
, N
GT
DM
, S
FM
, T
EM
, F
DT
A,
FP
S,
RU
NL
-
Confidence
Predict
Value
AN
N
Supervised
71.53
- - - - - -
k
-N
N
Supervi
sed
70.8 - - - - -
Mo
rph
olo
gic
al
Mult
i le
vel
app
roac
h, g
ray
sca
le
mo
rpho
logic
al a
nal
ysi
s
-
Confiden
ce Predict
Value
A
NN
Supervi
sed
72.2
6
- - - - -
S
V
M
supervis
ed
73.7
2
- - - - -
36
Ref
eren
ce
Inp
ut
Typ
e
Inp
ut
Sou
rce
Mod
ali
ty
Sam
ple
Siz
e
Gen
etic
s
Au
tom
ate
d/
Sem
i
Au
tom
ate
d
Syst
em
Seg
men
tati
on
Tec
hn
iqu
e
Fea
ture
s
Extr
act
ed
Features
Mach
ine
Lea
rnin
g
Alg
ori
thm
/
Cla
ssif
ier
Cla
ssif
ier
Typ
e
Acc
ura
cy
Posi
tive
Pre
dic
t
Valu
e
Sen
siti
vit
y
Fals
e P
osi
tive
Fals
e N
egati
ve
Sp
ecif
icit
y
Oth
er
Fea
ture
Sel
ecti
on
Tec
hn
iqu
e
Fea
ture
Red
uct
ion
Tec
hn
iqu
e
Ch
rist
odou
lou e
t a
l. (
201
0)
[90
]
Car
oti
d p
laq
ue
imag
e
Irv
ine
Lab
Ult
raso
und
274
×
Sem
i au
tom
ated
Man
ual
Mult
i re
gio
n H
isto
gra
m
- -
SOM 64.8 - - - - - -
k-NN supervised 63.1 - - - - -
37
Most commonly used classifiers for ischemic stroke risk estimation are ANN and SVM as
apparent from Table 2.2. Tsiaparas et al. [79] have used B mode ultrasound images of plaque
to achieve an accuracy rate of 79.3% using SVM for Texture features. Stoitsis et al. [80]
achieved classification accuracy of 84% based on texture features combined with motion
features using Fuzzy c-means clustering. Contours based classification using ANN and
Multilayer Back Propagation Network (MBPN) is used by Santhiyakumari et al. [81] to
achieve the highest accuracy of 96%. Contours, Texture and Motion features prove to give
best classification results.
We have discussed and compared many carotid imaging techniques, image processing and
classification techniques. All the discussed techniques have some advantages as well as
disadvantages. Factors that are of prime importance for automated analysis and classification
of the carotid images are the spatial resolution and quality of the image. The preferred
imaging technique should be low risk, noninvasive as well as low cost without much
compromise on the quality of the image. Carotid ultrasound is a preferred technique that has
all the stated properties. The challenge is to consider the features that contribute the most for
assessing the ischemic stroke risk.
CHAPTER NO. 3
PROPOSED APPROACH-
ISCHEMIC STROKE RISK
ESTIMATION USING CAROTID
IMAGING
38
Proposed Approach- Ischemic Stroke Risk
Estimation Using Carotid Imaging
CA atherosclerosis is one of the major reasons for ischemic stroke. A vast range of medical
imaging techniques are available for CA assessment including CDU [91, 92], CT [93], MRA
[94, 95], CA [96] and DSA [97]. Carotid ultrasound is widely used for diagnosis of carotid
diseases and risk estimation of stroke due to its noninvasive nature, being low cost, having
short examination time, aiding accurate approximation of carotid intima media thickness
(CIMT) values and continuous improvement in quality of ultrasound images.
Nighoghossian et. al. [98] has reviewed different CA assessment techniques and discussed
the future prospects of the research field. A detailed comparison of these medical imaging
techniques for CA assessment is given in the previous chapter.
IMT serves as an indicator of stroke risk and cardiovascular diseases. IMT is directly
associated with increased risk of stroke especially in elderly population without any history
of cardiovascular diseases. IMT is measured as the distance between the LI and MA
interfaces of the CA. The layers of carotid arterial walls can be seen in figure 3.1.
IntimaMedia
Adventitia
Lumen
Figure 3.1: Layers of carotid arterial wall
39
The inner layer is the intima which consists of endothelial cells. Central layer is the media
which consists of smooth cells, collagen, elastin and proteoglycan. Outermost wall is
adventitia which consists of fibroblasts, collagen and elastin. Each wall is separated by an
elastic membrane. In the middle of the CA is the blood known as lumen that flows between
the CA.
We have proposed SF based approach for ischemic stroke risk estimation from B-mode
ultrasound images. An active contour model based intima-media (IM) segmentation approach
is being proposed that uses coordinates obtained from lines extracted using HT on ultrasound
images and deforms the contours formed by minimizing energy function to get finalized
contours.
Here, parametric snake model has been used as active contour model. Coordinates of the
segments of lines being detected using HT can adapt to the variation in curve and thickness
of the blood vessel as they are taken to be part of co-centric circles. These circles can be big
enough to consist of the points that form straight lines or small enough to cover points that
form curved vessels. The IMT values calculated are used to estimate stroke risk using an
objective criteria based on scientific research [99, 100].
Section 1 describes the data and details of our proposed approach. Section 2 contains the
results of our proposed approach followed by discussion on the results achieved in section 3.
We conclude this chapter in section 4 with a comparison between our approach and other
approaches and future prospects of the proposed approach.
3.1 Materials and Methods
We have used the CA B-mode ultrasound dataset available at eHealth Laboratory website,
University of Cyprus [101]. The dataset includes 100 B-mode longitudinal ultrasound images
40
of 100 patients who were at the risk of atherosclerosis and already had developed symptoms
like stroke or TIA.
This dataset was recorded at the Cyprus Institute of Neurology and Genetics, Nicosia,
Cyprus and used for segmentation of CA IM [33]. Images were stored on a magneto optical
drive after logarithmic compression. The recorded images were resized using bicubic method
to16.66 pixels/mm. The details about the dataset used are given in the Table 3.1.
Table 3.1: Dataset details.
Attributes Values
No. of Patients 100
Male 58
Female 42
Age Group 26–95 years
Mean Age 54 years
Image Size 768x576 pixels
Image Type Grayscale
Image Extension .cri
Ultrasound Mode B-mode
Plane Longitudinal Ultrasound
Image
Scanner Type ATL HDI-3000
No. of Elements 64
Head Operating Frequency Range 4-7 MHz
Acoustic Aperture 10 x 8 mm
Transmission Focal Range 0.8-11 cm
41
For ischemic stroke risk estimation from the carotid ultrasound images, the main phases are
Preprocessing, Intima-Media Segmentation and Estimation. Figure 3.2 illustrates these steps
and their outputs.
In the Preprocessing phase (Phase-I) of the proposed approach, Image Enhancement, ROI
Coordinates Extraction and Noise Removal is performed on Original Image (ultrasound
image) to get Enhanced Image, ROI Coordinates and Filtered Image respectively.
Intima-Media Segmentation phase (Phase-II) includes Line Extraction, Candidate Line
Selection and Contour Extraction. Line Extraction process is executed on the Enhanced
Image using ROI Coordinates obtained from the ROI Coordinates Extraction process to get
the Extracted Lines. Extracted Lines undergo a process of Candidate Line Selection to
determine whether they are the part of CA wall or not. The output of this process is Selected
Lines. Original Image using ROI Coordinates is filtered by Noise Removal process to get
Filtered Image.
The Selected Lines and Filtered Image are then further processed by the Contour Extraction
process to get final Contours of the Intima-Media. In Estimation Phase (Phase-III), Features
are acquired by calculating IMT values in Feature Extraction process and afterwards, these
Features are fed to the Classification process for Stroke Risk Estimation.
Figure 3.3 shows the detailed working of ischemic stroke risk estimation from ultrasound
images of the proposed approach.
42
Original ImageOriginal Image PreprocessingIntima-Media Segmentation
Estimation
ContoursStroke RiskEstimation
Stroke RiskEstimation
Filtered Image
ROI coordinates
Enhanced Image
Figure 3.2: Block diagram of proposed approach.
43
Intima-Media SegmentationPreprocessing
Estimation
OriginalImage
OriginalImage
Image Enhancement Enhanced
Image
ROI Coordinates Extraction
Line Extraction
Contour Extraction
Candidate Line Selection
Classification
Extracted Lines
SelectedLines
Contours
Stroke Risk Estimation
Stroke Risk Estimation
Feature Extraction
Features
Noise Removal
Filtered Image
ROI Coordinates
Figure 3.3: Detailed working of proposed approach.
44
3.1.1 Phase-I: Preprocessing
The Original Image is processed for automatic extraction of Enhanced Image, ROI
Coordinates and Filtered Image. Contrast in the Original Image is enhanced using
Adaptive Histogram Equalization to get Enhanced Image. In the Image Enhancement
process a mapping function is used that adjusts the intensity of pixels. Difference
between the values of the pixel under consideration with the local variance of the
neighboring pixels is calculated. If the difference is greater than a fixed threshold then
the intensity of the pixel is replaced with the global intensity of that pixel.
Enhanced Image is first binarized during the ROI Coordinates Extraction process.
The resultant binary image is then scanned starting from the last row exploiting the
prior information that first occurrence of maximum number of white pixels is found
in the row that is closer to the Media Layer. The coordinates of the starting point of
the identified row are used to calculate ROI Coordinates. ROI of height ‘w’ is
comprised of w rows above and w rows below the identified row. A value of 15
determined through experimental analysis is used for w in our proposed approach.
The ROI Coordinates are the coordinates of first pixels of the first and last row of
ROI i.e. Crow-w, col and Crow+w, col respectively.
A sub image is formed on the basis of the Original Image and ROI Coordinates and is
filtered for speckle noise reduction to produce Filtered Image. Detail Preserving
Anisotropic Diffusion (DPAD) filter proposed by [102] is used to filter the sub image
for speckle noise. Δt is set at 0.2 with 100 iterations. Estimation of noise statistics is
made using C2
MAD. The gradient is calculated on a 5x5 window.
The algorithm for the Preprocessing phase is given as Algorithm 1.
45
Algorithm 1: Preprocessing phase Algorithm
Function Preprocessing (OriginalImage) : Enhanced Image, ROI coordinates,
Filtered Image
{
AHEImage = Call AdaptiveHistogramEqualization for Image
[ Point x1, y1 , Point x2 , y2 ] = Call ROI for AHEImage
FLTRImage = Call NoiseRemoval for Image, [ Point x1, y1 , Point x2 , y2 ]
return AHEImage, [ Point x1, y1 , Point x2 , y2 ], FLTRImage
}
Function AdaptiveHistogramEqualization (Image) : AHEImage
{
EImage e = Call HistogramEqualization for Image
∀row Є Image c : i
∀ column Є Image c : j
sd = Call StandardDeviation for c (i-1:i+1,j-1:j+1)
if | c(i,j) – sd | > threshold //threshold = 10
then d(i,j) = e (i,j);
else d(i,j) = c(i,j)
return d
}
46
Function ROI (AHEImage) : ROI coordinates
{
BinaryImage= Call Binarize for AHEImage
C row,col = maxwhite (BinaryImage)
return Crow+w, col , Crow-w, col // w is ½ (ROI
height)
}
Function NoiseRemoval (Image, [ Point x1, y1 , Point x2 , y2 ] ) : Filtered Image
{
// Method DPAD for Image row Point x1, y1 to row Point x2 , y2 as in [102]
}
Function StandardDeviation (Intensities of 8 Neighbours) : Standard Deviation
{
// Default Method
}
Function HistogramEqualization (Image) : EImage
{
// Default Method
}
47
Function Binarize (AHEImage) : BinaryImage
{
∀row Є AHEImage a : i
∀ column Є AHEImage a : j
if a[i] [j] <= threshold
then b [i] [j] = 0
else b[i] [j] = 255
return b
}
3.1.2 Phase-II: Intima-Media Segmentation
A sub image is obtained from Enhanced Image using ROI Coordinates and is
processed for Line Extraction. LI and MA contain some piecewise straight
boundaries. HT is applied to the sub image to get Extracted Lines. These Extracted
Lines are then processed by Candidate Line Selection process as described in
Algorithm 2 to find the lines that are part of the LI and MA. Candidate Line Selection
is considered as a 2D problem given a set of n points Pi (xi , yi) (i = 1…n) comprising
of the starting and ending points of each of the Extracted Lines, find the center points
C1(xc1 , yc1), C2(xc2 , yc2) and the radius r1 and r2 of two unknown circles that
pass closest to the points. LI and MA each are considered as part of circles. Even
straight LI and MA can be considered to be part of very big circles. The two best fit
48
circles are found using Conjugate method with Polak and Ribiere Factor [103]
implemented in Matlab 7.14 and is given as Algorithm 2. The circles are selected
such that they are co-centered within a range and are are covering maximum number
of points. Moreover the circles selected do not have points inside illumination region,
since the adventitia region is brighter than the media region [104].
The Selected Lines are processed for Contour Extraction. The coordinates of the
Selected Lines are used as seed points for accurate contour extraction using snakes
model. The gradient vector flow (GVF) snake model [105] is used in the proposed
approach that processes and deforms the initial contour built using the seed points.
Algorithm 2: Candidate Line Selection Algorithm
Function CandidateLineSelection (Extracted Lines) : Selected Lines
{
Circles [C 1:n], Member Lines [L1:n]= Call ExtractCircles for Extracted Lines
Selected Lines [L1:n] = Call LineSelection for Circles [C1:n], Member Lines
return Selected Lines [L1:n]
}
Function ExtractCircles (Extracted Lines) : Circles, Member Coordinates
{
// Method for finding best fit circles using conjugate gradient with Polak and Ribiere
factor [103]
}
49
Function LineSelection (Circles, Member Lines) : Selected Lines
{
maxcoordinates = Call CoordinatesCountOrdering for Circles and Member Lines
∀ Ci 1:n Є Circles
∀ Cj 1:n Є Circles
x = Call IsCocentric for Ci and Cj
y = Call CoversMaxCoordinates for Ci and Cj and maxcoordinates
z = Call IlluminationRegionCheck for Ci and Cj
if (x && y && z)
Ca = Ci
Cb = Cj
Selected Lines = S1S2 : S1 ∈Member Lines Ca, S2 ∈ Member Lines Cb
break
return Selected Lines
}
Function CoordinatesCountOrdering (Circles, Member Lines) : maxcoordinates
{
// return maximum number of coordinates on a circle from Circles
}
50
Function IsCocentric (Ci , Cj) : BOOL
{
if |Center Ci – Center Cj| <= threshold
return TRUE
else
return FALSE
}
Function CoversMaxCoordinates (Ci, Cj, maxcoordinates) : BOOL
{
if ( Coordinates count Ci or Coordinates count Cj >= max coordinates / 2)
return TRUE
else
return FALSE
}
Function IlluminationRegionCheck ( Ci, Cj ) : BOOL
{
if ( Illumination value of Ci and Cj is in range [104])
return FALSE
else
return TRUE
51
}
For the proposed approach, GVF field calculation and initial contour deformation is
implemented in Matlab 7.14. Weightline component = 0.30, Weightedge component = 0.30,
Weightterminal component = 0.30 and number of iterations = 30 after experimentation.
The initial values of snake parameters were set to α = 0.40, β = 0.20, γ = 1, κ = 0.15.
Figure 3.4 illustrates the Image processing steps performed in Phase-I and Phase-II on
the Original Image to get the Contours.
52
Original Image
Image Enhancement
ROI Coordinates Extraction
Contours
Line Extraction
Candidate Line Selection Contour Extraction
Preprocessing
Intima-Media Segmentation
Noise RemovalROI Coordinates
Figure 3.4: Image processing steps on carotid ultrasound images.
53
3.1.3 Phase -III: Estimation
After getting the CA Contours, distance values between contours are calculated and
formation of Features is done by the Feature Extraction process. Pixels are counted
downwards from upper contour (LI) till the lower contour (MA) is reached. A
stepping function is used so that the distance is calculated in steps for pixels located
on the upper contour. A step value of 5 pixels is chosen.
The distance values are used for calculating the features from the carotid images.
Four types of features are being calculated and taken into account for analyzing the
thickness data. These features include minimum, maximum, mean and standard
deviation values of the line distance data calculated using Equations (3.1), (3.2), (3.3)
and (3.4) respectively.
𝑑𝑖𝑠𝑡𝑎𝑛𝑐𝑒𝑚𝑖𝑛𝑝𝑖𝑥𝑒𝑙𝑠= min ( 𝑑𝑖𝑠𝑡𝑎𝑛𝑐𝑒𝑝𝑖𝑥𝑒𝑙𝑠1
𝑛 ) (3.1)
𝑑𝑖𝑠𝑡𝑎𝑛𝑐𝑒𝑚𝑎𝑥𝑝𝑖𝑥𝑒𝑙𝑠= max ( 𝑑𝑖𝑠𝑡𝑎𝑛𝑐𝑒𝑝𝑖𝑥𝑒𝑙𝑠1
𝑛 ) (3.2)
𝑑𝑖𝑠𝑡𝑎𝑛𝑐𝑒𝑚𝑒𝑎𝑛𝑝𝑖𝑥𝑒𝑙𝑠=
∑ 𝑑𝑖𝑠𝑡𝑎𝑛𝑐𝑒𝑝𝑖𝑥𝑒𝑙𝑠𝑖𝑛𝑖=1
𝑛 (3.3)
𝑑𝑖𝑠𝑡𝑎𝑛𝑐𝑒𝑠𝑑𝑝𝑖𝑥𝑒𝑙𝑠= √
∑ (𝑑𝑖𝑠𝑡𝑎𝑛𝑐𝑒𝑝𝑖𝑥𝑒𝑙𝑠𝑖−𝑑𝑖𝑠𝑡𝑎𝑛𝑐𝑒𝑚𝑒𝑎𝑛𝑝𝑖𝑥𝑒𝑙𝑠
)2 𝑛𝑖=1
𝑛 (3.4)
where,
𝑑𝑖𝑠𝑡𝑎𝑛𝑐𝑒𝑝𝑖𝑥𝑒𝑙𝑠 is distance calculated in pixels
𝑑𝑖𝑠𝑡𝑎𝑛𝑐𝑒𝑚𝑖𝑛𝑝𝑖𝑥𝑒𝑙𝑠 is Minimum value for the distance in pixels
54
𝑑𝑖𝑠𝑡𝑎𝑛𝑐𝑒𝑚𝑎𝑥𝑝𝑖𝑥𝑒𝑙𝑠 is Maximum value for the distance in pixels
𝑑𝑖𝑠𝑡𝑎𝑛𝑐𝑒𝑚𝑒𝑎𝑛𝑝𝑖𝑥𝑒𝑙𝑠 is Mean value for the distance in pixels
𝑑𝑖𝑠𝑡𝑎𝑛𝑐𝑒𝑠𝑑𝑝𝑖𝑥𝑒𝑙𝑠 is Standard deviation for the distance in pixels
𝑛 is total number of distance values
The distance values calculated are mapped from pixels to millimeters (mms). General
Mapping function is given in Equation (3.5).
