Artificial Intelligence Techniques for Ocular Pattern Classi...

8/13/2019 Artificial Intelligence Techniques for Ocular Pattern Classi...

1/200

FacultyofComputersandInformationComputerScienceDepartment

ArtificialIntelligenceTechniquesforOcular

PatternClassification

Amr

Ahmed

Sabry

Abdel

Rahman

Ghoneim

Supervisedby:ProfessorAtefZakiGhalwash

ProfessorofComputerScience,TheFacultyofComputers&Information,Helwan

University,

Cairo,

Egypt.

AssociateProfessorAliaaAbdelHaleimAbdelRazikYoussif

AssociateProfessorofComputerScience,TheFacultyOfComputers&Information,HelwanUniversity,Cairo,Egypt.

AssistantProfessorHosamElDeanMahmoudBakeir

AssistantProfessorofOphthalmology,TheFacultyofMedicine,Cairo

University,

Cairo,

Egypt.

A thesissubmitted toHelwanUniversity inaccordancewith therequirements forthedegreeofMasterofScienceinComputerScienceattheFacultyofComputers&Information,DepartmentofComputerScience.

May2007


2/200

References2

Thispageisleftblankintentionally


3/200

)285(

))286

InthenameofAllah,MostGracious,MostMercifulTheMessengerbelieveth inwhathathbeenrevealed tohimfromhisLord,asdo

themenoffaith.Eachone(ofthem)believethinAllah,Hisangels,HisBooks,and

His Messengers. Wemake nodistinction (they say)between one andanotherof

HisMessengers. Andtheysay: Wehear,andweobey,(weseek)Thyforgiveness,

ourLord,andtoTheeistheendofalljourneys. (285)OnnosouldothAllahplace

aburden greater than it canbear. It gets every good that it earns, and it suffers

everyill

that

it

earns.

(Pray:) Our

Lord!

Condemn

us

not

if

we

forget

or

fall

into

error;ourLord!Laynotonusaburden like thatwhichThoudidst layon those

beforeus;ourLord!laynotonusaburdengreaterthanwehavestrengthtobear.

Blot out our sins, and grant us forgiveness. Have mercy on us. Thou art our

Protector;helpusagainstthosewhostandagainstFaith.(286)TheholyQuran:Chapter2AlBaqarah285:286

: : .

AbuHuraira(Allahbepleasedwithhim)reportedAllahsMessenger(Maypeaceandblessingsbeuponhim)assaying:Whenamandies,hisactscometoanend,butthree,recurringcharity,orknowledge(bywhichpeople)benefit,orapiousson,whopraysforhim(forthedeceased).

SahihMuslim


4/200

References2

Thispageisleftblankintentionally


5/200

To

thememoryofmycolleagues

IbraheimArafat,

MustaphaGamal,

&

KhaledAbdelMoneim

Andtomybelovedcity,Cairo,acitythatnever

failsto

make

an

impression,

an

everlasting,

unique

impression


6/200

Theresearch

described

in

this

thesis

was

carried

out

at

the

Faculty

of

Computers

&

Information HelwanUniversity,Cairo,TheArabRepublicofEgypt.

Copyright 2007 by Amr S. Ghoneim. All rights reserved. No part of this

publication may be reproduced or transmitted in any form or by any means,

electronic or mechanical, including photocopy, recording, or any information

storageandretrievalsystem,withoutpermission inwriting from theauthor.Any

trademarksinthispublicationarepropertyoftheirrespectiveowners.