𝐼𝑀𝑇_𝑡ℎ𝑖𝑐𝑘𝑛𝑒𝑠𝑠𝑚𝑖𝑛𝑚𝑚,𝐼𝑀𝑇_𝑡ℎ𝑖𝑐𝑘𝑛𝑒𝑠𝑠𝑚𝑎𝑥𝑚𝑚
, 𝐼𝑀𝑇_𝑡ℎ𝑖𝑐𝑘𝑛𝑒𝑠𝑠𝑚𝑒𝑎𝑛𝑚𝑚and
𝐼𝑀𝑇_𝑡ℎ𝑖𝑐𝑘𝑛𝑒𝑠𝑠𝑠𝑑𝑚𝑚 are calculated using thickness mapping functions represented as
Equations (3.6), (3.7), (3.8) and (3.9) respectively.
Pixel density per mm is already known for the ultrasound images. In our case the
pixel density per mm is 16.66 pixels per mm. These thickness values form the
Features. The thickness mapping function is as below:
𝑚𝑎𝑝(𝑋𝑝𝑖𝑥𝑒𝑙𝑠) = 𝑋𝑝𝑖𝑥𝑒𝑙𝑠
𝑝𝑖𝑥𝑒𝑙_𝑑𝑒𝑛𝑠𝑖𝑡𝑦= 𝑋𝑚𝑚 (3.5)
𝐼𝑀𝑇_𝑡ℎ𝑖𝑐𝑘𝑛𝑒𝑠𝑠𝑚𝑖𝑛𝑚𝑚=
𝑑𝑖𝑠𝑡𝑎𝑛𝑐𝑒𝑚𝑖𝑛𝑝𝑖𝑥𝑒𝑙𝑠
𝑝𝑖𝑥𝑒𝑙_𝑑𝑒𝑛𝑠𝑖𝑡𝑦 (3.6)
𝐼𝑀𝑇_𝑡ℎ𝑖𝑐𝑘𝑛𝑒𝑠𝑠𝑚𝑎𝑥𝑚𝑚=
𝑑𝑖𝑠𝑡𝑎𝑛𝑐𝑒𝑚𝑎𝑥𝑝𝑖𝑥𝑒𝑙𝑠
𝑝𝑖𝑥𝑒𝑙_𝑑𝑒𝑛𝑠𝑖𝑡𝑦 (3.7)
𝐼𝑀𝑇_𝑡ℎ𝑖𝑐𝑘𝑛𝑒𝑠𝑠𝑚𝑒𝑎𝑛𝑚𝑚=
𝑑𝑖𝑠𝑡𝑎𝑛𝑐𝑒𝑚𝑒𝑎𝑛𝑝𝑖𝑥𝑒𝑙𝑠
𝑝𝑖𝑥𝑒𝑙_𝑑𝑒𝑛𝑠𝑖𝑡𝑦 (3.8)
𝐼𝑀𝑇_𝑡ℎ𝑖𝑐𝑘𝑛𝑒𝑠𝑠𝑠𝑑𝑚𝑚=
𝑑𝑖𝑠𝑡𝑎𝑛𝑐𝑒𝑠𝑑𝑝𝑖𝑥𝑒𝑙𝑠
𝑝𝑖𝑥𝑒𝑙_𝑑𝑒𝑛𝑠𝑖𝑡𝑦 (3.9)
where,
55
𝑋𝑝𝑖𝑥𝑒𝑙𝑠 is any X value in pixels
𝑋𝑚𝑚 is mapped X value in mm
𝑝𝑖𝑥𝑒𝑙_𝑑𝑒𝑛𝑠𝑖𝑡𝑦 is pixel density per mm
𝑑𝑖𝑠𝑡𝑎𝑛𝑐𝑒𝑚𝑖𝑛𝑝𝑖𝑥𝑒𝑙𝑠 is Minimum value for the distance in pixels
𝑑𝑖𝑠𝑡𝑎𝑛𝑐𝑒𝑚𝑎𝑥𝑝𝑖𝑥𝑒𝑙𝑠 is Maximum value for the distance in pixels
𝑑𝑖𝑠𝑡𝑎𝑛𝑐𝑒𝑚𝑒𝑎𝑛𝑝𝑖𝑥𝑒𝑙𝑠 is Mean value for the distance in pixels
𝑑𝑖𝑠𝑡𝑎𝑛𝑐𝑒𝑠𝑑𝑝𝑖𝑥𝑒𝑙𝑠 is Standard Deviation for the distance in pixels
𝐼𝑀𝑇_𝑡ℎ𝑖𝑐𝑘𝑛𝑒𝑠𝑠𝑚𝑖𝑛𝑚𝑚 is mapped Minimum value for IMT thickness in mm
𝐼𝑀𝑇_𝑡ℎ𝑖𝑐𝑘𝑛𝑒𝑠𝑠𝑚𝑎𝑥𝑚𝑚 is mapped Maximum value for IMT thickness in mm
𝐼𝑀𝑇_𝑡ℎ𝑖𝑐𝑘𝑛𝑒𝑠𝑠𝑚𝑒𝑎𝑛𝑚𝑚 is mapped Mean value for IMT thickness in mm
𝐼𝑀𝑇_𝑡ℎ𝑖𝑐𝑘𝑛𝑒𝑠𝑠𝑠𝑑𝑚𝑚 is mapped Standard Deviation value for IMT thickness in mm
𝐼𝑀𝑇_𝑡ℎ𝑖𝑐𝑘𝑛𝑒𝑠𝑠𝑚𝑖𝑛𝑚𝑚, IMT_thicknessmaxmm
, IMT_thicknessmeanmmcorrespond to
𝐼𝑀𝑇𝑚𝑖𝑛, 𝐼𝑀𝑇𝑚𝑎𝑥 and 𝐼𝑀𝑇𝑚𝑒𝑎𝑛values.
In the Classification process the subjects are classified into four major categories i.e.
No Stenosis, Mild Stenosis, Moderate Stenosis and Severe Stenosis based on well-
established criteria [99, 100]. This classification is further associated with the % risk
of ischemic stroke episode. The % risk is in terms of no risk, 25%, 50%, 75%
increased risk.
56
The IMT_thicknesssdmm is a marker of thickness irregularities. It is used to find out
the % risk for an individual examined for the risk of ischemic stroke episode. The
IMT_thicknessmaxmm is used for the major classification, further classification is
done on the basis of IMT_thicknesssdmm. The proposed decision tree for the
classification is given in Figure 3.5.
57
≥ 0.88 mm & ≤ 0.99 mm
Classified as No Stenosis
No Risk
Classified as Mild Stenosis
Classified as Mild
Stenosis with 25% Increased
Risk
Classified as Moderate Stenosis
Classified as Moderate Stenosis with 25% Increased
Risk
Classified as Moderate
Stenosis with 50%
Increased Risk
Classified as Severe
Stenosis with 75% Increased
Risk
<0.88 mm > 0.99 mm & ≤ 1.12 mm
< 0.1 mm≥ 0.1 mm & ≤ 0.89 mm
≥ 0.1 mm & ≤ 0.89 mm
> 0.9 mm< 0.1 mm
> 1.12 mm
Figure 3.5: Decision tree for classification of stenosis and ischemic stroke risk.
58
3.2 Results
To validate the results of the proposed approach, the Features extracted after Intima-Media
Segmentation and Stroke Risk Estimation results are investigated. The automatic
segmentation results are investigated that how much the automatic results differ from the
manual results. The evaluation metrics used are Intra-Observer Error (IOE) [29, 33],
coefficient of variation (CV) % [29, 33], wilcoxon matched pairs rank sum test[33], bland-
altman plots [33, 106] and figure of merit (FoM) %[34].
𝐼𝑀𝑇𝑚𝑖𝑛,𝐼𝑀𝑇maxand 𝐼𝑀𝑇mean have been obtained both for the proposed approach using the
equations (3.6), (3.7) and (3.8) and for manual measurements by two experts for each one of
100 images. Intra-observer error is calculated using the formula IOE = SDIMT / √2 where
SDIMT is the standard deviation for each 100 measurements. Difference as percentage of the
collective mean value is calculated using CV% = IOE / IMT̅̅ ̅̅ ̅ * 100 where IMT̅̅ ̅̅ ̅ is the mean
IMT for each 100 measurements.
Each set of measurement is checked whether a significant difference exists or not between all
segmented boundaries by using the Wilcoxon pairs rank sum test at p<0.05. Agreement
between automatic and manual results is evaluated using Bland-Altman plots with 95%
agreement.
The measurements for all 100 images are given as manual measurements (NM1, NM2) by the
experts and automatic measurements (NA) generated by the proposed approach. The
observed standard deviation, SDIMT for the 𝐼𝑀𝑇𝑚𝑒𝑎𝑛 values of manual for expert1NM1 is
0.67mm, expert2 NM2 is 0.65mm and for automatic measurements by proposed approach NA
is 0.68mm. The results computed for the 100 images are given in Table 3.2 for 𝐼𝑀𝑇𝑚𝑒𝑎𝑛,
𝐼𝑀𝑇𝑚𝑖𝑛, 𝐼𝑀𝑇𝑚𝑎𝑥for NM1, NM2 and NA along with the values for IOE and CV%. The
𝐼𝑀𝑇𝑚𝑒𝑎𝑛± SDIMT for NM1 is 0.67± 0.15, NM2 is 0.65±0.16 and 0.68±0.15 for NA.
59
Table 3.2: Comparison of manual and proposed approach measurements for 100 carotid artery ultrasound images
Metrics
Manual measurements (mm) Automatic measurements by
proposed approach NA (mm) Expert 1 NM1 Expert 2 NM2
IMT mean (SDIMT) 0.67 (0.15) 0.65 (0.16) 0.68 (0.15)
IMT min (SDIMT) 0.53 (0.13) 0.57 (0.15) 0.51 (0.11)
IMT max (SDIMT) 0.82 (0.20) 0.74 (0.16) 0.86 (0.16)
IOE 0.105 0.109 0.088
CV % 15.72 16.92 12.99
IOE Intra-observer error, CV% coefficient of variation.
60
The results of Wilcoxon rank sum test for both of the manual segmentation measurements
and automatic measurement is given in Table 3.3. It is observed that a significant difference
does not exist between the automatic and manual measurements suggesting that manual
measurements can be replaced by automatic measurements with confidence.
Table 3.3: Wilcoxon ranksum test computed for Expert 1, Expert 2 and Automatic measurements for 100 carotid artery ultrasound images
Wilcoxon ranksum test
Expert 1 Expert 2
Automatic Not Significant (0.3538) Not Significant (0.0532)
Expert 1 - Not Significant (0.2365)
p value is shown in parentheses (Significant difference at p < 0.05, Not significant
difference at p > 0.05).
Bland-Altman plots between the manual measurements by Expert1NM1, Expert2 NM2 and
Automatic measurements by proposed approach NA are given in Figure 3.6. Mean difference
is represented by the middle line. Upper and lower lines represent the limits of agreement
between the manual and automatic measurements i.e. mean of the measurements ± 2SD. The
difference of measurements of the proposed approach and expert 1 is 0.012 + 0.17 and 0.012
- 0.14 (Figure 3.6 (a)), and for the expert 2 is 0.034 + 0.31 and 0.0.34 – 0.24 (Figure 3.6 (b)).
FoM% is calculated for the proposed approach using equation 3.10.
FoM = 100 − |IMT̅̅ ̅̅ ̅̅ 𝐴𝑢𝑡𝑜𝑚𝑎𝑡𝑖𝑐− IMT̅̅ ̅̅ ̅̅ 𝑀𝑎𝑛𝑢𝑎𝑙
IMT̅̅ ̅̅ ̅̅ 𝑀𝑎𝑛𝑢𝑎𝑙| ∗ 100 (3.10)
where,
61
IMT̅̅ ̅̅�̅�𝑢𝑛𝑡𝑜𝑚𝑎𝑡𝑖𝑐is the average IMT value calculated for the 100 measurements done by the
proposed approach
IMT̅̅ ̅̅ ̅𝑀𝑎𝑛𝑢𝑎𝑙 is the mean IMT for the 100 measurements done manually by an expert
The FoM% is 98.5 for the proposed approach w.r.t expert 1 and 95.4 for the proposed
approach w.r.t expert 2.
Patients are classified for stroke risk by the proposed classification decision tree using the
automatic IMT values calculated by the proposed approach. The manual measurements by
the experts are evaluated by expert physician for stroke risk estimation. The risk estimation
results of both the manual measurements evaluated by a physician and the automatic
measurements calculated using the proposed approach and evaluated by the proposed
decision tree are compared.
The major classes for stroke risk estimation i.e. no stenosis, mild stenosis, moderate stenosis
and severe stenosis, are assigned values 1, 2, 3 and 4 respectively. The difference of
classification results is measured by using equation 3.11.
Df = |𝑀𝑎𝑛𝑢𝑎𝑙𝐶𝑙𝑎𝑠𝑠𝑉𝑎𝑙𝑢𝑒 − 𝐴𝑢𝑡𝑜𝑚𝑎𝑡𝑖𝑐𝐶𝑙𝑎𝑠𝑠𝑉𝑎𝑙𝑢𝑒| (3.11)
where,
ManualClassValue is the value of the class for manual measurement for an image
AutomaticClassValue is the value of the class for the automatic measurement for an image
62
(a) (b)
Figure 3.6: Bland-Altman plots of (a) Expert1 NM1 versus Automatic measurements by proposed approach NA (b) Expert2 NM2 versus Automatic
measurements by proposed approach NA.
63
The difference values were 0 where the classification results were same for the automatic
values and those by an expert physician for manual measurements. For the cases where the
results were not the same, the difference values were 1, 2 or 3. The results of these
differences are given in Table 3.4.
Table 3.4: Difference computed for classification results of Expert 1, Expert 2 and Automatic measurements for 100 carotid artery ultrasound images
Classification
Accuracy
Classification Difference (Df)
1 2 3
Automatic vs. Expert1 76 % 19% 5% 0
Automatic vs. Expert 2 68% 24% 8% 0
Expert 1 vs. Expert 2 69% 20% 6% 5%
3.3 Discussion
CIMT is considered to be a marker for early diagnosis and risk estimation of atherosclerosis.
Ultrasound images are generally used to measure CIMT. We have developed a fully
automatic image processing based approach to measure the CIMT. The proposed approach
can predict ischemic stroke risk based on IMT values and variation in the IMT values of
same individual. An IM segmentation approach is being proposed that uses an improved
snakes initialization method for the GVF snakes. The coordinates of the selected lines
extracted by HT are used for automatic initialization of the snakes.
The proposed approach is more reproducible as the CV% (12.99%) and IOE (0.08) of the
proposed approach are both smaller than that of the manual measurements by experts (see
Table 3.2). The manual measurements by expert 1 and expert 2 (0.67mm and 0.65mm
64
respectively) were smaller than the measurements by the proposed approach (0.68mm).
Similar results have been reported in other studies [29, 33]. Intra-observer variability is an
effective metric to measure the performance of a computer aided measurement approach. It
can be seen in Table 3.2 that the intra-observer variability calculated as IOE for both experts
is higher as compared to the IOE for the proposed approach.
Wilcoxon ranksum test results are given in Table 3.3. It is observed that there is no
significant difference between the automatic measurements made by the proposed approach
and the manual measurements made by experts. Bland-Altman plots given as Figure 3.6
show that almost all of the data points lie within the 2σ of the mean range. The difference
between the proposed approach and measurements by expert 1 is 0.012 mm and is 0.034mm
for the proposed approach and measurements by expert 2.
The dataset used has images that were recorded using a standard recording technique to
adjust the position of the probe such that the ultrasound beam is at right angle to the arterial
wall [33]. This improves the Intima-Media visualization.
Correct segmentation depends on the estimation and positioning of the initial snake contour.
All the images are correctly segmented by the proposed approach giving a segmentation
accuracy of 100%.
The classification results by an expert physician for both of the manual measurements and by
proposed decision tree for automatic measurements are compared. The classification results
of the proposed approach are found to be more close to that of the expert 1 than expert 2.
Similar findings are also produced by the Bland-Altman plots (See Figure 3.6). The reason
for these results is that expert 2 tends to give smaller values for the IMT measurements. The
classification accuracy for the proposed approach when compared with the diagnosis by the
physician for expert 1 measurements is 76% and is 68% when compared with expert 2. The
classification accuracy for the comparison between the two manual measurements is 69%.
65
The classification difference (Df) of 1 is observed for 19% of the images for the automatic
vs. expert1 measurements, for 24% of the images for the automatic vs. expert2 measurements
and for 20% of the images for the expert 1 vs. expert 2 measurements. Classification
difference (Df) of 2 is observed for 5% of the images for the automatic vs. expert1
measurements, for 8% of the images for the automatic vs. expert2 measurements and for 6%
of the images for the expert 1 vs. expert 2 measurements. Classification difference (Df) of 3
is observed for 0% of the images for the automatic vs. expert1 measurements, for 0% of the
images for the automatic vs. expert2 measurements and for 5% of the images for the expert 1
vs. expert 2 measurements.
The Classification difference (Df) of 3 means that the diagnosis is completely incorrect. For
example, if a individual has severe stenosis then he is diagnosed as having no stenosis and
estimation of stroke risk is ‘risk free’. Or an individual has no stenosis and is diagnosed as
having severe stenosis and estimation of stroke risk as increased to 75 %. The results of
proposed approach when compared with the diagnosis by expert physician for either of the
manual measurements did not produce any incorrect diagnosis. On the contrary the diagnosis
for both of the manual measurements when compared with each other had such incorrect
results for 5% of the cases.
3.4 Conclusion
In this research we have proposed an improved approach for Intima-Media segmentation
with improved snake initialization process. A classification scheme is also proposed to
associate the stenosis with ischemic stroke risk estimation. The proposed approach extracts
the contours in the ultrasound images using gradient vector flow snakes with an improved
snake initialization process. The seed points for this improved snake initialization process are
extracted using selected edges returned by the candidate line selection algorithm.
IMT is calculated from the extracted contours. SF are calculated using the IMT values and
66
are included in the feature set for risk estimation analysis. The proposed approach is tested
and clinically validated on a data set of 100 longitudinal ultrasound images of the CA. The
IOE of 0.088, a CV of 12.99%, Bland-Altman plots with small differences between experts
(0.01 and 0.03 for Expert 1 and Expert 2, respectively) and FoM of 98.5% are obtained.
We found no significant difference between the IMT measurements through proposed
approach and the manual measurements which shows the accuracy of our approach. Based
on the Bland-Altman test, CV% and FoM it can be observed that the proposed approach
measurements are interchangeable to manual measurements.
The proposed approach can be successfully used for measurement of IMT, complementing
the manual IMT measurements. The IMT values are then further used for a individual’s risk
estimation of stroke. The risk estimation for the measurements by proposed approach and
measurements taken manually are also found to be similar.
The proposed approach is better than the existing approaches in terms of FoM % and
variation (see Table 3.5). Variation is the difference of the manual and system means.
FoM = |IMT̅̅ ̅̅�̅�𝑢𝑡𝑜𝑚𝑎𝑡𝑖𝑐 − IMT̅̅ ̅̅ ̅
𝑀𝑎𝑛𝑢𝑎𝑙| (3.12)
where,
IMT̅̅ ̅̅�̅�𝑢𝑛𝑡𝑜𝑚𝑎𝑡𝑖𝑐is the average IMT value calculated for the 100 measurements done by the
proposed approach
IMT̅̅ ̅̅ ̅𝑀𝑎𝑛𝑢𝑎𝑙 is the mean IMT for the 100 measurements done manually by an expert
67
Table 3.5: Comparison of proposed approach with existing approaches
Approach Method Mode Dataset
Size
IMT Values (Mean ± SD) mm
Manual System |Variation| FoM %
Delsanto et al.,
2007 [22]
Snakes, Fuzzy C-
means method
Automatic 200 0.77± 0.22 0.71± 0.16 0.06 92.2
Loizou et al.,
2007 [33]
Snakes based
SF
Semi-
Automatic
100 0.65± 0.18 0.68±0.12 0.03 95.4
Faita et al., 2008
[13]
FOAM edge operator Automatic 150 0.56± 0.14 0.57± 0.14 0.01 98.2
Loizou et al.,
2009 [107]
Snakes Semi-
Automatic
100 0.71± 0.17 0.67± 0.12 0.04 94.4
68
Approach Method Mode Dataset
Size
IMT Values (Mean ± SD) mm
Manual System |Variation| FoM %
Molinari et al.,
2010 [24]
Integrated approach Automatic 182 0.92± 0.30 0.75± 0.39 0.17 81.5
Molinari et al.,
2010 [47]
Integrated approach Automatic 200 0.92± 0.30 0.75± 0.39 0.17 81.5
Molinari et al.,
2011 [108]
Integrated approach,
FOAM
Automatic 295 0.782 ± 0.281 0.750 ± 0.203 0.03 95.9
Meiburger et al.,
2011 [38]
Edge flow Automatic 300 0.818 ± 0.246 0.861 ± 0.276 0.04 94.7
Molinari et al.,
2012 [49]
FOAM Automatic 365 0.95 ± 0.39 0.91± 0.44 0.04 95.8
69
Approach Method Mode Dataset
Size
IMT Values (Mean ± SD) mm
Manual System |Variation| FoM %
Molinari et al.,
2012 [34]
FOAM,
Dual snakes
Automatic 665 0.836 ± 0.296 0.823± 0.239 0.01 98.4
Xu et al., 2012
[35]
HT, Dual snakes Semi-
Automatic
50 0.63± 0.14 0.65± 0.16 0.02 96.8
Menchón-Lara et
al., 2014 [109]
Morphological
operations, ANN
Automatic 60 0.64± 0.19 0.61±0.19 0.03 95.3
Proposed Integrated approach Automatic 100 0.672 ± 0.149 0.683 ± 0.148 0.01 98.5
70
The presented research may be further explored for estimation on ultrasound images of the
CA on large scale. Moreover, it may take into account the plaque texture properties,
phenotype data e.g. age, gender, stroke history, smoking, Body Mass Index (BMI), etc. and
genetic data for developing a substantially improved and extensive criterion for ischemic
stroke risk estimation.
CHAPTER NO. 4
PROPOSED APPROACH-
GENETIC DATA BASED
ISCHEMIC STROKE
CLASSIFICATION
71
Proposed Approach- Genetic Data Based Ischemic
Stroke Classification
In the last decade the genetic basis underlying the human diseases has been investigated
intensively. Different Genome-Wide Association Studies (GWAS) have been conducted to
find out the genetic reasons for different human diseases [110-113]. Such a study consists of
analyzing genotype data from affected (cases) and healthy (controls) individuals to identify
the Single Nucleotide Polymorphisms (SNPs) having a significant difference in frequencies
for the two groups. Thousands of SNPs associated with diseases have been identified [114-
117]. SNPs have been identified for diabetes [118-121], cancer [122-125], alzheimer’s [126-
128], autism [129, 130] etc.
Ischemic stroke is a common neurological multifactorial disorder. There are many risk
factors for ischemic stroke. Some of these factors can be changed and some cannot be
changed.
Changeable factors are the risk factors that can be treated, controlled or changed. Such
factors include high blood pressure, smoking, diabetes, artery diseases, atrial fibrillation,
sickle cell disease/ anemia, high blood cholesterol, physical inactivity and obesity.
On the contrary unchangeable risk factors are the one that cannot be changed, modified or
cannot be treated. They include age, genetic predisposition, race, gender, prior stroke, TIAs
and heart attack.
Other factors related to ischemic stroke risk are geographic location, drug and alcohol usage.
Ischemic stroke has a research based indication of genetic influence [112, 131-133]. Almost
half of the ischemic stroke cases are suspected to be genetic as the patients do not suffer from
the conventional risk factors.
Extensive research has been conducted to investigate the unknown reasons and their
72
relationship with genetics. Different studies have suggested an association of genetic factors
among young individuals. Similarly family history of stroke as an increase in risk for stroke
is also being investigated. It is found that family history is a major risk factor both in young
individuals and adults.
Living organisms are made up of cells. The instructions for growth and functionality are
stored in the nucleus of the cell. Proteins are produced inside a cell that enables a cell to
perform special functions. There are 100 trillion cells in a human body.
DeoxyriboNucleic Acid (DNA) is the genetic material of a cell in all living things. DNA is a
molecule having the instructions, also known as the blueprint of a human being. Detailed set
of instructions are encoded in DNA. There are two main types of DNA namely;
1. Genomic DNA
2. Mitochondrial DNA
Genomic DNA/ Nuclear DNA comprise of the whole genome of an organism. Nuclear DNA
undergoes recombination. Mitochondrial DNA is the present in the mitochondria. It is always
maternally inherited and remains unchanged from parent to child. It is used to investigate
maternal hereditary diseases.
The DNA is in the shape of a twisted ladder known as “double helix” with the ladder rungs
made of base pairs bonded together with hydrogen. Each of the bases can be either of the
four letters A (Adenine), C (Cytosine), T (Thymine) and G (Guanine). There is a special rule
for pairing of the bases i.e. A always pairs with T and C with G.
Sequence of the base letters is known as DNA strand. e.g.
ATGCTCGAATAAATGTCAATTTGA. These letters combine together to make words. For
example, ATG CTC GAA TAA ATG TCA ATT TGA. Words combine to make
sentences. These sentences are known as genes.
Genes instruct the cell to produce molecules known as proteins. These proteins are
73
responsible for enabling a cell to collectively work together with other cells to perform
special functions for example heart cells work together to make the heart pump blood to all
parts of body. There are almost 25000 genes in human body. Chromosomes are compact
units of DNA. DNA is packaged in chromosomes. Each chromosome has two DNA chains.
There are 46 chromosomes (2 sets of 23 chromosomes) in humans. One set is inherited from
mother and other from the father. A specific position in chromosome is known as marker or
locus. There are two types of markers which include SNPs and macrosatellite markers. SNPs
are short DNA sequences that surround a single base-pair change. Macrosatellite markers are
long DNA sequences that surround base pair changes.
If the instructions in a gene (DNA sequence) are changed then the gene is known as mutated
gene. Mutated genes are responsible for malfunctioning of cells and for different disorders.
During the process of copying DNA just before the division of cell an error occurs in almost
every 100,000 nucleotides. This error can be when a base is substituted by another base. The
error can also be caused by deletion or addition of a base. These changes are mostly being
repaired by the cell itself. If the change in the DNA is not repaired then it is passed on to the
child through the parents containing the changed cell.
Heredity is the passing of traits from parents to children. Genes encode instructions to define
our traits that are notable features or qualities in a individual. Every human has different
combinations of traits that make him/her unique. These qualities and features are passed from
generation to generation. One generation inherits these traits from previous generation and
passes onto the next generation. Environmental conditions affect the traits and can partially
or completely change them. There are several types of traits.
1. Physical Traits
2. Behavioral Traits
Physical traits form one’s appearance. Skin color, eyes color, height, hair type and color are
74
some of the examples of physical traits. Environmental factors like exposure to sunlight or
certain chemicals etc. can change these. Behavioral traits form one’s personality and
disposition. Examples include guard dog’s instinct to guard houses or herds of cattle. But we
can train a watch dog to be a play dog. This implies that environmental conditions can
change behavioral traits. Predisposition to certain disease traits increases a individual’s risk
of certain genetically transmissible diseases. Hereditary diseases include heart disease, sickle
cell anaemia, cancer, mental disorders, etc. Disease risk can be reduced by preventive
measures and healthy lifestyle.
Allele is the genetic information for each trait. These are DNA sequences within a marker or
locus. There are two alleles for each trait. Each one of the allele is a copy from mother and
father. These copies can be identical or different from each other. Individuals who have two
of the same alleles for a particular trait are known as homozygous. The identical alleles can
be used to predict an individual’s traits. Heterozygous defines the individual who have two
different alleles for a trait. One of the alleles is masked by the other allele. The masked allele
is called “recessive” and the one that masks it is “dominant”.
Well-defined physical traits can be easily traced through generations. For such traits the
alleles are known. These include eye color, skin color, thumb extension etc. Incomplete
dominance traits are not easy to trace through generations. In such a case the alleles interact
together to produce a particular trait. Single-gene traits are influenced by a single gene. Traits
that are formulated by more than one gene are known as complex traits.
Research has proved that genetic mutations can cause more than 4,000 diseases.
Environmental factors as well as multi-gene variations also play a major role in disease risk
elevation. Usually a human cell has 5 to 10 mutated genes. But having a mutated gene
doesn’t always indicate that an individual will develop a certain disease. The problem occurs
when either the diseased gene is dominant or when both copies of the recessive gene are
75
mutated. Individuals who have either of the two conditions are diseased.
Individuals with mutation in the recessive gene are called carriers. When only the recessive
gene is mutated then the normal copy of the gene takes over the functionality. If both the
parents are carriers then there is 25% chance of the child developing the disease. One of the
variations is X-linked diseases. In such a case if there is a mutation on recessive gene of X
chromosome then only males might develop a genetic disease but females might not as they
have two copies of X chromosome.
Another variation can be getting extra or missing chromosomes in the cell during the cell
division in the reproduction phase. This can also cause genetic diseases. The two alleles
combinely at a specific locus on each pair of the chromosomes is known as genotype.
Haplotypes are sequence of multiple alleles on a chromosome.
The International Haplotype Map Project (HapMap) along with genotyping methods has
provided opportunities for association studies. Different linkage and candidate association
studies have been done to identify the candidate genes and mutations that are prospective risk
for ischemic stroke. There are many candidate genes significantly associated with ischemic
stroke [131, 134, 135] but quite a few have been replicated.
Different SNPs and one haplogroup that have been reported in literature to be associated with
ischemic stroke risk are shown in Figure 4.1. Table 4.1 shows the genes and their SNP id as
well as the respective SNPs that are identified by meta-analysis of genes for stroke risk
estimation.
Multiple researches have been conducted to associate SNPs with disease risk in individuals
using their SNP profiles [136-138]. The disease risk is evaluated using SNP profile and then
compared with the actual status of the individual (case/control).
76
Genes
MTHFR
F5
PRKCH
PROC
CYP11B2
APOE e4
Near NINJ2
9p21.3
rs1801133
rs1801131
rs12121543
rs13306553
rs9651118
rs1801133
rs2274976
rs1801131
rs6025
rs2230500
rs2246700
rs3783799
rs12587610
rs3825655
rs1401296
rs1799998
rs7412
rs429358
rs1333049
rs12425791
rs11833579
Haplogroup
Figure 4.1: Identified SNPs causing stroke risk
77
Table 4.1: Stroke associated genes/ locus, SNP id/ haplotype and corresponding SNPs
Gene
Symbol/
Locus
Gene Name Chromosome
Cytogenetic
Location
SNP id/
Haplotype
SNP
MTHFR
[133, 139-
142]
Methylenetetrahydrofolate
reductase (NAD(P)H)
1 1p36.3
rs1801133
TTGAAGGAGAAGGTGTCTGCG
GGAG[C/T]CGATTTCATCATCAC
GCAGCTTTTC
rs1801131
TGGGGGGAGGAGCTGACCAGT
GAAG[A/C]AAGTGTCTTTGAAG
TCTTCGTTCTT
rs12121543
GCCACCACATGCCCAGGAGGCC
ATT[A/C]CTGTAAATTCTGCCCC
TGACTCCTC
78
Gene
Symbol/
Locus
Gene Name Chromosome
Cytogenetic
Location
SNP id/
Haplotype
SNP
rs13306553
rs9651118
rs1801133
ACATGCAGAGGTGAACTGCACC
ATG[C/T]CCTTGCTCCTTTTGTAT
CACCCACT
ACTTTTCACAGCGCTTGCCTGT
TTA[C/T]TATCTCAGGTGAGTTA
AGACATCAT
TTGAAGGAGAAGGTGTCTGCG
GGAG[C/T]CGATTTCATCATCAC
GCAGCTTTTC
GAGGCCTTTGCCCTGTGGATTG
79
Gene
Symbol/
Locus
Gene Name Chromosome
Cytogenetic
Location
SNP id/
Haplotype
SNP
rs2274976
rs1801131
AGC[A/G]GTGGGGAAAGCTGTA
TGAGGAGGAG
TGGGGGGAGGAGCTGACCAGT
GAAG[A/C]AAGTGTCTTTGAAG
TCTTCGTTCTT
F5 [143-
145]
Coagulation factor V
(proaccelerin, labile
factor)
1 1q23 rs6025
TGTAAGAGCAGATCCCTGGACA
GGC[A/G]AGGAATACAGGTATTT
TGTCCTTGA
80
Gene
Symbol/
Locus
Gene Name Chromosome
Cytogenetic
Location
SNP id/
Haplotype
SNP
PRKCH
[112, 146,
147]
Protein kinase C, eta. 14 14q23.1
rs2230500
TTTGCCATAGGTGATGCTTGCA
AGA[A/G]TAAAAGAAACAGGAG
ACCTCTATGC
rs2246700
GATCCTAAATGGGGAAAAGGCA
TTT[A/T]ATGGCTCTAGAGAGGG
TCCTGGGGA
rs3783799
GCCTGGGGACAATGAAGGATCT
GAG[A/G]CGTTATCAGCTGGAAT
AAATTCTGA
rs12587610 CATATTATATATGGTGGTTAAGAT
81
Gene
Symbol/
Locus
Gene Name Chromosome
Cytogenetic
Location
SNP id/
Haplotype
SNP
T[A/G]GGGCTCTGGAATCAGATT
TGGATTT
rs3825655
TGCTGATGGGAGGTGAAACTGA
AGC[A/G]ACAAGCAACACATTC
CTGTTTTATG
PROC
[148, 149]
Protein C (inactivator of
coagulation factors Va and
VIIIa)
2 2q12-q14 rs1401296
AACCGCGCCCGGGGCTGGAAG
CACC[C/T]GCCGAATGGCACAG
GGCCAGTGCCC
CYP11B2
[133, 150,
151]
Cytochrome P450,
family11, subfamily B,
polypeptide 2
8 8q21-q22 rs1799998
AAAGTCTATTAAAAGAATCCAA
GGC[C/T]CCCTCTCATCTCACGA
TAAGATAAA
82
Gene
Symbol/
Locus
Gene Name Chromosome
Cytogenetic
Location
SNP id/
Haplotype
SNP
APOE- e4
[152-154]
Apolipoprotein E version
e4
19 19q13.2
rs7412
CCGCGATGCCGATGACCTGCAG
AAG[C/T]GCCTGGCAGTGTACC
AGGCCGGGGC
rs429358
GCTGGGCGCGGACATGGAGGA
CGTG[C/T]GCGGCCGCCTGGTG
CAGTACCGCGG
Near NINJ2
[155, 156]
- 12 12p13 rs12425791
CCTGGTAAAAAGATTTTGTGCC
AAC[A/G]GTTCTTGGTTTCTCCT
CTGACAACC
83
Gene
Symbol/
Locus
Gene Name Chromosome
Cytogenetic
Location
SNP id/
Haplotype
SNP
rs11833579
CTTTCTGGAAAACCTTATTTCGG
AT[A/G]CCAGAAGCAAAATATTA
ACTATTTA
9p21.3
[157-159]
- 9 9p21.3 rs1333049
CATACTAACCATATGATCAACAG
TT[C/G]AAAAGCAGCCACTCGC
AGAGGTAAG
84
4.1 Materials and Methods
4.1.1 Data
We have compiled data from various databases and studies on ischemic stroke [160-163].
Our compiled data includes DNA data for 249 patients with stroke and 268 controls. The
details about the dataset are given in Table 4.2.
Table 4.2: Details of dataset
Attributes Controls Cases
No. of Subjects 268 249
Male 128 134
Female 140 115
Mean Age (years) 69.5 71.8
4.1.2 Method
Genotype data for all subjects for 15 SNPs and 1 haplotype (listed in Table 4.1) is organized
for analysis. The alleles for selected SNPs and haplotype for all subjects is considered. Table
4.3 shows the genetic data for eight randomly chosen subjects as an example. The attribute
‘class’ in the Table 4.3 shows whether the individual is a normal control or an ischemic
stroke patient.
85
Table 4.3: Sample genetic data for randomly chosen subjects
SNP id/
Haplotype
Allele
Subject 1 Subject 2 Subject 3 Subject 4 Subject 5 Subject 6 Subject 7 Subject 8
rs1801133 CT CC CT CT CT CT CC CC
rs1801131 AA AA AC AC AA AC AA AA
rs12121543
rs13306553
rs9651118
rs1801133
rs2274976
rs1801131
CC CC AC AC CC AC CC CC
TT TT TT CT TT CT TT TT
CT CT TT TT CT TT CC CC
CT CC CT CT CT CT CC CC
GG GG GG AG GG AG GG GG
AA AA AC AC AA AC AA AA
86
SNP id/
Haplotype
Allele
Subject 1 Subject 2 Subject 3 Subject 4 Subject 5 Subject 6 Subject 7 Subject 8
rs6025 GG GG GG GG GG - GG GG
rs2230500 GG GG GG GG GG GG GG GG
rs2246700 AT AT AT AT AA - AA AT
rs3783799 GG GG AG AG GG - GG GG
rs12587610 GG AG AG AA GG - AG AG
rs3825655 CC - CT - - - CC CC
rs1401296 CT CT TT CT CT - CT TT
rs1799998 CT CT CC CT CT CT TT CT
87
SNP id/
Haplotype
Allele
Subject 1 Subject 2 Subject 3 Subject 4 Subject 5 Subject 6 Subject 7 Subject 8
rs7412 CT CC CC CC CC CC CC CC
rs429358 TT TT TT TT TT - TT TT
rs12425791 AG AG GG GG AG AG GG AA
rs11833579 AG AG AG GG AG AG GG AA
rs1333049 GG CG GG CG CC CG GG CG
Class* 1 1 0 0 1 1 0 1
* Class 0 = Normal Subject Control Class, Class 1 = Ischemic Stroke Subject Case Class
This data is used for classification purposes.