7/200

Contents vii

C o n t e n t s

Abstract & Keywords. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xiv

Acknowledgment. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xvii

Declaration. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xx

List of Publications. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xxi

List of Figures. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xxii

List of Tables. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xxviii

Awards . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xxix

Medical Glossary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xxx

1 Introduction 1

1.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

1.2 Objectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3

1.3 Thesis Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

1.4 Thesis Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

1.5 Eye Anatomy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6

1.6 Fundus Photography and Eye Diseases . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

1.6.1 Diabetic Retinopathies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10

1.6.2 Glaucoma . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11

1.6.3 Detecting Retina Landmarks . . . . . . . . . . . . . . . . . . . . . . . . . . . 12

1.6.4 Fundus Photography Datasets . . . . . . . . . . . . . . . . . . . . . . . . . . 12


8/200

Contents

viii

2 Preprocessing 13

2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13

2.2 Fundamentals of Retinal Digital Image Representation . . . . . . . . . . . . . . 13

2.3 Mask Generation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15

2.4 Illumination Equalization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16

2.5 Contrast Enhancement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18

2.5.1 Green Band Processing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18

2.5.2 Histogram Equalization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20

2.5.3 Local Contrast Enhancement . . . . . . . . . . . . . . . . . . . . . . . . . . 21

2.5.4 Adaptive Histogram Equalization . . . . . . . . . . . . . . . . . . . . . . 23

2.5.5 Background Subtraction of Retinal Blood Vessels (BSRBV)

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .24

2.5.6 Estimation of Background Luminosity and Contrast

Variability (EBLCV) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24

2.6 Color Normalization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25

2.6.1 Gray-World Normalization . . . . . . . . . . . . . . . . . . . . . . . . . . . 26

2.6.2 Comprehensive Normalization . . . . . . . . . . . . . . . . . . . . . . . . . 27

2.6.3 Histogram Equalization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29

2.6.4 Histogram Specification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29

2.7 Image Quality Assessment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31

2.7.1 Convolution of Global Intensity Histograms . . . . . . . . . . . . . . 32

2.7.2 Edge Magnitude and Local Pixel Intensity Distributions . . . . . 33

2.7.3 Asymmetry of Histograms Derived from Edge Maps . . . . . . . 34


9/200

Contents ix

2.7.4 Chromaticity Values Distribution . . . . . . . . . . . . . . . . . . . . . . 37

2.7.5 Clarity and Field Definition . . . . . . . . . . . . . . . . . . . . . . . . . . . 38

2.8 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42

3 Automatic Localization of the Optic Disc 43

3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43

3.2 Properties of the Optic Disc . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45

3.3 Automatic Optic Disc Localization: A Literature Review . . . . . . . . . . . . . 46

3.3.1 Optic Disc Localization versus Disc Boundary Detection . . . . 46

3.3.2 Existing Automatic OD-Localization Algorithms Review . . . . 46

3.3.3 Alternative OD-Localization Algorithms . . . . . . . . . . . . . . . . . 58

3.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59

4 Automatic Segmentation of the Retinal Vasculature 61

4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61

4.2 Properties of the Retinal Vasculature . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62

4.3 Automatic Retinal Vasculature Segmentation: A Literature Review . . . . 62

4.3.1 Detection of Blood Vessels in Retinal Images using TwoDimensional Matched Filters . . . . . . . . . . . . . . . . . . . . . . . . . .