88
4.1.3 Classification
We have used a machine learning tool WEKA (freely available at
http://www.cs.waikato.ac.nz/ml/weka) for the classification of genetic data. Classifiers are
chosen from the bayes, functions, lazy, meta and trees group of classifiers. Multiple
classifiers from different groups are chosen because classifiers produce different
classification results depending on the type of application. Nine classifiers are chosen for the
classification purposes namely, Bayes Net [164], Naïve Bayes [165, 166], IBk [167],
AdaBoostM1 [168], Classification via Regression [169], J48 [170], Random Forest [171],
Bagging [170, 172] and Multilayer Perceptron (MLP) [173].
4.1.3.1 Bayes Net
Bayes Net is a statistical model in which the conditional dependencies of variables using a
directed acyclic graph. They are generally used to answer probabilistic queries about the
variables. These probabilities are used to classify data.
We have used simple estimator algorithm for finding the conditional probability tables of the
bayes network and K2 for searching network structures.
4.1.3.2 Naïve Bayes
Naïve Bayes classifiers are probabilistic classifiers. They are based on probabilistic models
built by applying Bayes’ theorem to feature set with strong independence assumptions. The
features from the feature set contribute independently for classification and any possible
correlations among the features are abandoned. These classifiers require a small training set
for parameter estimation.
Normal distribution is used for numeric attributes.
4.1.3.3 IBk
IBk is based on k-nearest-neighbor. Nearest neighbors can be specified or can be
89
automatically calculated using leave-one-out cross-validation with a specified upper limit.
The similarity function in IBk calculates the similarity between training instances and the
depicted instances. The similarity results and classification performance records are fed to a
classification function to get the result for classification of an instance. The concept
description is updated after each classification.
We have used 1 as the number of neighbors to use without the classifier to output additional
info to the console. ‘No distance weighting’ method is used. Mean absolute error is used to
do cross-validation for regression. Nearest neighbor search is done using Euclidean distance.
No limit is set to the number of training instances.
4.1.3.4 AdaBoostM1
AdaBoost stands for adaptive boosting and M1 is a version of the adaptive boosting
algorithm. It works in aggregation with many other learning algorithms to improve its
performance. The results of the learning algorithms referred as weak learners are aggregated
to form a weighted sum which is the result of AdaBoostM1 classifier.
Individual learners might be weak but as long as each one produces slightly better results
than random guessing, the resultant final model congregates to a strong learner. It overcomes
the curse of dimensionality by selecting only the features that improve the prediction power
of the model.
We have used Decision stump as a base classifier. The number of iterations to be performed
is set to 10. Seed is set to 1. Reweighting is used and weight threshold for weight pruning is
set to 100.
4.1.3.5 Classification via Regression
In classification via regression a linear regression function is estimated based on the dataset.
Weights are calculated for different parameters having the machine learning objective for the
weights to reduce to least squares fitting. The resulting regression function is used for
90
classification of new data.
We have used M5P as base classifier with the option to generate model tree/rule. The
minimum numbers of instances allowed at a leaf node are set to 4.
4.1.3.6 J48
J48 is Weka implementation of C4.5 algorithm which is used to produce a decision tree
which is used for classification. It is a statistical classifier that builds decision tree based on
entropy using a set of training data. The data attribute labeled at each node is the one that
maximally splits the data into one class or the other depending on the information gain of
that attribute.
We have used a confidence factor of 0.25 with the minimum number of instances per leaf set
to 2. numFolds is set to 3.
4.1.3.7 Random Forest
It is an ensemble learning based classifier. Many decision trees are made at the training time,
the classification of an instance is the result of the mode of the classification of these trees or
mean prediction of each of the trees. It is a combination of bagging and random selection of
features.
The classification of a new object is made by feeding the input vector to each of the decision
tree in the random forest. Each tree votes for the instance and the decision for classification
is done for the class having majority votes. The maximum depth of trees is set to unlimited
and number of trees to be generated equal to 10.
4.1.3.8 Bagging
Bagging is a model averaging approach also known as bootstrap aggregating. It improves
stability and accuracy of classification algorithm. New training sets are generated by
sampling from the original training set uniformly and with replacement known as bootstrap
91
samples. Multiple models are fitted on the bootstrap samples. These models are combined by
averaging voting for classification of an instance.
We have set the size of each bag as a percent of the training set size equal to 100. Fast
decision tree learner is used as a base classifier without any restriction on the maximum tree
depth. The minimum total weight of the instances in a leaf are set to 2 with the minimum
proportion of the variance on all the data that needs to be present at a node in order for
splitting to be performed in regression trees set to 0.001 and number of folds equal to 3. 10
iterations are to be performed with seed equal to 1.
4.1.3.9 Multilayer Perceptron (MLP)
It is an ANN based model that maps inputs onto outputs. A MLP is a directed graph having
multiple layers of nodes and each layer is fully connected to the next layer. Network is
trained via backpropagation which is a supervised learning technique. It can even distinguish
the instances which are not linearly separable.
All of these classifiers belong to different classifier families differentiated by their
algorithmic nature. We have used 10 fold and 15 fold cross validation and 66% split for the
performance analysis of our scheme and the classifiers. We have used MLP with an autobuild
option to add and connect up hidden layers in the network. Learning rate is set to 0.3 and
momentum equal to 0.2 with normalized attributes and normalized numeric classes. The reset
option is set to true. The number of epochs to train through is set to 500. Validation threshold
is set to 20.
4.2 Results and Discussion
The results of the genetic data classification using different classifiers are summarized in
Table 4.4. Measures used to compare results are accuracy, sensitivity and specificity for the
92
selected classifiers using 10 fold, 15 fold cross validation and 66% split. For 10 fold cross
validation, MLP gave the best accuracy of 88.16%. MLP produced almost the same results
when applied using 10 fold, 15 fold cross validation and 66% split. IBk gave the worst
accuracy for 10 fold cross validation when applied to the selected SNPs data. For 15 fold
cross validation the best results are produced using AdaboostM1 (88.01%). MLP produced
an accuracy of 87.8% for 66% split test. Naïve bayes produced the overall lowest result of
80.61% when selected SNPs data was tested with 66% split.
MLP, AdaboostM1 and classification via regression gave best results for SNPs data. IBk and
naïve bayes did not perform very well for the classification of the data. Figures 4.2, 4.3 and
4.4 present the comparison of accuracy, specificity and sensitivity respectively.
93
Table 4.4: Classification results for genetic data using different classifiers
Classifier Cross Validation Split Test
10 Fold 15 Fold 66% split
ACC SPE SEN ACC SPE SEN ACC SPE SEN
BayesNet 84.46 82.54 87.06 84.38 82.54 86.89 82.14 78.57 87.71
Naïve Bayes 83.42 82.16 85.14 83.42 82.41 84.79 80.61 77.5 85.47
IBk 82.46 80.62 84.97 82.9 80.74 85.84 82.57 79.28 87.71
AdaBoostM1 87.42 88.58 85.84 88.01 88.19 87.76 86.49 85.71 87.71
ClassificationViaRegression 87.27 91.01 82.17 87.79 91.4 82.87 85.62 89.64 79.32
J48 84.6 91.01 75.87 83.94 90 75.7 82.57 88.57 73.18
RandomForest 84.97 88.58 80.07 84.75 89.21 78.67 83.22 85.36 79.89
94
Classifier Cross Validation Split Test
10 Fold 15 Fold 66% split
ACC SPE SEN ACC SPE SEN ACC SPE SEN
Bagging 84.97 90.12 77.97 85.34 90.5 78.32 83.22 87.86 75.98
MultilayerPerceptron 88.16 89.34 86.54 87.93 88.32 87.41 87.8 88.21 87.15
ACC = Accuracy, SPE = Specificity, SEN = Sensitivity
95
For the comparison of accuracies it is evident from Figure 4.2 that MLP gives the best
accuracy using the 15 fold cross validation on SNP data. AdaBoostM1and classification via
regression also produce close to best results i.e. 88.01% and 87.79% respectively using 15
fold cross validation. 15 fold cross validation gives better results for all classifiers for SNP
data.
Figure 4.3 shows that classification via regression achieves the highest specificity of 91.4%
using 15 fold cross validation on the SNPs data. The lowest specificity is produced by naïve
bayes on SNP data using 66% split test. 66% split test produced lowest results for all
classifiers. However, the results produced by cross validation for both 10 fold and 15 fold
provided high specificity.
Figure 4.4 gives the comparison of % sensitivities of different classifiers. It is evident from
the figure (Figure 4.4) that adaboostM1 produces highest sensitivity of 87.76% for the SNPs
data when performed using 15 fold cross validation. AdaboostM1 and IBk are very close to
the best, both with the sensitivities of 87.71% using 66% split test.
96
Figure 4.2: Comparison of % Accuracy of Classifiers using Genetic Data
76
78
80
82
84
86
88
90
Pe
rce
nta
ge A
ccu
racy
Classifiers
Accuracies of Classifiers
10-Fold CV
15-Fold CV
66% Split
97
Figure 4.3: Comparison of % Specificity of Classifiers using Genetic Data
70
75
80
85
90
95
Pe
rce
nta
ge S
pe
cifi
city
Classifiers
Specificities of Classifiers
10-Fold CV
15-Fold CV
66% Split
98
Figure 4.4: Comparison of % Sensitivity of Classifiers using Genetic Data
65
70
75
80
85
90
Pe
rce
nta
ge S
en
siti
vity
Classifiers
Sensitivities of Classifiers
10-Fold CV
15-Fold CV
66% Split
99
4.3 Conclusion
Genotype data is analyzed for ischemic stroke risk estimation in this chapter. The data for the
SNPs known for ischemic stroke risk is arranged for classification. Different classification
models are used to analyze and classify data. Highest accuracy is achieved using MLP.
Research is still going on to unfold the genes and SNPs responsible for ischemic strokes.
Stroke is a complex disease which can occur due to any or all of the multiple risk factors.
Risk gene allele in one population might not be present in other population and thus not
responsible for ischemic stroke in that population. Despite all these considerations our
proposed approach has shown an accuracy of 88.16%.
CHAPTER NO. 5
ANALYSIS & DISCUSSION-
CORRELATING PHYLOGENETIC
TREES
100
Analysis & Discussion- Correlating Phylogenetic
Trees
A phylogenetic tree is a branch diagram that is being traditionally used for representing the
evolutionary relations among different species and organisms. Similarity and dissimilarity in
different physical traits are used to build such trees. A more practical approach nowadays is
to generate these trees using the genes or proteins sequences for evolutionary relationships.
In such trees the similarity and dissimilarity is calculated based on the differences in genes or
protein sequences. These trees serve as a tool to reconstruct the evolutionary linkage between
groups of organisms and also to estimate the time of divergence between them. The number
of changes can be estimated using the phylogenetic trees.
Phylogenetic tree is also known as dendogram. The trees can be rooted or unrooted trees.
Rooted phylogenetic trees are also known as cladogram. These trees are built for the objects
that have descended from a common ancestor. The ancestor comes on the root. The paths
from roots to nodes represent the evolutionary time from the ancestor to the object. The
unrooted tree also known as phenogram is a tree where the objects are known to be related
but the ancestor is not confirmed or known. The path between such nodes does not tell about
the evolutionary time involved between the organisms or the objects.
Two approaches to build trees are discussed as under:
1. Traditional Approach:
Traditional approach used characterizing the organisms either through morphology
of organisms or through fossil record of the impressions left by an organism.
Morphological features include simply the physical characteristics of an organism
e.g. shape of beak, feathers, tail, number of legs, etc.
These features are helpful when the organisms under discussion are not extinct. In
101
the case where using the morphological features is not feasible, then the fossil
records are used. The fossilized remains are used for the estimation of age by
estimating the age of the rock surrounding the fossil imprints.
2. Modern Approach
Protein, Messenger RNA (mRNA), Ribosomal RNA (rRNA), genome sequenced
data etc. are used to characterize different organisms. This data is used to construct
phylogenetic trees. Large molecules consisting of one or more long chains of amino
acid residues are known as proteins. They are responsible for most of the functions
in organisms. They are responsible for DNA replication, stimuli response, molecule
transportation and as a catalyst of metabolic reactions etc. mRNA are molecules
responsible to carry genetic information from the DNA to the ribosome. rRNA is
RNA part of the ribosome. It is needed for protein synthesis in all organisms.
5.1 Phylogenetic Tree Construction Methods
Most commonly used methods for phylogenetic tree construction either fall into the distance
based methods or character based methods. Distance methods include Unweighted Pair
Group Method using arithmetic Averages (UPGMA) [174], Neighbor Joining (NJ) [175] and
Fitch and Margoliash algorithms [68]. Character based methods include MP [176, 177] and
ML [178, 179] methods for construction of phylogenetic trees. The most commonly used
methods are discussed here.
5.1.1 Unweighted Pair Group Method using Arithmetic
Averages (UPGMA)
It clusters nodes at each stage of the tree and forms a new node on the tree. The tree is built
bottom up. The tree is built on the assumption that all the nodes are equidistant from the root.
At each level the length of the branch is determined by the difference in the heights of the
102
nodes at each end of the branch. Equally distant nodes pose a problem for the molecular
clock hypothesis because UPGMA assumes a constant rate of evolution. It produces a rooted
tree that is the true picture of the pairwise similarity or dissimilarity matrix. At each step two
nearest clusters are combined to make a higher level cluster.
5.1.2 Neighbor Joining (NJ)
It works just like the UPGMA with a difference that new distance matrix is calculated at each
of the iterations. The distance matrix is calculated using Hamming distance between each
node. This method is not based on the additivity of the nodes and works even if the nodes are
not additive. This method produces and unrooted tree. If a common ancestor has to be
assigned to the tree then it has to be selected as an outgroup.
5.1.3 Maximum Parsimony (MP)
MP searches through all possible tree structures and assigning cost to each tree. The most
parsimonious tree is chosen to be the one that requires least changes to explain the aligned
data. This tree is the one that explains the evolutionary pattern of the objects analyzed.
5.2.4 Maximum Likelihood (ML)
It is a computationally intensive approach that optimizes the likelihood of observed data
given a tree and a nucleotide evolution model. It is based on probability theory. It tries to find
a tree that has the highest probability under a specific evolution model. This method has the
disadvantage that it relies on the assumption that the evolution model is accurate and correct.
If it is provided with a false model the resultant tree will not be consistent. It can generate
multiple trees having the same probabilities.
103
5.2 Softwares and Tools Available for Phylogenetic Tree
Construction
A number of software tools are available for the construction of phylogenetic trees. Some of
the famous tools and the implementation methods of these tools are given in Table 5.1.
Table 5.1: Phylogenetic tools and their implementation methods
Tool Description
Method
Distance Character
NJ UPGMA MP ML
ClustalW2
[180]
It is used for multiple sequence alignment and
phylogenetic tree construction √ √
MEGA [181] It is an integrated tool for sequence alignment
and phylogenetic tree construction and analysis √ √
PAUP [182] It is a program for inferring phylogenetic trees
on the basis of parsimony. √
PAUP*
[183]
It is the version 4 onwards program for PAUP.
It has an additional support for distance matrix
and likelihood based methods for phylogenetic
tree construction.
√ √ √ √
PHYLIP
[184]
It is a computational package for constructing
and analyzing the phylogenetic trees. √ √ √ √
BioNumerics
[185]
It is a software for analysis with a wide range
of Bioinformatics applications. One of its
application is the inference of phylogenetics.
√ √ √ √
NJ= Neighbor-Joining, MP= Maximum Parsimony, ML= Maximum Likelihood, UPGMA= Unweighted Pair
Group Method using Arithmetic Averages.
104
5.3 Bioinformatics Databanks
There are many bioinformatics databanks. Some of the famous databanks include:
1. European Molecular Biology Laboratories (EMBL) [186]
It is nucleotide sequence resource that contains submissions of DNA and RNA
sequences by different researchers.
2. Database of Genomic Variants (DGV) [187]
This database contains control data from studies that investigate the genomic
variation associations with phenotype data.
3. European Genome Phenome Archive (EGA) [188]
It has genotype data from various case controls, population and family studies.
4. dbSNP - NCBI [189]
This database consists of Single Nucleotide Polymorphisms and their relation with
heritable phenotypes.
5. The SNP Consortium Ltd [190]
This site has Single nucleotide polymorphisms (SNPs) which are common DNA
sequence variations among individuals and have great significance for biomedical
research.
6. HGBASE [191]
HGBASE summarizes all known sequence variations in the human genome. They
facilitate researches on genotypes effects on common diseases, drug responses, and
other complex phenotypes.
7. HAPMAP [192]
The International HapMap Project is a partnership of scientists and funding agencies
from Canada, China, Japan, Nigeria, United Kingdom and United States to develop
a public resource that will help researchers find genes associated with human
105
disease and response to pharmaceuticals.
8. 1000 Genomes [193]
It is a catalog of human genetic variations.
9. ICGC [194]
ICGC has a comprehensive description of genomic, transcriptomic and epigenomic
changes in 50 different tumor types and/or subtypes. This data is of clinical and
societal importance all over the world.
10. COSMIC [195]
COSMIC is designed to store and display somatic mutation information and related
details. It contains information relating to human cancers.
11. HGMD [196]
The Human Gene Mutation Database (HGMD) keeps record of published gene
lesions responsible for human inherited disease.
12. OMIM [197]
Online Mendelian Inheritance in Man (OMIM) is a catalog of human genes and
genetic disorders. The database contains textual information, pictures, and reference
information.
13. GeneTests [198]
GeneTests is a medical genetics information resource developed for physicians,
other healthcare providers, and researchers.
14. Genomic Variants[187]
The Database of Genomic Variants is a catalog of control data for studies aiming to
correlate genomic variation with phenotypic data.
15. Mitelman Database of Chromosome Aberrations in Cancer [199]
Mitelman Database of Chromosome Aberrations in Cancer is a repository that
106
relates chromosomal aberrations to tumor characteristics. The relations are based
either on individual cases or associations.
16. Genetic Association Database [200]
The Genetic Association Database is an archive of human genetic association
studies of complex diseases and disorders.