63

4.3.2 Existing Automatic Retinal Vasculature Segmentation

Algorithms Review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

66

4.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74

5 Automatic Detection of Hard Exudates 75

5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75

5.2 Properties of the Diabetic Retinopathy Lesions . . . . . . . . . . . . . . . . . . . . 76


10/200

Contents

x

5.3 Automatic Diabetic Retinopathy Lesions Detection: A Literature Review

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .78

5.3.1 Existing Automatic Bright Lesions Detection Algorithms

Review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .79

5.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84

6 Comparative Studies 85

6.1 A Comparative Study of Mask Generation Methods . . . . . . . . . . . . . . . . . 85

6.1.1 Purpose . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85

6.1.2 Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85

6.1.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86

6.1.4 Observations and Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . 86

6.2 A Comparative Study of Illumination Equalization Methods . . . . . . . . . . 88

6.2.1 Purpose . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88

6.2.2 Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88

6.2.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89


6.3 A Comparative Study of Contrast Enhancement Methods . . . . . . . . . . . . 90

6.3.1 Purpose . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90

6.3.2 Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90

6.3.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91


11/200

Contents xi


6.4 A Comparative Study of Color Normalization Methods . . . . . . . . . . . . . . 95

6.4.1 Purpose . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95

6.4.2 Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95

6.4.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96


7 The Developed Automatic DR Screening System Components 100

7.1 Optic Disc Localization by Means of a Vessels Direction Matched Filter

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .100

7.1.1 Purpose . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100

7.1.2 Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101

7.1.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103


7.2 Retinal Vasculature Segmentation using a Large-Scale Support Vector

Machine . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .109

7.2.1 Purpose . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109

7.2.2 Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109

7.2.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112



12/200

Contents

xii

7.3 Hard Exudates Detection using a Large-Scale Support Vector Machine .

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .118

7.3.1 Purpose . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118

7.3.2 Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118

7.3.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121


8 Conclusions & Future Work 125

8.1 Preprocessing of Digital Retinal Fundus Images . . . . . . . . . . . . . . . . . . . . 125

8.2 Retinal Landmarks Classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126

8.3 Diabetic Retinopathies Classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128

Appendix A Eye-Related Images 129

A.1 Fundus Images . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129

A.2 Fluorescein Angiograms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130

A.3 Indocyanine Green Dye . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131

A.4 Hartmann-Shack Images . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131

A.5 Iris Images . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133

Appendix B Diabetes, Diabetic Retinopathy and their Prevalence in Egypt 135

B.1 Diabetes Mellitus . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135

B.2 Diabetes in Egypt . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136

B.3 Diabetic Retinopathy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138


13/200

Contents xiii

B.4 Diabetic Retinopathy in Egypt . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139

Appendix C Fundus Photography Datasets 142

C.1 DRIVE Database . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142

C.2 STARE Project Database . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143

Appendix D Color Models 147

D.1 HSI Color Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147

D.1.1 Converting Colors from RGB to HIS . . . . . . . . . . . . . . . . . . . . 148

D.1.2 Converting Colors from HSI to RGB . . . . . . . . . . . . . . . . . . . . 148

D.2 Chromaticity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149

Appendix E A Threshold Selection Method from Gray-Level Histograms 151

References 154

(Abstract & Keywords in Arabic) (Cover page in Arabic)


14/200

Abstract & Keywordsxiv

Abstract

Diabetes is a disease that affects about 5.5% of the global population. In

Egypt, nearly 9 million (over 13% of the population 20 years) will have

diabetes by the year 2025, while recent surveys from Oman and Pakistan

suggest that this may be a regional phenomenon. Consequently, about 10%

of all diabetic patients have diabetic retinopathy (DR); one of the most

prevalent complications of diabetes and which is the primary cause of

blindness in the Western World, and this is likely to be true in Hong Kong,

and Egypt. Moreover, diabetic population is expected to have a 25 times

greater risk of going blind than non-diabetic. Due to the growing number of

patients, and with insufficient ophthalmologists to screen them all, automatic

screening can reduce the threat of blindness by 50%, provide considerable

cost savings, and decrease the pressure on available infrastructures and

resources.

Retinal photography is significantly more effective than direct

ophthalmoscopy in detecting DR. Digital fundus images do not require

injecting the body by fluorescein or indocyanine green dye, thus not

requiring a trained personnel. Digital fundus images are routinely analyzed

by screening systems, and owing to the acquisition process, these images are

very often of poor quality that hinders further analysis. State-of-the-art

studies still struggle with the issue of preprocessing in retinal images, mainly

due to the lack of literature reviews and comparative studies. Furthermore,

available preprocessing methods are not being evaluated on large benchmark

publicly-available datasets.

The first part of this dissertation discusses four major preprocessing

methodologies described in literature (mask generation, illumination

equalization, contrast enhancement, and color normalization), and their

effect on detecting retinal anatomy. In each methodology, a comparative

performance measure based on proposed appropriate metrics is

accomplished among available methods, using two publicly available fundus

datasets. In addition, we proposed the comprehensive normalizationand a

local contrast enhancement proceeded by illumination equalizationwhich


15/200

Abstract & Keywordsxv

recorded acceptable results when applied for color normalization and

contrast enhancement respectively.

Detecting retinal landmarks (the vasculature, optic disc, and macula)

give a framework from which automatedanalysis and human interpretation

of the retina proceed.And therefore, it will highly aid the future detection

and hence quantification of diseases in the mentioned regions. In addition,

recognizing main components can be used for criteria that allow the

discarding of images that have a too bad quality for assessment of

retinopathy.