5.4 Materials and Methods
5.4.1 Data
The genotype data of 1417 individuals is downloaded from publically accessable HapMap
database HapMap Genome Browser release #28, The International Hapmap Project,
available at http://hapmap.ncbi. nlm.nih.gov
ftp://ftp.ncbi.nlm.nih.gov/hapmap/genotypes/2010-08_phaseII+III/ website (Jul 12, 2014).
The raw genotype files from different populations were merged together to generate a
combined genotype file. The details of the population data are given in Table 5.2. There are a
total of 11 populations.
Table 5.2: Details of selected HapMap population data
Population Number of
Individuals Name Detail
ASW African ancestry in Southwest USA 87
CEU
Utah residents with Northern and
Western European ancestry
174
CHB Han Chinese in Beijing, China 139
107
Population Number of
Individuals Name Detail
CHD
Chinese in Metropolitan Denver,
Colorado
109
GIH Gujarati Indians in Houston, Texas 101
JPT Japanese in Tokyo, Japan 116
LWK Luhya in Webuye, Kenya 110
MEX
Mexican ancestry in Los Angeles,
California
86
MKK Maasai in Kinyawa, Kenya 184
TSI Toscans in Italy 102
YRI Yoruban in Ibadan, Nigeria 209
Total 1417
5.4.2 Method
A meta-analysis has been conducted to identify the risk of different SNPs for ischemic
stroke. The risk associated with the SNPs, genes and genotype or haplotype values are given
in Table 5.3.
108
Table 5.3: Stroke associated genes/ locus, SNP id/ haplotype and corresponding allele risk
Gene/ Locus SNP id Genotype/ Haplotype Allele Risk
MTHFR
rs1801133 TT 4
rs1801131 CC 1
rs12121543
C-T-T-T-G-A 1
rs13306553
rs9651118
rs1801133
rs2274976
rs1801131
F5 rs6025
AA 9
AG 2.7
PRKCH
rs2230500
AA 1.4
AG 1.4
rs2246700
AA 1
AT 1
rs3783799
AA 1.4
AG 1.4
rs12587610
AG 1
GG 1
rs3825655 CC 1
109
Gene/ Locus SNP id Genotype/ Haplotype Allele Risk
CT 1
PROC rs1401296 CT 1
CYP11B2 rs1799998
CT 1
TT 1
APOE- E4
rs7412
CC 1
CT 1
rs429358 CT 1.4
Near NINJ2
rs12425791
AA 1.4
AG 1.4
rs11833579
AA 1.4
AG 1.4
9p21.3 rs1333049
CC 1.15
CG 1.15
Data for the SNPs given in Table 5.3 are extracted from the genotype data of 1417 samples
of different populations. The allele frequencies for the SNPs calculated for each population
are given in Table 5.4. Frequency values are given for each population against each SNP
allele. A value of zero (‘0’) indicates that a particular population does not have that particular
allele value. On the contrary a dash (‘-’) represents that the population under discussion does
not have that particular SNP.
110
Table 5.4: Allele frequencies for the SNPs from sample population data
Gene/ Locus SNP id Genotype/ Haplotype
Populations
ASW CEU CHB CHD GIH JPT LWK MEX MKK TSI YRI
MTHFR
rs1801133 TT 5 14 34 13 2 17 0 16 1 26 1
rs1801131 CC 1 17 3 8 16 2 5 4 16 13 4
rs12121543 C-T-T-T-G-A 15 51 37 12 3 17 - 15 1 26 0
rs13306553
rs9651118
rs1801133
rs2274976
rs1801131
F5 rs6025
AA - 1 1 0 0 0 0 0 - 0 0
AG - 4 0 1 1 0 1 3 - 1 0
PRKCH rs2230500 AA - 0 0 0 - 0 0 - - 0 0
111
Gene/ Locus SNP id Genotype/ Haplotype
Populations
ASW CEU CHB CHD GIH JPT LWK MEX MKK TSI YRI
AG - 0 7 6 - 3 1 - - 1 0
rs2246700
AA - - 14 - - 15 - - - - 20
AT - - 23 - - 26 - - - - 36
rs3783799
AA - - 1 - - 3 - - - - 0
AG - - 14 - - 16 - - - - 0
rs12587610
AG - 47 20 - - 6 - - - - 13
GG - 5 13 - - 29 - - - - 75
rs3825655
CC - 67 24 - - 15 - - - - 76
CT - 5 15 - - 22 - - - - 2
PROC rs1401296 CT - 40 26 - - 25 - - - - 29
CYP11B2 rs1799998 CT 26 85 62 57 49 54 39 40 46 55 53
112
Gene/ Locus SNP id Genotype/ Haplotype
Populations
ASW CEU CHB CHD GIH JPT LWK MEX MKK TSI YRI
TT 60 51 67 35 28 44 68 30 129 19 145
APOE- E4
rs7412
CC 68 - 109 92 93 101 95 75 154 90 160
CT 16 - 24 15 7 12 11 9 22 9 36
rs429358 CT - - 0 0 - 1 - - - - 2
Near NINJ2
rs12425791
AA 1 4 9 5 10 10 1 11 30 8 2
AG 16 50 43 42 38 54 13 43 3 38 34
rs11833579
AA 5 8 17 10 11 14 4 17 6 10 8
AG 25 56 56 57 44 63 29 38 50 40 76
9p21.3 rs1333049
CC 2 36 30 28 20 33 6 16 10 23 6
CG 39 82 73 48 53 53 46 52 70 60 60
113
The frequencies are then used to calculate weighted risk for each population. All the frequencies
are converted into percentages by using Equation 5.1. Percentage conversion is done so that the
populations and their risks could be compared. After percentage conversion weighted risk for
each allele is calculated for each population using Equation 5.2. Finally aggregate weighted risk
for each population is calculated using Equation 5.3. If a population does not have a particular
SNP then the weighted average is calculated excluding that allele.
𝑃𝑒𝑟𝑐𝑒𝑛𝑡 𝑆𝑁𝑃 𝑎𝑙𝑙𝑒𝑙𝑒 𝑓𝑟𝑒𝑞𝑢𝑒𝑛𝑐𝑦 = 𝑆𝑁𝑃 𝑎𝑙𝑙𝑒𝑙𝑒 𝑓𝑟𝑒𝑞𝑢𝑒𝑛𝑐𝑦
𝑃𝑜𝑝𝑢𝑙𝑎𝑡𝑖𝑜𝑛 𝑠𝑎𝑚𝑝𝑙𝑒 𝑠𝑖𝑧𝑒× 100 (5.1)
𝑊𝑒𝑖𝑔ℎ𝑡𝑒𝑑 𝑆𝑁𝑃 𝑎𝑙𝑙𝑒𝑙𝑒 𝑟𝑖𝑠𝑘 = 𝑃𝑒𝑟𝑐𝑒𝑛𝑡 𝑆𝑁𝑃 𝑎𝑙𝑙𝑒𝑙𝑒 𝑓𝑟𝑒𝑞𝑢𝑒𝑛𝑐𝑦 × 𝐴𝑙𝑙𝑒𝑙𝑒 𝑟𝑖𝑠𝑘 (5.2)
𝐴𝑔𝑔𝑟𝑒𝑔𝑎𝑡𝑒 𝑤𝑒𝑖𝑔ℎ𝑡𝑒𝑑 𝑟𝑖𝑠𝑘 = ∑ 𝑊𝑒𝑖𝑔ℎ𝑡𝑒𝑑 𝑆𝑁𝑃 𝑎𝑙𝑙𝑒𝑙𝑒 𝑟𝑖𝑠𝑘𝑛
𝑖=1∑ 𝐴𝑙𝑙𝑒𝑙𝑒 𝑟𝑖𝑠𝑘𝑛
𝑖=1⁄ (5.3)
where,
n = number of SNP alleles
The detailed SNP allele frequency percentage for each SNP allele and the aggregate weighted
risks for each population are given in Table 5.5.
114
Table 5.5: Percent allele frequencies for the SNPs from sample population data
Gene/
Locus
SNP id
Genotype/
Haplotype
Risk
Percent Allele Frequency for Populations
ASW CEU CHB CHD GIH JPT LWK MEX MKK TSI YRI
MTHFR
rs1801133 TT 4 5.74 8.54 24.82 11.93 1.98 15.04 0 18.6 0.54 25.74 0.49
rs1801131 CC 1 1.15 10.3 2.19 7.34 15.84 1.77 4.55 4.65 8.7 12.75 1.97
rs12121543
C-T-T-T-G-
A
1 17.24 29.31 26.62 11.01 2.97 14.66 - 17.44 0.54 25.49 0
rs13306553
rs9651118
rs1801133
rs2274976
rs1801131
F5 rs6025 AA 9 - 0.61 0.73 0.00 0 0 0 0 - 0 0
115
Gene/
Locus
SNP id
Genotype/
Haplotype
Risk
Percent Allele Frequency for Populations
ASW CEU CHB CHD GIH JPT LWK MEX MKK TSI YRI
AG 2.7 - 2.42 0.00 0.92 0.99 0 0.92 3.49 - 0.98 0
PRKCH
rs2230500
AA 1.4 - 0 0.00 0.00 - 0 0 - - 0 0
AG 1.4 - 0 5.22 5.50 - 2.68 0.92 - - 0.98 0
rs2246700
AA 1 - - 31.11 - - 33.33 - - - - 22.22
AT 1 - - 51.11 - - 57.78 - - - - 40
rs3783799
AA 1.4 - - 2.22 - - 6.67 - - - - 0
AG 1.4 - - 31.11 - - 35.56 - - - - 0
rs12587610
AG 1 - 52.81 44.44 - - 13.33 - - - - 14.61
GG 1 - 5.62 28.89 - - 64.44 - - - - 84.27
116
Gene/
Locus
SNP id
Genotype/
Haplotype
Risk
Percent Allele Frequency for Populations
ASW CEU CHB CHD GIH JPT LWK MEX MKK TSI YRI
rs3825655
CC 1 - 93.06 60.00 - - 37.5 - - - - 97.44
CT 1 - 6.94 37.50 - - 55 - - - - 2.56
PROC rs1401296 CT 1 - 44.44 57.78 - - 56.82 - - - - 32.22
CYP11B2 rs1799998
CT 1 30.23 52.15 46.27 53.27 50.52 49.54 35.78 46.51 25 56.12 26.37
TT 1 69.77 31.29 50.00 32.71 28.87 40.37 62.39 34.88 70.11 19.39 72.14
APOE-
E4
rs7412
CC 1 79.07 - 81.95 85.98 92.08 89.38 89.62 89.29 87.5 90.91 81.22
CT 1 18.6 - 18.05 14.02 6.93 10.62 10.38 10.71 12.5 9.09 18.27
rs429358 CT 1.4 - - 0.00 - - 2.27 - - - - 2.22
Near rs12425791 AA 1.4 1.15 2.42 6.57 4.59 9.9 8.85 0.94 12.79 16.3 7.84 0.99
117
Gene/
Locus
SNP id
Genotype/
Haplotype
Risk
Percent Allele Frequency for Populations
ASW CEU CHB CHD GIH JPT LWK MEX MKK TSI YRI
NINJ2 AG 1.4 18.39 30.3 31.39 38.53 37.62 47.79 12.26 50 1.63 37.25 16.75
rs11833579
AA 1.4 5.81 4.88 12.69 9.35 11.11 12.61 3.77 20.24 3.3 9.8 3.94
AG 1.4 29.07 34.15 41.79 53.27 44.44 56.76 27.36 45.24 27.47 39.22 37.44
9p21.3 rs1333049
CC 1.15 2.3 21.82 21.90 25.69 19.8 29.2 5.45 18.6 5.43 22.55 2.96
CG 1.15 44.83 49.7 53.28 44.04 52.48 46.9 44.82 60.47 38.04 58.82 29.56
Aggregate Weighted Risk 20.64 15.67 20.91 15.14 14.7 20.98 10.07 18.84 18.14 16.85 14.19
118
The aggregate weighted risk is used to calculate distance matrix. A pairwise difference is
calculated for each population. Distance matrix shown in Table 5.6 is calculated using Equation
5.4.
𝑑𝑎,𝑏 = |𝐴𝑔𝑔𝑟𝑒𝑔𝑎𝑡𝑒 𝑤𝑒𝑖𝑔ℎ𝑡𝑒𝑑 𝑟𝑖𝑠𝑘𝑎 − 𝐴𝑔𝑔𝑟𝑒𝑔𝑎𝑡𝑒 𝑤𝑒𝑖𝑔ℎ𝑡𝑒𝑑 𝑟𝑖𝑠𝑘𝑏| (5.4)
where,
da,b is a cell in the distance matrix.
a and b are any two populations from the set of 11 populations.
The calculated distance matrix is used to create a phylogenetic tree for the populations (Figure
5.1). The tree is generated using Matlab 7.14 by neighbor joining method.
119
Table 5.6: Distance matrix for all populations
Distance ASW CEU CHB CHD GIH JPT LWK MEX MKK TSI YRI
ASW 4.97 0.27 5.5 5.94 0.34 10.57 1.8 2.5 3.79 6.45
CEU 4.97 5.24 0.53 0.97 5.31 5.6 3.17 2.47 1.18 1.48
CHB 0.27 5.24 5.77 6.21 0.07 10.84 2.07 2.77 4.06 6.72
CHD 5.5 0.53 5.77 0.44 5.84 5.07 3.7 3 1.71 0.95
GIH 5.94 0.97 6.21 0.44 6.28 4.63 4.14 3.44 2.15 0.51
JPT 0.34 5.31 0.07 5.84 6.28 10.91 2.14 2.84 4.13 6.79
LWK 10.57 5.6 10.84 5.07 4.63 10.91 8.77 8.07 6.78 4.12
MEX 1.8 3.17 2.07 3.7 4.14 2.14 8.77 0.7 1.99 4.65
MKK 2.5 2.47 2.77 3 3.44 2.84 8.07 0.7 1.29 3.95
TSI 3.79 1.18 4.06 1.71 2.15 4.13 6.78 1.99 1.29 2.66
YRI 6.45 1.48 6.72 0.95 0.51 6.79 4.12 4.65 3.95 2.66
120
Figure 5.1: Phylogenetic tree using our distance matrix
121
5.5 Results and Discussion
JPT population has the highest aggregate weighted risk of 20.98 while LWK has the lowest
aggregate weighted risk of 10.07. Figure 5.2 shows the percent SNP allele frequency of all 11
populations. Different populations have different frequencies of the SNP alleles. ASW and MKK
are the least to be described by the genetic data and SNPs. JPT, CHH and CHB are best
described by the genetic data and SNPs chosen for analysis.
Combined allele frequencies for all populations are shown for each SNP are shown in Figure
5.3. It is obvious from the figure (Figure 5.3) that data for rs7412 on APOE-e4 gene is present in
most of the populations. AA allele for rs2230500 is not present in all the populations. AA allele
for rs6025 is present in CEU and CHB only. CT allele for rs429358 is present only in JPT and
YRI.
The phylogenetic tree generated using the calculated distance matrix (Figure 5.1) shows that
MEX, MKK and TSI are closely related w.r.t. distance matrix based on aggregate weighted risk.
Similarly, CHB, JPT and ASW form another group of closely related populations. CHD, GIH,
YRI and CEU are the third closely related group. Population LWK is distant from other
populations.
Another phylogenetic tree is constructed based on FST distances calculated by Altshuler et. al.
[201] in Figure 5.4. The phylogenetic relationships observed in the second tree (Figure 5.4)
shows different results when compared with the one in Figure 5.1. First group of closely related
populations are CEU, TSI, MEX and GIH. Second group comprises of CHB, CHD and JPT. The
third group comprises of LWK, YRI, ASW and MKK.
122
Figure 5.2: Comparison of % allele frequency of all sample populations
0
20
40
60
80
100
120
ASW CEU CHB CHD GIH JPT LWK MEX MKK TSI YRI
rs1801133
rs1801131
Haplotype
rs6025 AA
rs6025 AG
rs2230500 AA
rs2230500 AG
rs2246700 AA
rs2246700 AT
rs3783799 AA
rs3783799 AG
rs12587610 AG
rs12587610 GG
rs3825655 CC
rs3825655 CT
rs1401296
rs1799998 CT
rs1799998 TT
rs7412 CC
rs7412 CT
123
Figure 5.3: Combined allele frequencies of all sample populations
0
100
200
300
400
500
600
700
800
900
1000
YRI
TSI
MKK
MEX
LWK
JPT
GIH
CHD
CHB
CEU
ASW
124
Figure 5.4: Constructed phylogenetic tree using FST matrix as calculated by Altshuler et. al. [201]
125
One of reasons for the difference in both trees is that different populations might be genetically
close but not w.r.t. ischemic stroke risk. Another reason for this difference is that when analyzing
the genetic distances in general, too many SNPs and a large amount of genetic data are involved.
While for ischemic risk estimation we have used only a few selected SNPs. This clearly
indicates that genetic closeness of populations is not an indication that their risk for some
disease would be the same as well.
5.6 Conclusion
This research addresses the correlation of phylogenetic trees with ischemic stroke risk. Two
types of phylogenetic trees are generated; one based on genetic distance matrix and the other
based on ischemic stroke risk difference matrix. A strenuous comparison of genetic distance and
ischemic stroke risk difference matrix in the form of phylogenetic trees is presented in this
chapter. Both trees show different relationships among different populations. These relationships
indicate that different populations might be close genetically but they might have differences as
far as disease risks are concerned. The reason for this difference is that genetic distances are
calculated using data of all genes while the ischemic stroke risk difference matrix is calculated
using selected SNPs only.
CHAPTER NO. 6
CONCLUSION & FUTURE
RECOMENDATIONS
126
Conclusion and Future Recommendations
6.1 Conclusion
The real-time analysis of genetic data and ultrasound imaging provides a quick means by which
to qualitatively analyze data and draw meaningful interpretation. This subsequently helps in
analyzing the risk factors and to take into account the preventive measures for the stroke. The
trees can be used for the representation of multi-dimensional data sets and accessing the genetic
distance of a population based on ischemic stroke risk.
Multi-dimensional feature sets – Image and Genetic features are used for the early diagnosis and
assessment of the stroke risk that would be beneficial for the process of correct identification of
individuals at high risk of ischemic stroke. This research is an effort to improve and contribute
in the visual assessment procedure conducted by the medical personals. This research not only
facilitates the medical personals but also the community at large by elucidating the risk factors.
The proposed approach plays an important role by contributing to the area of Computer Aided
Diagnostics and Preventive Studies. The research facilitates in meaningful interpretation of
genetic and image based data that ultimately helps in critical analysis of the risk factors and the
preventive measures for ischemic strokes.
6.2 Future Recommendations
This research can be enhanced in future by incorporating the texture features of plaque into the
feature set already used. This may strengthen the risk estimation task. In addition to this multiple
CA imaging modalities such as CTA, MRA, CAG, DSA etc. can also be used for improvement
of the accuracy rate of the classification algorithms.
127
References
1 Organization, W. H. Top 10 causes of death: Fact sheet no. 310: Who 2011. Ref Type:
Report L U. (2011).
2 Naghavi, M., Wang, H., Lozano, R., Davis, A., Liang, X., Zhou, M., Vollset, S. E. V.,
Abbasoglu Ozgoren, A., Norman, R. E., and Vos, T. Global, regional, and national age–
sex specifi c all-cause and cause-specifi c mortality for 240 causes of death, 1990–2013:
A systematic analysis for the global burden of disease study 2013. The lancet,
385(9963): 117-171. (2015).
3 Association, A. H. A. A. S. Types of stroke Retrieved 12/02, 2014, from
http://www.strokeassociation.org/STROKEORG/AboutStroke/TypesofStroke/Types-of-
Stroke_UCM_308531_SubHomePage.jsp
4 Cerebrovascular accident - national library of medicine - pubmed health. (2012), 2012,
from http://www.ncbi.nlm.nih.gov/pubmedhealth/PMH0001740/
5 Deep venous thrombosis: Medlineplus medical encyclopedia. Retrieved 20/04, 2015,
from http://www.nlm.nih.gov/medlineplus/ency/article/000156.htm
6 Stroke.org. (2014). What is stroke? Retrieved 10/12/2014, 2015, from
http://www.stroke.org/site/PageServer?pagename=TYPE
7 Goldschmidt-Clermont, P. J., Dong, C., Seo, D. M., and Velazquez, O. C.
Atherosclerosis, inflammation, genetics, and stem cells: 2012 update. Current
atherosclerosis reports, 14(3): 201-210. (2012).
8 National Heart Lung and Blood Institute, N. What is atherosclerosis? Retrieved
10/12/2014, 2014, from http://www.nhlbi.nih.gov/health/health-
topics/topics/atherosclerosis/
9 Bots, M. L., Hoes, A. W., Koudstaal, P. J., Hofman, A., and Grobbee, D. E. Common
carotid intima-media thickness and risk of stroke and myocardial infarction the rotterdam
128
study. Circulation, 96(5): 1432-1437. (1997).
10 Simons, L. A., McCallum, J., Friedlander, Y., and Simons, J. Risk factors for ischemic
stroke dubbo study of the elderly. Stroke, 29(7): 1341-1346. (1998).
11 Touboul, P.-J., Labreuche, J., Vicaut, E., and Amarenco, P. Carotid intima-media
thickness, plaques, and framingham risk score as independent determinants of stroke
risk. Stroke, 36(8): 1741-1745. (2005).
12 Johnson, H., Turke, T. L., Grossklaus, M., Dall, T., Carimi, S., Koenig, L. M.,
Aeschlimann, S. E., Korcarz, C. E., and Stein, J. H. Detection of subclinical
atherosclerosis by office-based carotid ultrasound increases prescription of preventive
therapies. Journal of the American College of Cardiology, 55(10): A58. E557. (2010).
13 Faita, F., Gemignani, V., Bianchini, E., Giannarelli, C., Ghiadoni, L., and Demi, M. Real-
time measurement system for evaluation of the carotid intima-media thickness with a
robust edge operator. Journal of Ultrasound in Medicine, 27(9): 1353-1361. (2008).
14 Gariepy, J., Massonneau, M., Levenson, J., Heudes, D., and Simon, A. Evidence for in
vivo carotid and femoral wall thickening in human hypertension. Groupe de prévention
cardio-vasculaire en médecine du travail. Hypertension, 22(1): 111-118. (1993).
15 Liguori, C., Paolillo, A., and Pietrosanto, A. An automatic measurement system for the
evaluation of carotid intima-media thickness. Instrumentation and Measurement, IEEE
Transactions on, 50(6): 1684-1691. (2001).
16 Pignoli, P., and Longo, T. Evaluation of atherosclerosis with b-mode ultrasound imaging.
The Journal of nuclear medicine and allied sciences, 32(3): 166-173. (1987).
17 Rocha, R., Campilho, A., Silva, J., Azevedo, E., and Santos, R. Segmentation of the
carotid intima-media region in b-mode ultrasound images. Image and Vision Computing,
28(4): 614-625. (2010).
18 Selzer, R. H., Hodis, H. N., Kwong-Fu, H., Mack, W. J., Lee, P. L., Liu, C.-r., and Liu,
C.-h. Evaluation of computerized edge tracking for quantifying intima-media thickness
129
of the common carotid artery from b-mode ultrasound images. Atherosclerosis, 111(1):
1-11. (1994).
19 Stein, J. H., Korcarz, C. E., Mays, M. E., Douglas, P. S., Palta, M., Zhang, H., LeCaire,
T., Paine, D., Gustafson, D., and Fan, L. A semiautomated ultrasound border detection
program that facilitates clinical measurement of ultrasound carotid intima-media
thickness. Journal of the American Society of Echocardiography, 18(3): 244-251.
(2005).
20 Touboul, P.-J., Prati, P., Scarabin, P.-Y., Adrai, V., Thibout, E., and Ducimetière, P. Use of
monitoring software to improve the measurement of carotid wall thickness by b-mode
imaging. Journal of hypertension, 10: S37-S42. (1992).
21 Bastida-Jumilla, M. C., Menchón-Lara, R. M., Morales-Sánchez, J., Verdú-Monedero,
R., Larrey-Ruiz, J., and Sancho-Gómez, J. L. Segmentation of the common carotid artery
walls based on a frequency implementation of active contours. Journal of digital
imaging, 26(1): 129-139. (2013).
22 Delsanto, S., Molinari, F., Giustetto, P., Liboni, W., Badalamenti, S., and Suri, J. S.
Characterization of a completely user-independent algorithm for carotid artery
segmentation in 2-d ultrasound images. Instrumentation and Measurement, IEEE
Transactions on, 56(4): 1265-1274. (2007).
23 Golemati, S., Stoitsis, J., Sifakis, E. G., Balkizas, T., and Nikita, K. S. Using the hough
transform to segment ultrasound images of longitudinal and transverse sections of the
carotid artery. Ultrasound in medicine & biology, 33(12): 1918-1932. (2007).
24 Molinari, F., Zeng, G., and Suri, J. S. Intima-media thickness: Setting a standard for a
completely automated method of ultrasound measurement. Ultrasonics, Ferroelectrics,
and Frequency Control, IEEE Transactions on, 57(5): 1112-1124. (2010).
25 Petroudi, S., Loizou, C., Pantziaris, M., Pattichis, M., and Pattichis, C. 2011. A fully
automated method using active contours for the evaluation of the intima-media thickness
in carotid us images. Engineering in Medicine and Biology Society, EMBC, 2011 Annual
130
International Conference of the IEEE. p. 8053-8057
26 Mahmoud, A., Morsy, A., and de Groot, E. 2010. A new gradient-based algorithm for
edge detection in ultrasonic carotid artery images. Engineering in Medicine and Biology
Society (EMBC), 2010 Annual International Conference of the IEEE. p. 5165-5168
27 Gustavsson, T., Abu-Gharbieh, R., Hamarneh, G., and Liang, Q. 1997. Implementation
and comparison of four different boundary detection algorithms for quantitative
ultrasonic measurements of the human carotid artery. Computers in Cardiology 1997. p.
69-72
28 Holdfeldt, P., Viberg, M., and Gustavsson, T. 2008. A new method based on dynamic
programming for boundary detection in ultrasound image sequences. Engineering in
Medicine and Biology Society, 2008. EMBS 2008. 30th Annual International Conference
of the IEEE. p. 3072-3074
29 Liang, Q., Wendelhag, I., Wikstrand, J., and Gustavsson, T. A multiscale dynamic
programming procedure for boundary detection in ultrasonic artery images. Medical
Imaging, IEEE Transactions on, 19(2): 127-142. (2000).
30 Cheng, D.-C., and Jiang, X. Detections of arterial wall in sonographic artery images
using dual dynamic programming. Information Technology in Biomedicine, IEEE
Transactions on, 12(6): 792-799. (2008).
31 Cheng, D.-c., Schmidt-Trucksass, A., Cheng, K.-s., Sandrock, M., Pu, Q., and Burkhardt,
H. 1999. Automatic detection of the intimal and the adventitial layers of the common
carotid artery wall in ultrasound b-mode images using snakes. Image Analysis and
Processing, 1999. Proceedings. International Conference on. p. 452-457
32 Cheng, D.-c., Schmidt-Trucksäss, A., Cheng, K.-s., and Burkhardt, H. Using snakes to
detect the intimal and adventitial layers of the common carotid artery wall in sonographic
images. Computer methods and programs in biomedicine, 67(1): 27-37. (2002).
33 Loizou, C. P., Pattichis, C. S., Pantziaris, M., Tyllis, T., and Nicolaides, A. Snakes based
segmentation of the common carotid artery intima media. Medical & biological
131
engineering & computing, 45(1): 35-49. (2007).
34 Molinari, F., Meiburger, K. M., Saba, L., Zeng, G., Acharya, U. R., Ledda, M.,
Nicolaides, A., and Suri, J. S. Fully automated dual-snake formulation for carotid intima-
media thickness measurement a new approach. Journal of Ultrasound in Medicine,
31(7): 1123-1136. (2012).
35 Xu, X., Zhou, Y., Cheng, X., Song, E., and Li, G. Ultrasound intima–media segmentation
using hough transform and dual snake model. Computerized Medical Imaging and
Graphics, 36(3): 248-258. (2012).
36 Bastida-Jumilla, M., Morales-Sánchez, J., Verdú-Monedero, R., Larrey-Ruiz, J., and
Sancho-Gómez, J.-L. 2010. Detection of the intima and media walls of the carotid artery
with geodesic active contours. Image Processing (ICIP), 2010 17th IEEE International
Conference on. p. 2213-2216
37 Moursi, S. G., and El-Sakka, M. R. 2008. Active contours initialization for ultrasound
carotid artery images. Computer Systems and Applications, 2008. AICCSA 2008.
IEEE/ACS International Conference on. p. 629-636
38 Meiburger, K. M., Molinari, F., Zeng, G., Saba, L., and Suri, J. S. 2011. Carotid
automated ultrasound double line extraction system (cadles) via edge-flow. Engineering
in Medicine and Biology Society, EMBC, 2011 Annual International Conference of the
IEEE. p. 575-578
39 Molinari, F., Liboni, W., Giustetto, P., Badalamenti, S., and Suri, J. S. Automatic
computer-based tracings (act) in longitudinal 2-d ultrasound images using different
scanners. Journal of Mechanics in Medicine and Biology, 9(04): 481-505. (2009).
40 Delsanto, S., Molinari, F., Liboni, W., Giustetto, P., Badalamenti, S., and Suri, J. S. 2006.
User-independent plaque characterization and accurate imt measurement of carotid artery
wall using ultrasound. Engineering in Medicine and Biology Society, 2006. EMBS'06.
28th Annual International Conference of the IEEE. p. 2404-2407
41 Molinari, F., Delsanto, S., Giustetto, P., Liboni, W., Badalamenti, S., and Suri, J. User-
132
independent plaque segmentation and accurate intima-media thickness measurement of
carotid artery wall using ultrasound. (2008).
42 Destrempes, F., Meunier, J., Giroux, M.-F., Soulez, G., and Cloutier, G. Segmentation in
ultrasonic b-mode images of healthy carotid arteries using mixtures of nakagami
distributions and stochastic optimization. Medical Imaging, IEEE Transactions on,
28(2): 215-229. (2009).
43 Destrempes, F., Soulez, G., Giroux, M.-F., Meunier, J., and Cloutier, G. 2009.
Segmentation of plaques in sequences of ultrasonic b-mode images of carotid arteries
based on motion estimation and nakagami distributions. Ultrasonics Symposium (IUS),
2009 IEEE International. p. 2480-2483
44 Golemati, S., Stoitsis, J., Balkizas, T., and Nikita, K. 2006. Comparison of b-mode, m-
mode and hough transform methods for measurement of arterial diastolic and systolic
diameters. Engineering in Medicine and Biology Society, 2005. IEEE-EMBS 2005. 27th
Annual International Conference of the. p. 1758-1761
45 Golemati, S., Tegos, T. J., Sassano, A., Nikita, K. S., and Nicolaides, A. N. Echogenicity
of b-mode sonographic images of the carotid artery work in progress. Journal of
Ultrasound in Medicine, 23(5): 659-669. (2004).
46 Stoitsis, J., Golemati, S., Kendros, S., and Nikita, K. 2008. Automated detection of the
carotid artery wall in b-mode ultrasound images using active contours initialized by the
hough transform. Engineering in Medicine and Biology Society, 2008. EMBS 2008. 30th
Annual International Conference of the IEEE. p. 3146-3149
47 Molinari, F., Zeng, G., and Suri, J. S. An integrated approach to computer-based
automated tracing and its validation for 200 common carotid arterial wall ultrasound
images a new technique. Journal of Ultrasound in Medicine, 29(3): 399-418. (2010).
48 Delsanto, S., Molinari, F., Giustetto, P., Liboni, W., and Badalamenti, S. 2006. Culex-
completely user-independent layers extraction: Ultrasonic carotid artery images
segmentation. Engineering in Medicine and Biology Society, 2005. IEEE-EMBS 2005.
133
27th Annual International Conference of the. p. 6468-6471
49 Molinari, F., Pattichis, C. S., Zeng, G., Saba, L., Acharya, U. R., Sanfilippo, R.,
Nicolaides, A., and Suri, J. S. Completely automated multiresolution edge snapper—a
new technique for an accurate carotid ultrasound imt measurement: Clinical validation
and benchmarking on a multi-institutional database. Image Processing, IEEE
Transactions on, 21(3): 1211-1222. (2012).
50 Molinari, F., Zeng, G., and Suri, J. S. 2010. Inter-greedy technique for fusion of different
carotid segmentation boundaries leading to high-performance imt measurement.
Engineering in Medicine and Biology Society (EMBC), 2010 Annual International
Conference of the IEEE. p. 4769-4772
51 Molinari, F., Meiburger, K. M., Acharya, U. R., Zeng, G., Rodrigues, P., Saba, L.,
Nicolaides, A., and Suri, J. 2011. Cares 3.0: A two stage system combining feature-based
recognition and edge-based segmentation for cimt measurement on a multi-institutional
ultrasound database of 300 images. Engineering in Medicine and Biology Society,
EMBC, 2011 Annual International Conference of the IEEE. p. 5149-5152
52 Mao, F., Gill, J., Downey, D., and Fenster, A. Segmentation of carotid artery in
ultrasound images: Method development and evaluation technique. Medical Physics,
27(8): 1961-1970. (2000).
53 Abolmaesumi, P., Sirouspour, M. R., and Salcudean, S. 2000. Real-time extraction of
carotid artery contours from ultrasound images. Computer-Based Medical Systems,
2000. CBMS 2000. Proceedings. 13th IEEE Symposium on. p. 181-186
54 Gill, J., Ladak, H., Steinman, D., and Fenster, A. 2000. Segmentation of ulcerated
plaque: A semi-automatic method for tracking the progression of carotid atherosclerosis.
Engineering in Medicine and Biology Society, 2000. Proceedings of the 22nd Annual
International Conference of the IEEE. p. 669-672
55 Cohen, L. D. On active contour models and balloons. CVGIP: Image understanding,
53(2): 211-218. (1991).
134
56 Hamou, A. K., and El-Sakka, M. R. 2004. A novel segmentation technique for carotid
ultrasound images. Acoustics, Speech, and Signal Processing, 2004.
Proceedings.(ICASSP'04). IEEE International Conference on. p. iii-521-524 vol. 523
57 Abdel-Dayem, A. R., and Ei-Sakka, M. 2004. A novel morphological-based carotid
artery contour extraction. Electrical and Computer Engineering, 2004. Canadian
Conference on. p. 1873-1876
58 Weszka, J. S., and Rosenfeld, A. A comparative study of texture measures for terrain
classification. NASA STI/Recon Technical Report N, 76: 13470. (1975).
59 Amadasun, M., and King, R. Textural features corresponding to textural properties.
Systems, Man and Cybernetics, IEEE Transactions on, 19(5): 1264-1274. (1989).
60 Laws, K. I. 1980. Rapid texture identification. 24th Annual Technical Symposium. p.
376-381
61 Wu, C.-M., Chen, Y.-C., and Hsieh, K.-S. Texture features for classification of ultrasonic
liver images. Medical Imaging, IEEE Transactions on, 11(2): 141-152. (1992).
62 Mandelbrot, B. B. The fractal geometry of nature/revised and enlarged edition. New
York, WH Freeman and Co., 1983, 495 p., 1. (1983).
63 Robinson, D., and Foulds, L. R. Comparison of phylogenetic trees. Mathematical
Biosciences, 53(1): 131-147. (1981).
64 Floßmann, E., Schulz, U. G., and Rothwell, P. M. Systematic review of methods and
results of studies of the genetic epidemiology of ischemic stroke. Stroke, 35(1): 212-227.
(2004).
65 Jerrard-Dunne, P., Cloud, G., Hassan, A., and Markus, H. S. Evaluating the genetic
component of ischemic stroke subtypes a family history study. Stroke, 34(6): 1364-1369.
(2003).
66 Matarin, M., Brown, W. M., Singleton, A., Hardy, J. A., and Meschia, J. F. Whole
135
genome analyses suggest ischemic stroke and heart disease share an association with
polymorphisms on chromosome 9p21. Stroke, 39(5): 1586-1589. (2008).
67 Li, J., Lang, J., Zeng, Z., and McCullough, L. D. Akt1 gene deletion and stroke. Journal
of the neurological sciences, 269(1): 105-112. (2008).
68 Fitch, W. M., and Margoliash, E. Construction of phylogenetic trees. Science,
155(3760): 279-284. (1967).
69 Sacco, R. L. Risk factors and outcomes for ischemic stroke. Neurology, 45(2 Suppl 1):
S10-14. (1995).
70 Tang, J., and Moret, B. M. Scaling up accurate phylogenetic reconstruction from gene-
order data. Bioinformatics, 19(suppl 1): i305-i312. (2003).
71 Nei, M., Tajima, F., and Tateno, Y. Accuracy of estimated phylogenetic trees from
molecular data. Journal of Molecular Evolution, 19(2): 153-170. (1983).
72 Rambaut, A., and Grass, N. C. Seq-gen: An application for the monte carlo simulation of
DNA sequence evolution along phylogenetic trees. Computer applications in the
biosciences: CABIOS, 13(3): 235-238. (1997).
73 Parry-Smith, D., and Attwood, T. K. Somap: A novel interactive approach to multiple
protein sequences alignment. Computer applications in the biosciences: CABIOS, 7(2):
233-235. (1991).
74 Milinkovitch, M. C., LeDuc, R. G., Adachi, J., Farnir, F., Georges, M., and Hasegawa,
M. Effects of character weighting and species sampling on phylogeny reconstruction: A
case study based on DNA sequence data in cetaceans. Genetics, 144(4): 1817. (1996).