The second part of the dissertation deals with the Optic Disc (OD)detection as a main step while developing automated screening systems for

diabetic retinopathy. We present in this study a method to automatically

detect the position of the OD in digital retinal fundus images based on

matching the expected directional pattern of the retinal blood vessels. Hence,

a simple matched filter is proposed to roughly match the direction of the

vessels at the OD vicinity. The proposed method was evaluated using a

subset of the STARE project's dataset, containing 81fundus images of both

normal and diseased retinas, and initially used by literature OD detection

methods. The OD-center was detected correctly in 80 out of the 81 images(98.77%). In addition, the OD-center was detected correctly in all of the 40

images (100%) using the publicly available DRIVE dataset.

The third part of the dissertation deals with the Retinal vasculature

(RV) segmentation as another basic foundation while developing retinal

screening systems, since the RV acts as the main landmark for further

analysis. Recently, supervised classification proved to be more efficient and

accurate for the segmentation process. Moreover, novel features have been

used in literature methods, showing high separability between vessels/non-vessels classes. This work utilizes the large-scale support vector machine for

automatic segmentation of RV, using for the pixel features a mixture of the

2D-Gabor wavelet, Top-hat, and Hessian-based enhancements. The

presented method noticeably reduces the number of training pixels since

2000 instead of 1 million pixels, as presented in recent literature studies, are

only needed for training. As a result, the average training time drops to 3.75

seconds instead of the 9 hours that was previously recorded in literature. For

classifying an image, 30 seconds were only needed. Small training sets and


16/200

Abstract & Keywordsxvi

efficient training time are critical for systems that always need readjustment

and tuning with various datasets. The publicly available benchmark DRIVE

dataset was used for evaluating the performance of the presented method.Experiments reveal that the area under the receiver operating characteristic

curve (AUC) reached a value 0.9537 which is highly comparable to

previously reported AUCs that range from 0.7878 to 0.9614.

Finally, the fourth part of the presented work deals with the

automated detection of hard exudates, as a main manifestation of diabetic

retinopathies. Methods dealing with the detection of retinal bright-lesions in

general were reviewed. Then, a method based on the response of the large-

scale support vector machine for the RV segmentation was proposed. Themethod uses to a closed inverted version of the large-scale support vector

machine response to determine the potential exudates regions according to

the properties of each region. The response image was segmented into

regions using the watersheds algorithm. The proposed method achieved and

accuracy of 100% for detecting hard exudates, an accuracy of 90% for

indicating images not including any form of bright-lesions.

Keywords Medical Image Processing, Artificial Intelligence,

Preprocessing, Retinal/Fundus Image Analysis, Optic Disc Localization,

Vessels Segmentation, Exudates Detection, Diabetic Retinopathies

Detection.


17/200

Acknowledgment xvii

Acknowledgment

For their tireless support, I thankMom,Dad,my sisterGhada,and

mybrotherAyman. Iamsure thesuccessof thisworkwouldmake

themdelighted.

Formany years of priceless advice andwisdomwhich shapedmy

development as a researcher, I thankmy supervisors Prof.AtefZ.

Ghalwashand

Prof.

Aliaa

A.

Youssif.

Iowe

them

much

for

their

assistance in all aspects of my work, providing me great help on

varioustopics,exchangingideasonbothacademicandnonacademic

matters,andforawonderfulfriendship.Thankyouverymuch!

Special thanks are given to the consultant ophthalmologist Dr.

HosamElDeanBakeirforhiskindsupervisionofmedicalexpertise.

For taking the time tobelieve inmeand investing theirmindsand

hearts into teaching me how to be a better person, I thank my

teachers Lecturers & Assistants at the Faculty of Computers &

Information(FCI),HelwanUniversitywhohaveshapedmymindand

madeathoughtfulimpactonmydevelopmentovertheyears.Andof

them, I wish to express my gratitude to Dr. Mohamed A. Belal

(currently at the Faculty of Science and Information Technology, Al

Zaytoonah

University,Amman,

Jordan)

for

supporting

me

in

ComputationalIntelligence.AndIwishalsotoexpressmygratitude

toDr.WaleedA.YousefwhosupportedmeinLinearAlgebra.

IwouldliketothankmanyofmycolleaguestheTeachingAssistants

at both departments in my faculty for their support and

understanding.


18/200

Acknowledgmentxviii

I would like to deeply thank my true friends from all grades,

especiallymy

class

(the

2002

graduates),

and

both

the

2005

and

2006

graduates, always givingme support, encouragement, and helping

mekeepmyspiritsup.AndIshallneverforgetmystudents(hopefully

the2007and2008graduates) forbeingsopatient,understandingand

loving. I acknowledge thatwithout allmyFCIHelwan friends,my

lifeisempty.Youkeptmegoingonguys!!

I am also indebted to many researchers who supported me with

variousresources

during

all

stages

of

writing,

and

so

Special

thanks

aredueto:

Ayman S.Ghoneim; a TeachingAssistant at the Operations

Research Dept., the Faculty of Computers & Information

Cairo Univ. (now aM.Sc. student at the School of Information

Technology & Electrical Engineering (ITEE), Australian Defense

Force Academy (ADFA), the University of New South Wales

(UNSW),

Canberra,

Australia.)

Amany AbdelHaleim & Bassam Morsy; both are Teaching

Assistants at the Information Systems/Computer Science

Departments respectively, the Faculty of Computers &

Information Helwan Univ. (currently students at the

DepartmentofElectricalandComputerEngineering,Universityof

Victoria,Victoria,BritishColumbia,Canada.)

GhadaKhoriba&AymanEzzat;bothareTeachingAssistants

atthe

Computer

Science

Department,

the

Faculty

of

Computers & Information Helwan Univ. (currently PhD.

students at the Graduate School of Systems and Information

Engineering,TsukubaUniversity,Tsukuba,Japan.)

Mohamed Ali; a Teaching Assistant at the Bioengineering

Dept.,theFacultyofEngineering HelwanUniversity.


19/200

Acknowledgment xix

Theworkofmanyresearchers inthisfieldmade thisworkpossible,

butmydirectcontactswithsomeofthemwhokindlyansweredsome

technicalquestions

pertaining

to

their

work

have

influenced

my

thinkingagreatdeal,andsoIwishtoexpressmysincerethanksto:

Chanjira Sinthanayothin; the National Electronics and

ComputerTechnologyCentre(NECTEC),Thailand.

Subhasis Chaudhuri; Professor & Head, Dept. of Electrical

Engineering, Indian Institute of Technology (IIT), Powai,

Bombay,India.

SarahA. Barman; Senior Lecturer,Digital Imaging Research

Centre,Kingston

University,

London,

UK.

NormanKatz;CEO of the IPConsulting (a custom enterprise

softwarecompany),SanDiego,California,USA.

LangisGagnon;AssociateProfessor,DepartmentofComputer

andElectricalEngineering,UniversitLaval,Quebec,Canada.

Meindert Niemeijer; Ph.D. Student, Image Sciences Institute

(ISI), University Medical Center Utrecht, Utrecht, the

Netherlands.

Andfinally,StephenR.Aylward;Anassistantprofessorinthe

DepartmentofRadiologyandanadjunctassistantprofessorin

the Department of Computer Science, College of Arts &

Sciences at theUniversity ofNorthCarolina atChapelHill,

USA.HewastheAssociateEditorresponsibleforcoordinating

thereviewofmy first IEEETransactionsonMedical Imaging

paper.


20/200

Chapter 1 Introductionxx

Declaration

I declare that the work in this dissertation was carried out inaccordancewith theRegulationsofHelwanUniversity.Thework isoriginalexceptwhereindicatedbyspecialreferenceinthetextandnopartofthedissertationhasbeensubmittedforanyotherdegree.Thedissertationhasnotbeenpresented toanyotherUniversity forexamination

either

in

the

Arab

Republic

of

Egypt

or

abroad.

AmrS.Ghoneim


21/200

List of Publications xxi

ListofPublicationsInpeerreviewedjournals:[1] Aliaa A. A. Youssif, Atef Z. Ghalwash, and Amr S. Ghoneim, Optic

Disc Detection from Normalized Digital Fundus Images by Means of a

Vessels Direction Matched Filter,IEEETransactions on Medical Imaging,

accepted for publication (in press).