75 Baxevanis, A. D. The importance of biological databases in biological discovery.
Current Protocols in Bioinformatics: 1.1. 1-1.1. 6. (2006).
76 Sigaud, O., and Wilson, S. W. Learning classifier systems: A survey. Soft Computing,
11(11): 1065-1078. (2007).
136
77 Kotsiantis, S. B., Zaharakis, I., and Pintelas, P.: ‘Supervised machine learning: A review
of classification techniques’, in Editor (Ed.)^(Eds.): ‘Book Supervised machine learning:
A review of classification techniques’ (2007, edn.), pp.
78 Zhu, X. Semi-supervised learning literature survey. (2005).
79 Tsiaparas, N., Golemati, S., Andreadis, I., Stoitsis, J., Valavanis, I., and Nikita, K.
Assessment of carotid atherosclerosis from b-mode ultrasound images using directional
multiscale texture features. Measurement Science and Technology, 23(11): 114004.
(2012).
80 Stoitsis, J., Golemati, S., Nikita, K., and Nicolaides, A. 2004. Characterization of carotid
atherosclerosis based on motion and texture features and clustering using fuzzy c-means.
Engineering in Medicine and Biology Society, 2004. IEMBS'04. 26th Annual
International Conference of the IEEE. p. 1407-1410
81 Santhiyakumari, N., Rajendran, P., and Madheswaran, M. Medical decision-making
system of ultrasound carotid artery intima–media thickness using neural networks.
Journal of digital imaging, 24(6): 1112-1125. (2011).
82 Kemény, V., Droste, D. W., Hermes, S., Nabavi, D. G., Schulte-Altedorneburg, G.,
Siebler, M., and Ringelstein, E. B. Automatic embolus detection by a neural network.
Stroke, 30(4): 807-810. (1999).
83 Kyriacou, E., Pattichis, M. S., Pattichis, C. S., Mavrommatis, A., Christodoulou, C. I.,
Kakkos, S., and Nicolaides, A. Classification of atherosclerotic carotid plaques using
morphological analysis on ultrasound images. Applied Intelligence, 30(1): 3-23. (2009).
84 Christodoulou, C. I., Pattichis, C. S., Pantziaris, M., and Nicolaides, A. Texture-based
classification of atherosclerotic carotid plaques. Medical Imaging, IEEE Transactions
on, 22(7): 902-912. (2003).
85 Kyriacou, E., Pattichis, C. S., Pattichis, M. S., Mavrommatis, A., Panagiotou, S.,
Christodoulou, C. I., Kakkos, S., and Nicolaides, A. Classification of atherosclerotic
carotid plaques using gray level morphological analysis on ultrasound images Artificial
137
intelligence applications and innovations (pp. 737-744): Springer.(2006).
86 Kyriacou, E., Pattichis, C., Christodoulou, C., Karaolis, M., and Nicolaides, A. 2005. An
integrated teleconsultation system for the evaluation of the risk of stroke. Proceedings of
the 5th International Network Conference INC2005. p.
87 Kyriacou, E. C., Pattichis, C. S., Karaolis, M. A., Loizou, C. P., Christodoulou, C. I.,
Pattichis, M. S., Kakkos, S., and Nicolaides, A. An integrated system for assessing stroke
risk. Engineering in Medicine and Biology Magazine, IEEE, 26(5): 43-50. (2007).
88 Abdolmaleki, P., Mokhtari-Dizaji, M., Montazeri, M., Saberi, H., and Nikanjam, N.
2005. Applying the logistic regression model to predict the carotid artery stenosis using
the sequential color doppler ultrasound image processing. Proceedings of Third Iranian
Conference on Machine Vision and Image Processing and Applications (MVIP), Tehran,
Iran. p. 312-317
89 Lambrou, A., Papadopoulos, H., Kyriacou, E., Pattichis, C. S., Pattichis, M. S.,
Gammerman, A., and Nicolaides, A. Evaluation of the risk of stroke with confidence
predictions based on ultrasound carotid image analysis. International Journal on
Artificial Intelligence Tools, 21(04). (2012).
90 Christodoulou, C., Pattichis, C., Kyriacou, E., and Nicolaides, A. Image retrieval and
classification of carotid plaque ultrasound images. The Open Cardiovascular Imaging
Journal, 2: 18-28. (2010).
91 French-Sherry, E., Bassiouny, H., Pooley, T., Marcoux, M., Castilla, M., Watson, D., and
Lozanski, L. Assessing stroke risk with carotid duplex ultrasound scanning. JOURNAL
OF CRITICAL ILLNESS, 13: 448-460. (1998).
92 Perez, I. Appropriateness of carotid duplex ultrasound in acute stroke management. PhD,
King's College, London. (1999).
93 Chalela, J. A., Kidwell, C. S., Nentwich, L. M., Luby, M., Butman, J. A., Demchuk, A.
M., Hill, M. D., Patronas, N., Latour, L., and Warach, S. Magnetic resonance imaging
and computed tomography in emergency assessment of patients with suspected acute
138
stroke: A prospective comparison. Lancet, 369(9558): 293-298. (2007).
94 Amin-Hanjani, S., Du, X., Zhao, M., Walsh, K., Malisch, T. W., and Charbel, F. T. Use of
quantitative magnetic resonance angiography to stratify stroke risk in symptomatic
vertebrobasilar disease. Stroke, 36(6): 1140-1145. (2005).
95 Lee, L. J., Kidwell, C. S., Alger, J., Starkman, S., and Saver, J. L. Impact on stroke
subtype diagnosis of early diffusion-weighted magnetic resonance imaging and magnetic
resonance angiography. Stroke, 31(5): 1081-1089. (2000).
96 Zanette, E. M., Fieschi, C., Bozzao, L., Roberti, C., Toni, D., Argentino, C., and Lenzi,
G. L. Comparison of cerebral angiography and transcranial doppler sonography in acute
stroke. Stroke, 20(7): 899-903. (1989).
97 Hassan, A. E., Rostambeigi, N., Chaudhry, S. A., Khan, A. A., Zacharatos, H., Khatri, R.,
Uzun, G., and Qureshi, A. I. Combination of noninvasive neurovascular imaging
modalities in stroke patients: Patterns of use and impact on need for digital subtraction
angiography. Journal of Stroke and Cerebrovascular Diseases, 22(7): e53-e58. (2013).
98 Nighoghossian, N., Derex, L., and Douek, P. The vulnerable carotid artery plaque:
Current imaging methods and new perspectives. Stroke, 36(12): 2764-2772. (2005).
99 Lorenz, M. W., Markus, H. S., Bots, M. L., Rosvall, M., and Sitzer, M. Prediction of
clinical cardiovascular events with carotid intima-media thickness a systematic review
and meta-analysis. Circulation, 115(4): 459-467. (2007).
100 van der Meer, I. M., Bots, M. L., Hofman, A., del Sol, A. I., van der Kuip, D. A., and
Witteman, J. C. Predictive value of noninvasive measures of atherosclerosis for incident
myocardial infarction the rotterdam study. Circulation, 109(9): 1089-1094. (2004).
101 E-health laboratory, computer science department. (2015) Retrieved 26 March 2015,
2015, from http://www.medinfo.cs.ucy.ac.cy/
102 Aja-Fernández, S., and Alberola-López, C. On the estimation of the coefficient of
variation for anisotropic diffusion speckle filtering. Image Processing, IEEE
139
Transactions on, 15(9): 2694-2701. (2006).
103 Maisonobe, L. Finding the circle that best fits a set of points. October 25th. (2007).
104 Rocha, R., Campilho, A., Silva, J., Azevedo, E., and Santos, R. Segmentation of
ultrasound images of the carotid using ransac and cubic splines. Computer methods and
programs in biomedicine, 101(1): 94-106. (2011).
105 Xu, C., and Prince, J. L. Snakes, shapes, and gradient vector flow. Image Processing,
IEEE Transactions on, 7(3): 359-369. (1998).
106 Bland, J. M., and Altman, D. Statistical methods for assessing agreement between two
methods of clinical measurement. The lancet, 327(8476): 307-310. (1986).
107 Loizou, C. P., Pattichis, C. S., Nicolaides, A. N., and Pantziaris, M. Manual and
automated media and intima thickness measurements of the common carotid artery.
Ultrasonics, Ferroelectrics, and Frequency Control, IEEE Transactions on, 56(5): 983-
994. (2009).
108 Molinari, F., Acharya, R. U., Zeng, G., Meiburger, K. M., Rodrigues, P. S., Saba, L., and
Suri, J. S. Cares 2.0: Completely automated robust edge snapper for cimt measurement in
300 ultrasound images—a two stage paradigm. Journal of Medical Imaging and Health
Informatics, 1(2): 150-163. (2011).
109 Menchón-Lara, R.-M., Bastida-Jumilla, M.-C., Morales-Sánchez, J., and Sancho-Gómez,
J.-L. Automatic detection of the intima-media thickness in ultrasound images of the
common carotid artery using neural networks. Medical & biological engineering &
computing, 52(2): 169-181. (2014).
110 Bush, W. S., and Moore, J. H. Genome-wide association studies. PLoS computational
biology, 8(12): e1002822. (2012).
111 Hirschhorn, J. N., and Daly, M. J. Genome-wide association studies for common diseases
and complex traits. Nature Reviews Genetics, 6(2): 95-108. (2005).
140
112 Traylor, M., Farrall, M., Holliday, E. G., Sudlow, C., Hopewell, J. C., Cheng, Y.-C.,
Fornage, M., Ikram, M. A., Malik, R., and Bevan, S. Genetic risk factors for ischaemic
stroke and its subtypes (the metastroke collaboration): A meta-analysis of genome-wide
association studies. The Lancet Neurology, 11(11): 951-962. (2012).
113 Kato, N., Takeuchi, F., Tabara, Y., Kelly, T. N., Go, M. J., Sim, X., Tay, W. T., Chen, C.-
H., Zhang, Y., and Yamamoto, K. Meta-analysis of genome-wide association studies
identifies common variants associated with blood pressure variation in east asians.
Nature genetics, 43(6): 531-538. (2011).
114 Maurano, M. T., Humbert, R., Rynes, E., Thurman, R. E., Haugen, E., Wang, H.,
Reynolds, A. P., Sandstrom, R., Qu, H., and Brody, J. Systematic localization of common
disease-associated variation in regulatory DNA. Science, 337(6099): 1190-1195. (2012).
115 Schaub, M. A., Boyle, A. P., Kundaje, A., Batzoglou, S., and Snyder, M. Linking disease
associations with regulatory information in the human genome. Genome research, 22(9):
1748-1759. (2012).
116 Jeck, W. R., Siebold, A. P., and Sharpless, N. E. Review: A meta‐analysis of gwas and
age‐associated diseases. Aging cell, 11(5): 727-731. (2012).
117 Campo, M., Randhawa, A. K., Dunstan, S., Farrar, J., Caws, M., Bang, N. D., Lan, N. N.,
Chau, T. T. H., Horne, D. J., and Thuong, N. T. Common polymorphisms in the cd43
gene region are associated with tuberculosis disease and mortality. American journal of
respiratory cell and molecular biology(ja). (2014).
118 Steinthorsdottir, V., Thorleifsson, G., Sulem, P., Helgason, H., Grarup, N., Sigurdsson,
A., Helgadottir, H. T., Johannsdottir, H., Magnusson, O. T., and Gudjonsson, S. A.
Identification of low-frequency and rare sequence variants associated with elevated or
reduced risk of type 2 diabetes. Nature genetics, 46(3): 294-298. (2014).
119 Black, M. H., Wu, J., Takayanagi, M., Wang, N., Taylor, K. D., Haritunians, T., Trigo, E.,
Lawrence, J. M., Watanabe, R. M., and Buchanan, T. A. Variation in pparg is associated
with longitudinal change in insulin resistance in mexican americans at risk for type 2
141
diabetes. The Journal of Clinical Endocrinology & Metabolism. (2015).
120 Dayeh, T. A., Olsson, A. H., Volkov, P., Almgren, P., Rönn, T., and Ling, C. Identification
of cpg-snps associated with type 2 diabetes and differential DNA methylation in human
pancreatic islets. Diabetologia, 56(5): 1036-1046. (2013).
121 de Miguel-Yanes, J. M., Shrader, P., Pencina, M. J., Fox, C. S., Manning, A. K., Grant, R.
W., Dupuis, J., Florez, J. C., D'Agostino, R. B., and Cupples, L. A. Genetic risk
reclassification for type 2 diabetes by age below or above 50 years using 40 type 2
diabetes risk single nucleotide polymorphisms. Diabetes Care, 34(1): 121-125. (2011).
122 Michailidou, K., Hall, P., Gonzalez-Neira, A., Ghoussaini, M., Dennis, J., Milne, R. L.,
Schmidt, M. K., Chang-Claude, J., Bojesen, S. E., and Bolla, M. K. Large-scale
genotyping identifies 41 new loci associated with breast cancer risk. Nature genetics,
45(4): 353-361. (2013).
123 Cowper-Sal, R., Zhang, X., Wright, J. B., Bailey, S. D., Cole, M. D., Eeckhoute, J.,
Moore, J. H., and Lupien, M. Breast cancer risk-associated snps modulate the affinity of
chromatin for foxa1 and alter gene expression. Nature genetics, 44(11): 1191-1198.
(2012).
124 Wu, C., Miao, X., Huang, L., Che, X., Jiang, G., Yu, D., Yang, X., Cao, G., Hu, Z., and
Zhou, Y. Genome-wide association study identifies five loci associated with
susceptibility to pancreatic cancer in chinese populations. Nature genetics, 44(1): 62-66.
(2012).
125 Spurdle, A. B., Thompson, D. J., Ahmed, S., Ferguson, K., Healey, C. S., O'Mara, T.,
Walker, L. C., Montgomery, S. B., Dermitzakis, E. T., and Fahey, P. Genome-wide
association study identifies a common variant associated with risk of endometrial cancer.
Nature genetics, 43(5): 451-454. (2011).
126 Hollingworth, P., Harold, D., Sims, R., Gerrish, A., Lambert, J.-C., Carrasquillo, M. M.,
Abraham, R., Hamshere, M. L., Pahwa, J. S., and Moskvina, V. Common variants at
abca7, ms4a6a/ms4a4e, epha1, cd33 and cd2ap are associated with alzheimer's disease.
142
Nature genetics, 43(5): 429-435. (2011).
127 Lambert, J.-C., Ibrahim-Verbaas, C. A., Harold, D., Naj, A. C., Sims, R., Bellenguez, C.,
Jun, G., DeStefano, A. L., Bis, J. C., and Beecham, G. W. Meta-analysis of 74,046
individuals identifies 11 new susceptibility loci for alzheimer's disease. Nature genetics,
45(12): 1452-1458. (2013).
128 Jonsson, T., Stefansson, H., Steinberg, S., Jonsdottir, I., Jonsson, P. V., Snaedal, J.,
Bjornsson, S., Huttenlocher, J., Levey, A. I., and Lah, J. J. Variant of trem2 associated
with the risk of alzheimer's disease. New England Journal of Medicine, 368(2): 107-116.
(2013).
129 Baron-Cohen, S., Murphy, L., Chakrabarti, B., Craig, I., Mallya, U., Lakatošová, S.,
Rehnstrom, K., Peltonen, L., Wheelwright, S., and Allison, C. A genome wide
association study of mathematical ability reveals an association at chromosome 3q29, a
locus associated with autism and learning difficulties: A preliminary study. PloS one,
9(5): e96374. (2014).
130 Gilman, S. R., Iossifov, I., Levy, D., Ronemus, M., Wigler, M., and Vitkup, D. Rare de
novo variants associated with autism implicate a large functional network of genes
involved in formation and function of synapses. Neuron, 70(5): 898-907. (2011).
131 Bevan, S., Traylor, M., Adib-Samii, P., Malik, R., Paul, N. L., Jackson, C., Farrall, M.,
Rothwell, P. M., Sudlow, C., and Dichgans, M. Genetic heritability of ischemic stroke
and the contribution of previously reported candidate gene and genomewide associations.
Stroke, 43(12): 3161-3167. (2012).
132 Holliday, E. G., Maguire, J. M., Evans, T.-J., Koblar, S. A., Jannes, J., Sturm, J. W.,
Hankey, G. J., Baker, R., Golledge, J., and Parsons, M. W. Common variants at 6p21. 1
are associated with large artery atherosclerotic stroke. Nature genetics, 44(10): 1147-
1151. (2012).
133 Yadav, S., Hasan, N., Marjot, T., Khan, M. S., Prasad, K., Bentley, P., and Sharma, P.
Detailed analysis of gene polymorphisms associated with ischemic stroke in south asians.
143
PloS one, 8(3): e57305. (2013).
134 Della-Morte, D., Guadagni, F., Palmirotta, R., Testa, G., Caso, V., Paciaroni, M., Abete,
P., Rengo, F., Ferroni, P., and Sacco, R. L. Genetics of ischemic stroke, stroke-related
risk factors, stroke precursors and treatments. Pharmacogenomics, 13(5): 595-613.
(2012).
135 Leu, H.-B., Chung, C.-M., Chuang, S.-Y., Bai, C.-H., Chen, J.-R., Chen, J.-W., and Pan,
W.-H. Genetic variants of connexin37 are associated with carotid intima-medial
thickness and future onset of ischemic stroke. Atherosclerosis, 214(1): 101-106. (2011).
136 Milne, R., and Antoniou, A. Genetic modifiers of cancer risk for brca1 and brca2
mutation carriers. Annals of oncology, 22(suppl 1): i11-i17. (2011).
137 Webster, J. A., Gibbs, J. R., Clarke, J., Ray, M., Zhang, W., Holmans, P., Rohrer, K.,
Zhao, A., Marlowe, L., and Kaleem, M. Genetic control of human brain transcript
expression in alzheimer disease. The American Journal of Human Genetics, 84(4): 445-
458. (2009).
138 MacInnis, R. J., Antoniou, A. C., Eeles, R. A., Severi, G., Olama, A., Amin, A.,
McGuffog, L., Kote‐Jarai, Z., Guy, M., and O'Brien, L. T. A risk prediction algorithm
based on family history and common genetic variants: Application to prostate cancer
with potential clinical impact. Genetic epidemiology, 35(6): 549-556. (2011).
139 Sazci, A., Ergul, E., Tuncer, N., Akpinar, G., and Kara, I. Methylenetetrahydrofolate
reductase gene polymorphisms are associated with ischemic and hemorrhagic stroke:
Dual effect of mthfr polymorphisms c677t and a1298c. Brain research bulletin, 71(1):
45-50. (2006).
140 Li, P., and Qin, C. Methylenetetrahydrofolate reductase (mthfr) gene polymorphisms and
susceptibility to ischemic stroke: A meta-analysis. Gene, 535(2): 359-364. (2014).
141 Zhou, B.-S., Bu, G.-Y., Li, M., Chang, B.-G., and Zhou, Y.-P. Tagging snps in the mthfr
gene and risk of ischemic stroke in a chinese population. International journal of
144
molecular sciences, 15(5): 8931-8940. (2014).