[2] Aliaa A. A. Youssif, Atef Z. Ghalwash, and Amr S. Ghoneim,

Automatic Segmentation of the Retinal Vasculature using a Large-Scale

Support Vector Machine, IEEE Transactions on Medical Imaging,

Submitted for publication.

Ininternationalconferenceproceedings:[3] Aliaa A. A. Youssif, Atef Z. Ghalwash, and Amr S. Ghoneim,

Comparative Study of Contrast Enhancement and Illumination Equalization

Methods for Retinal Vasculature Segmentation, in: Proceedings of the

Third CairoInternational Biomedical Engineering Conf. (CIBEC06), Cairo

Egypt, December 21-24, 2006.

[4] Aliaa A. A. Youssif, Atef Z. Ghalwash, and Amr S. Ghoneim, A

Comparative Evaluation of Preprocessing Methods for Automatic Detection

of Retinal Anatomy, in: Proceedings of the fifth International Conference

on Informatics and Systems (INFOS2007), Cairo Egypt, pp. 24-31, March

24-26, 2007.


22/200

List of Figuresxxii

L i s t o f F i g u r e s

Figure 1.1

A typical generic automatic eye screening system, the figure

highlights the modules that will be included in our research (the light

grey-blocks point out modules that are out of the scope of this work.)

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

5

Figure 1.2Simplified diagram of a horizontal cross section of the human eye.

[4] . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7

Figure 1.3(a) A typical retinal image from the right eye. (b) Diagram of the

retina. [1] . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .8

Figure 1.4 (a) A normal optic disc. (b) Glaucomatous optic disc. [15] . . . . . . . . 11

Figure 2.1

(a) The area of the retina captured in the photograph with respect tothe different FOV's. (b) The computation of scale according the

FOV-geometry. [19] . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

15

Figure 2.2 A typical mask for a fundus image. [24] . . . . . . . . . . . . . . . . . . . . . . . 15

Figure 2.3

(a) Typical retinal image. [24] (b) The green image (green band)of

'a'. (c) The smoothed local average intensity image of 'b'using a 40 40window. (d) Illumination equalized version of 'b'usingEq. 2.1.

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

17

Figure 2.4(a) A typical RGB colored fundus image. (b) Red component image.

(c) Green component. (d) Blue component. [27] . . . . . . . . . . . . . . . .19

Figure 2.5

(a) Typical retinal image. [24] (b) Contrast enhanced version of 'a'

by applying histogram equalization to eachR, G, B bandseparately.

(c) Histogram equalization applied to theIntensitycomponent of 'a'.

(d) Color local contrast enhancement of each R, G, B band of 'a'

separately. (e) Color local contrast enhancement of the Intensity

component of 'a'. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

22

Figure 2.6

(a) The inverted green channel of the retinal image shown in '2.3(a)'.

(b) Illumination equalized version of 'a'. (c) Adaptive histogram

equalization applied to the image 'b'usingEq. 2.9. . . . . . . . . . . . . . .

23

Figure 2.7

(a) The background subtraction of retinal blood vessels and (b) the

estimation of background luminosity and contrast variability

enhancements, both methods are applied to the retinal image shown

in '2.3(a)'. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

25


23/200

List of Figures xxiii

Figure 2.8

The gray-world (a, c)and comprehensive (b, d)normalizations of

2.5(a) and 2.3(a) respectively. The normalization process was

repeated for 5 iterations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

28

Figure 2.9

The summarized procedure shown in blue for applying

histogram specification to a given input image using the histogram

of a reference image. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

30

Figure 2.10

(a) Reference image. [24] (b) Typical retinal image. (c) Colornormalized version of 'b' using the histogram of 'a' by applying

histogram specification separately using each R, G, B bandof both

images. (d) Histogram specification applied using the Intensitycomponent of both images. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

31

Figure 2.11Scatter plot showing the separability of the three classes "Good

image", "Fair image" and "Bad image". [37] . . . . . . . . . . . . . . . . . . .35

Figure 2.12

The Rayleigh distribution, which is a continuous probability

distribution that usually arises when a two dimensional vector (e.g.

wind velocity) has its two orthogonal components normally and

independently distributed. The absolute value (e.g. wind speed)will

then have a Rayleigh distribution. [40] . . . . . . . . . . . . . . . . . . . . . . . .

36

Figure 2.13Chromaticity space (r and g values)plots. (a) No normalization (b)

Histogram specification. [22] . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .39

Figure 2.14

(a) The method of macular vessel length detection. (b) The filed

definition metrics [41]. Constraints are expressed in multiples of

disc diameters DD. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

40

Figure 3.1

(a) Swollen nerve, showing a distorted size and shape. (b) Nerve that

is completely obscured by hemorrhaging. (c) Bright circular lesion

that looks similar to an optic nerve. (d) Retina containing lesions of

the same brightness as the nerve. [18, 27] . . . . . . . . . . . . . . . . . . . . . .

44

Figure 3.2A fundus diagnosed of having high severity retinal/sub-retinal

exudates. [27] . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .45

Figure 3.3A typical healthy fundus image, it shows the properties of a normal

optic disc. [24] . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .45

Figure 3.4

(a) Adaptive local contrast enhancement applied to the intensityimage of the retinal fundus in figure 3.3. (b) The variance image of

'a'. (c) The average variances of 'b'. (d) The OD location (white

cross) determined as the area of highest average variation in

intensity values. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

47

Figure 3.5

An example for the training set of images. Ten intensity images,

manually cropped around the OD from the DRIVE-dataset [24], &

can be used to create the OD model. . . . . . . . . . . . . . . . . . . . . . . . . . .

49


24/200

List of Figuresxxiv

Figure 3.6 The OD template image used by Alireza Osareh [1, 10]. . . . . . . . . . . 50

Figure 3.7 Five-level pyramidal decomposition applied to the green band of thefundus image in Figure 3.3. (a) (e) Image at the first, second, third,

fourth and fifth level correspondingly. . . . . . . . . . . . . . . . . . . . . . . . .

51

Figure 3.8Closing applied to the green band of the fundus image in Figure 3.3

in order to suppress the blood vessels. . . . . . . . . . . . . . . . . . . . . . . . .52

Figure 3.9 The fuzzy convergence image of a retinal blood vasculature. [18] . . . 54

Figure 3.10 A schematic drawing of the vessel orientations. [19] . . . . . . . . . . . . . 55

Figure 3.11

Complete model of vessels direction. For sake of clarity, directions

(gray segments)are shown only on an arbitrary grid of points. [53] .

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

56

Figure 3.12 The OD localization filter (template)used by [43]. . . . . . . . . . . . . . . 57

Figure 4.1One of the 12 different kernels that have been used to detect vessel

segments along the vertical direction. [30] . . . . . . . . . . . . . . . . . . . . .65

Figure 4.2

(a) The maximum responses after applying the 12 kernels proposed

by Chaudhuri et al.[30] to the retinal image shown in 2.6(c). (b)

The corresponding binarized image using the threshold selection

method proposed by Otsu [63]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

66

Figure 4.3 The eight impulse response arrays of Kirschs method. [12] . . . . . . . 67

Figure 4.4

(a) A typical fundus image from the STARE dataset. (b) The first

manual RV segmentation by Adam Hoover. (c) The second manualRV segmentation by Valentina Kouznetsova. (d) The results of

applying thepiecewise threshold probing of a MFR. [27] . . . . . . . . .

68

Figure 4.5

The first (a) and second (b) manual segmentations for the retinal

image in 2.5(a), and the corresponding results of the RVsegmentation methods by Chaudhuri et al.[30] (c), Zana and Klein

[65] (d), Niemeijer et al.[50] (e), and Staal et al.[7] (f). . . . . . . . . . .

70

Figure 4.6

(a)(d) The maximum 2D-Gabor wavelet response for scales a= 2,

3, 4, and 5 pixels respectively, applied to the retinal image in2.5(a). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

72

Figure 4.7(a) Top-hat and (b) Top-hat Hessian-based enhancements, both

applied to the retinal image in 2.5(a). . . . . . . . . . . . . . . . . . . . . . . . .73


25/200

List of Figures xxv

Figure 5.1

(a)(d) STARE images showing different symptoms of DR.