142 Somarajan, B., Kalita, J., Mittal, B., and Misra, U. Evaluation of mthfr c677t
polymorphism in ischemic and hemorrhagic stroke patients. A case–control study in a
northern indian population. Journal of the neurological sciences, 304(1): 67-70. (2011).
143 Ridker, P. M., Hennekens, C. H., Lindpaintner, K., Stampfer, M. J., Eisenberg, P. R., and
Miletich, J. P. Mutation in the gene coding for coagulation factor v and the risk of
myocardial infarction, stroke, and venous thrombosis in apparently healthy men. New
England Journal of Medicine, 332(14): 912-917. (1995).
144 Lynch, J. K., Nelson, K. B., Curry, C. J., and Grether, J. K. Cerebrovascular disorders in
children with the factor v leiden mutation. Journal of child neurology, 16(10): 735-744.
(2001).
145 Chabrier, S., Husson, B., Dinomais, M., Landrieu, P., and Tich, S. N. T. New insights
(and new interrogations) in perinatal arterial ischemic stroke. Thrombosis research,
127(1): 13-22. (2011).
146 Wu, L., Shen, Y., Liu, X., Ma, X., Xi, B., Mi, J., Lindpaintner, K., Tan, X., and Wang, X.
The 1425g/a snp in prkch is associated with ischemic stroke and cerebral hemorrhage in
a chinese population. Stroke, 40(9): 2973-2976. (2009).
147 Kubo, M., Hata, J., Ninomiya, T., Matsuda, K., Yonemoto, K., Nakano, T., Matsushita,
T., Yamazaki, K., Ohnishi, Y., and Saito, S. A nonsynonymous snp in prkch (protein
kinase c η) increases the risk of cerebral infarction. Nature genetics, 39(2): 212-217.
(2007).
148 van der Bom, J. G., Bots, M. L., Haverkate, F., Slagboom, P. E., Meijer, P., de Jong, P. T.,
Hofman, A., Grobbee, D. E., and Kluft, C. Reduced response to activated protein c is
associated with increased risk for cerebrovascular disease. Annals of internal medicine,
125(4): 265-269. (1996).
149 Simioni, P., De Ronde, H., Prandoni, P., Saladini, M., Bertina, R., and Girolami, A.
Ischemic stroke in young patients with activated protein c resistance a report of three
145
cases belonging to three different kindreds. Stroke, 26(5): 885-890. (1995).
150 Munshi, A., Sharma, V., Kaul, S., Rajeshwar, K., Babu, M. S., Shafi, G., Anila, A.,
Balakrishna, N., Alladi, S., and Jyothy, A. Association of the− 344c/t aldosterone
synthase (cyp11b2) gene variant with hypertension and stroke. Journal of the
neurological sciences, 296(1): 34-38. (2010).
151 Saidi, S., Mahjoub, T., and Almawi, W. Y. Aldosterone synthase gene (cyp11b2)
promoter polymorphism as a risk factor for ischaemic stroke in tunisian arabs. Journal
of Renin-Angiotensin-Aldosterone System. (2010).
152 Kessler, C., Spitzer, C., Stauske, D., Mende, S., Stadlmüller, J., Walther, R., and Rettig,
R. The apolipoprotein e and β-fibrinogen g/a-455 gene polymorphisms are associated
with ischemic stroke involving large-vessel disease. Arteriosclerosis, thrombosis, and
vascular biology, 17(11): 2880-2884. (1997).
153 Peng, D.-Q., Zhao, S.-P., and Wang, J.-L. Lipoprotein (a) and apolipoprotein e ε4 as
independent risk factors for ischemic stroke. European Journal of Cardiovascular Risk,
6(1): 1-6. (1999).
154 Couderc, R., Mahieux, F., Bailleul, S., Fenelon, G., Mary, R., and Fermanian, J.
Prevalence of apolipoprotein e phenotypes in ischemic cerebrovascular disease. A case-
control study. Stroke, 24(5): 661-664. (1993).
155 Ikram, M. A., Seshadri, S., Bis, J. C., Fornage, M., DeStefano, A. L., Aulchenko, Y. S.,
Debette, S., Lumley, T., Folsom, A. R., and van den Herik, E. G. Genomewide
association studies of stroke. New England Journal of Medicine, 360(17): 1718-1728.
(2009).
156 Tong, Y., Zhang, Y., Zhang, R., Geng, Y., Lin, L., Wang, Z., Liu, J., Li, X., Cao, Z., and
Xu, J. Association between two key snps on chromosome 12p13 and ischemic stroke in
chinese han population. Pharmacogenetics and genomics, 21(9): 572-578. (2011).
157 Gschwendtner, A., Bevan, S., Cole, J. W., Plourde, A., Matarin, M., Ross‐Adams, H.,
146
Meitinger, T., Wichmann, E., Mitchell, B. D., and Furie, K. Sequence variants on
chromosome 9p21. 3 confer risk for atherosclerotic stroke. Annals of neurology, 65(5):
531-539. (2009).
158 Anderson, C. D., Biffi, A., Rost, N. S., Cortellini, L., Furie, K. L., and Rosand, J.
Chromosome 9p21 in ischemic stroke population structure and meta-analysis. Stroke,
41(6): 1123-1131. (2010).
159 Smith, J. G., Melander, O., Lövkvist, H., Hedblad, B., Engström, G., Nilsson, P., Carlson,
J., Berglund, G., Norrving, B., and Lindgren, A. Common genetic variants on
chromosome 9p21 confers risk of ischemic stroke a large-scale genetic association study.
Circulation: Cardiovascular Genetics, 2(2): 159-164. (2009).
160 Meschia, J. F., Brott, T. G., Brown, R. D., Crook, R. J., Frankel, M., Hardy, J., Merino, J.
G., Rich, S. S., Silliman, S., and Worrall, B. B. The ischemic stroke genetics study (isgs)
protocol. BMC neurology, 3(1): 4. (2003).
161 Heiss, G., Sharrett, A. R., Barnes, R., Chambless, L., Szklo, M., and Alzola, C. Carotid
atherosclerosis measured by b-mode ultrasound in populations: Associations with
cardiovascular risk factors in the aric study. American journal of epidemiology, 134(3):
250-256. (1991).
162 Hofman, A., Breteler, M. M., van Duijn, C. M., Krestin, G. P., Pols, H. A., Stricker, B. H.
C., Tiemeier, H., Uitterlinden, A. G., Vingerling, J. R., and Witteman, J. C. The rotterdam
study: Objectives and design update. European journal of epidemiology, 22(11): 819-
829. (2007).
163 Bellenguez, C., Bevan, S., Gschwendtner, A., Spencer, C., Burgess, A., Pirinen, M.,
Jackson, C., Traylor, M., Strange, A., and Su, Z. International stroke genetics consortium
(isgc); wellcome trust case control consortium 2 (wtccc2). Genome-wide association
study identifies a variant in hdac9 associated with large vessel ischemic stroke. Nat
Genet, 44(3): 328-333. (2012).
164 Goldszmidt, M. Bayesian network classifiers. Wiley Encyclopedia of Operations
147
Research and Management Science. (2010).
165 Rish, I. 2001. An empirical study of the naive bayes classifier. IJCAI 2001 workshop on
empirical methods in artificial intelligence. p. 41-46
166 Murphy, K. P. Naive bayes classifiers. University of British Columbia. (2006).
167 Ali, S., and Smith, K. A. On learning algorithm selection for classification. Applied Soft
Computing, 6(2): 119-138. (2006).
168 Freund, Y., and Schapire, R. E. 1996. Experiments with a new boosting algorithm.
ICML. p. 148-156
169 Thornton, C., Hutter, F., Hoos, H. H., and Leyton-Brown, K. 2013. Auto-weka:
Combined selection and hyperparameter optimization of classification algorithms.
Proceedings of the 19th ACM SIGKDD international conference on Knowledge
discovery and data mining. p. 847-855
170 Quinlan, J. R. 1996. Bagging, boosting, and c4. 5. AAAI/IAAI, Vol. 1. p. 725-730
171 Breiman, L. Random forests. Machine learning, 45(1): 5-32. (2001).
172 Breiman, L. Bagging predictors. Machine learning, 24(2): 123-140. (1996).
173 Ruck, D. W., Rogers, S. K., Kabrisky, M., Oxley, M. E., and Suter, B. W. The multilayer
perceptron as an approximation to a bayes optimal discriminant function. Neural
Networks, IEEE Transactions on, 1(4): 296-298. (1990).
174 Sneath, P., and Sokal, R. Unweighted pair group method with arithmetic mean.
Numerical Taxonomy: 230-234. (1973).
175 Saitou, N., and Nei, M. The neighbor-joining method: A new method for reconstructing
phylogenetic trees. Molecular biology and evolution, 4(4): 406-425. (1987).
176 Eck, R., and Dayhoff, M. Atlas of protein sequence and structure national biomedical
research foundation. Springs Silver, Maryland. (1966).
148
177 Fitch, W. M. Toward defining the course of evolution: Minimum change for a specific
tree topology. Systematic Biology, 20(4): 406-416. (1971).
178 Felsenstein, J. Evolutionary trees from DNA sequences: A maximum likelihood
approach. Journal of Molecular Evolution, 17(6): 368-376. (1981).
179 Felsenstein, J. Evolutionary trees from gene frequencies and quantitative characters:
Finding maximum likelihood estimates. Evolution: 1229-1242. (1981).
180 Larkin, M. A., Blackshields, G., Brown, N., Chenna, R., McGettigan, P. A., McWilliam,
H., Valentin, F., Wallace, I. M., Wilm, A., and Lopez, R. Clustal w and clustal x version
2.0. Bioinformatics, 23(21): 2947-2948. (2007).
181 Tamura, K., Stecher, G., Peterson, D., Filipski, A., and Kumar, S. Mega6: Molecular
evolutionary genetics analysis version 6.0. Molecular biology and evolution, 30(12):
2725-2729. (2013).
182 Swofford, D. L. Phylogenetic analysis using parsimony. Champaign, IL: Illinois Natural
History Survey. (1993).
183 Swofford, D. L. Paup 4.0 for macintosh: Phylogenetic analysis using parsimony
(software and user's book for macintosh). Sinauer Associates, Incorporated. pp. (2004).
184 Felsenstein, J.: ‘Phylogeny inference package’, in Editor (Ed.)^(Eds.): ‘Book Phylogeny
inference package’ (Version, 2006, edn.), pp.
185 Bionumerics seven. (2011) Retrieved 02/12/2014, 2014, from http://www.applied-
maths.com/bionumerics
186 Kanz, C., Aldebert, P., Althorpe, N., Baker, W., Baldwin, A., Bates, K., Browne, P., van
den Broek, A., Castro, M., and Cochrane, G. The embl nucleotide sequence database.
Nucleic Acids Research, 33(suppl 1): D29-D33. (2005).
187 MacDonald, J. R., Ziman, R., Yuen, R. K., Feuk, L., and Scherer, S. W. The database of
genomic variants: A curated collection of structural variation in the human genome.
149
Nucleic Acids Research, 42(D1): D986-D992. (2014).
188 Church, D. M., Lappalainen, I., Sneddon, T. P., Hinton, J., Maguire, M., Lopez, J.,
Garner, J., Paschall, J., DiCuccio, M., and Yaschenko, E. Public data archives for
genomic structural variation. Nature genetics, 42(10): 813-814. (2010).
189 Sherry, S. T., Ward, M.-H., Kholodov, M., Baker, J., Phan, L., Smigielski, E. M., and
Sirotkin, K. Dbsnp: The ncbi database of genetic variation. Nucleic Acids Research,
29(1): 308-311. (2001).
190 Thorisson, G. A., and Stein, L. D. The snp consortium website: Past, present and future.
Nucleic Acids Research, 31(1): 124-127. (2003).
191 Brookes, A. J., Lehväslaiho, H., Siegfried, M., Boehm, J. G., Yuan, Y. P., Sarkar, C. M.,
Bork, P., and Ortigao, F. Hgbase: A database of snps and other variations in and around
human genes. Nucleic Acids Research, 28(1): 356-360. (2000).
192 Gibbs, R. A., Belmont, J. W., Hardenbol, P., Willis, T. D., Yu, F., Yang, H., Ch'ang, L.-Y.,
Huang, W., Liu, B., and Shen, Y. The international hapmap project. Nature, 426(6968):
789-796. (2003).
193 Clarke, L., Zheng-Bradley, X., Smith, R., Kulesha, E., Xiao, C., Toneva, I., Vaughan, B.,
Preuss, D., Leinonen, R., and Shumway, M. The 1000 genomes project: Data
management and community access. Nature methods, 9(5): 459-462. (2012).
194 Consortium, I. C. G.: ‘Icgc cancer genome projects’, in Editor (Ed.)^(Eds.): ‘Book Icgc
cancer genome projects’ (2013, edn.), pp.
195 Bamford, S., Dawson, E., Forbes, S., Clements, J., Pettett, R., Dogan, A., Flanagan, A.,
Teague, J., Futreal, P. A., and Stratton, M. The cosmic (catalogue of somatic mutations in
cancer) database and website. British journal of cancer, 91(2): 355-358. (2004).
196 Stenson, P. D., Ball, E. V., Mort, M., Phillips, A. D., Shiel, J. A., Thomas, N. S.,
Abeysinghe, S., Krawczak, M., and Cooper, D. N. Human gene mutation database
(hgmd®): 2003 update. Human mutation, 21(6): 577-581. (2003).
150
197 Amberger, J., Bocchini, C. A., Scott, A. F., and Hamosh, A. Mckusick's online mendelian
inheritance in man (omim®). Nucleic Acids Research, 37(suppl 1): D793-D796. (2009).
198 Pagon, R. A. Genetests: An online genetic information resource for health care providers.
Journal of the Medical Library Association, 94(3): 343. (2006).
199 Mitelman, F., Johansson, B., and Mertens, F. Mitelman database of chromosome
aberrations in cancer. Cancer Genome Anatomy Project. pp. (2007).
200 Becker, K. G., Barnes, K. C., Bright, T. J., and Wang, S. A. The genetic association
database. Nature genetics, 36(5): 431-432. (2004).
201 Altshuler, D. M., Gibbs, R. A., Peltonen, L., Dermitzakis, E., Schaffner, S. F., Yu, F. L.,
Bonnen, P. E., de Bakker, P. I. W., Deloukas, P., Gabriel, S. B., Gwilliam, R., Hunt, S.,
Inouye, M., Jia, X. M., Palotie, A., Parkin, M., Whittaker, P., Chang, K., Hawes, A.,
Lewis, L. R., Ren, Y. R., Wheeler, D., Muzny, D. M., Barnes, C., Darvishi, K., Hurles,
M., Korn, J. M., Kristiansson, K., Lee, C., McCarroll, S. A., Nemesh, J., Keinan, A.,
Montgomery, S. B., Pollack, S., Price, A. L., Soranzo, N., Gonzaga-Jauregui, C., Anttila,
V., Brodeur, W., Daly, M. J., Leslie, S., McVean, G., Moutsianas, L., Nguyen, H., Zhang,
Q. R., Ghori, M. J. R., McGinnis, R., McLaren, W., Takeuchi, F., Grossman, S. R.,
Shlyakhter, I., Hostetter, E. B., Sabeti, P. C., Adebamowo, C. A., Foster, M. W., Gordon,
D. R., Licinio, J., Manca, M. C., Marshall, P. A., Matsuda, I., Ngare, D., Wang, V. O.,
Reddy, D., Rotimi, C. N., Royal, C. D., Sharp, R. R., Zeng, C. Q., Brooks, L. D.,
McEwen, J. E., and Int HapMap, C. Integrating common and rare genetic variation in
diverse human populations. Nature, 467(7311): 52-58. (2010).
viii
Plagiarism Report
ix
List of Publications and Reprints
1 Farhan, S., Fahiem, M. A., andTauseef, H. An ensemble-of-classifiers based approach for
early diagnosis of alzheimer’s disease: Classification using structural features of brain
images. Computational and mathematical methods in medicine, 2014. (2014).
2 Tauseef, H., Fahiem, M. A., Farhan, S., andTahir, F. A review of image and phylogenetic
analysis based techniques for ischemic stroke risk estimation. Life Science Journal,
10(7s). (2013).
3 Farhan, S., Fahiem, M. A., Tahir, F., andTauseef, H. A comparative study of
neuroimaging and pattern recognition techniques for estimation of alzheimer’s. Life
Science Journal, 10(7s). (2013).
4 Tahir, F., Fahiem, M. A., Tauseef, H., andFarhan, S. A survey of multispectral high
resolution imaging based drug surface morphology validation techniques. Life Science
Journal, 10(7): 1050-1059. (2013).
5 Aftab, Z., andTuaseef, H. Enhancing pixel oriented visualization by merging circle view
and circle segment visualization techniques Multi-disciplinary trends in artificial
intelligence (pp. 101-109): Springer.(2012).
6 Tauseef, H., Fahiem, M. A., andFarhan, S. 2009. Recognition and translation of hand
gestures to urdu alphabets using a geometrical classification. Visualisation, 2009. VIZ'09.
Second International Conference in. p. 213-217
7 Tauseef, H., Farhan, S., andFahiem, M. A. 2009. A systematic approach for selecting a
suitable software architecture evaluation method. Software Engineering Research and
Practice. p. 295-299
8 Farhan, S., Fahiem, M. A., andTauseef, H. 2009. Geometrical features based approach
for the classification and recognition of handwritten characters. Visualisation, 2009.
VIZ'09. Second International Conference in. p. 185-190
9 Fahiem, M. A., Haq, S. A., Saleemi, F., andTauseef, H. 2009. 3d reconstruction:
x
Estimating depth of hole from 2d camera perspectives. Proceedings of the European
Computing Conference. p. 213-221
10 Farhan, S., Tauseef, H., andFahiem, M. A. 2009. Adding agility to architecture tradeoff
analysis method for mapping on crystal. Software Engineering, 2009. WCSE'09. WRI
World Congress on. p. 121-125
11 Tauseef, H., Fahiem, M. A., andSaleemi, F. 2007. Target recognition task based online
system for refractive error measurement using font transformations. Proceedings of the
7th Conference on 7th WSEAS International Conference on Applied Computer Science-
Volume 7. p. 162-167
12 Tauseef, H., Fahiem, M. A., Farhan, S.; Image Based Sign Language Translation: Urdu
Text. LAP LAMBERT Academic Publishing, Germany, 2011, ISBN-13: 978-3-8454-
0969-6, ISBN-10: 384540969X.
13 Farhan, S., Fahiem, M. A., Tauseef, H.; Hand Written Character Recognition: Non-
Cursive Scripts. LAP LAMBERT Academic Publishing, Germany, 2011, ISBN-13: 978-
3-8454-0104-1, ISBN-10: 3845401044X.