(e) Drusen, a macular degeneration disease usually confused with

bright DR lesions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

77

Figure 5.2Manual image segmentation. (a) An abnormal image. (b) Manually

segmented exudates (in green). (c) Close-up view of exudates. [1] . .80

Figure 5.3 Tissue layers within the ocular fundus. [85] . . . . . . . . . . . . . . . . . . . . 83

Figure 6.1

Results of comparing Mask Generation methods using the STARE.

(1st row) three typical images from the STARE, (2nd, 3rd, and 4th

rows) are the results of applying the mask generation methods of

Gagnon et al. [23], Goatman et al. [22], and Frank ter Haar [19]respectively. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

87

Figure 6.2

Results of comparing Illumination Equalization methods using theDRIVE. Top row images are typical gray-scale images, while the

bottom images represent the highest 2% intensity pixels per image.(a) green-band of a typical DRIVE image. (b) and (c) are

illumination equalized by [19] and [25] respectively. . . . . . . . . . . . . .

89

Figure 6.3

Reviewed normalizations of a typical fundus image. (a) Intensity

image. (b) Green-band image. (c) Histogram equalization.(d) Adaptive local contrast enhancement. (e) Adaptive histogram

equalization. (f) Desired average intensity. (g) Division by an over-

smoothed version. (h) Background subtraction of retinal bloodvessels. (i) Estimation of background luminosity and contrast

variability. (j) Adaptive local contrast enhancement applied to g

instead of a. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

92

Figure 6.4

(a)(j) are the results of applying a RV segmentation method to theimages Fig. 6.3 (a)(j) respectively. (k) A manual segmentation of

Fig. 6.3(a)(used as a gold-standard). . . . . . . . . . . . . . . . . . . . . . . . . .

93

Figure 6.5 ROC curves of the compared contrast enhancements methods. . . . . . 94

Figure 6.6

Color normalization chromaticity plots. (a) Before applying any

normalization methods. (b) Gray-world normalization.

(c) Comprehensive normalization. (d) Histogram equalization.(e) Histogram specification (matching). Red ellipse and elements

plots represent the non-vessels cluster, while the blue represent the

vessels cluster.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

97

Figure 7.1The proposed vessels direction at the OD vicinity matched filter. .

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .103


26/200

List of Figuresxxvi

Figure 7.2

The proposed method applied to the fundus image in 6.8(i). (a) ROI

mask generated. (b) Green-band image. (c) Illumination equalized

image. (d) Adaptive histogram equalization. (e) Binary vessel/non-

vessel image. (f) Thinned version of the preceding binary image.(g) The intensity mask. (h) Final OD-center candidates. (i) OD

detected successfully using the proposed method (White cross, right-

hand side). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

104

Figure 7.3

Results of the proposed method using the STARE dataset (white

cross represents the estimated OD center). (a) The only case where

the OD detection method failed. (b)(h) The results of the proposed

method on the images shown in [53]. . . . . . . . . . . . . . . . . . . . . . . . . .

105

Figure 7.4Results of the proposed method using the DRIVE dataset (white

cross represents the estimated OD center). . . . . . . . . . . . . . . . . . . . . .106

Figure 7.5

The pixels features. (a) A typical digital fundus image from the

DRIVE. (b) The inverted green-channel of a padded using [66].

(c)(f) The maximum 2D-Gabor wavelet response for scales a = 2,

3, 4, and 5 pixels respectively. (g) Top-hat enhancement. (h) Top-hat

Hessian-based enhancement. (i) Green-band Hessian-basedenhancement. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

113

Figure 7.6

(1strow) Using three typical retinal images from the DRIVE, results

of the LS-SVM classifier trained using 2000 pixels with the final set

of features. (2nd row) The corresponding ground-truth manualsegmentation. (3rdrow) Another manual segmentation available for

the DRIVE images. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

115

Figure 7.7

ROC curve for classification on the DRIVE dataset using the LS-

SVM classifier trained using 2000 pixels with the final set of

features. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

116

Figure 7.8

The first steps while detecting Exudates. (1st row) Typical STARE

images containing exudates. (2nd row) The LS-SVM RV

segmentation soft-responses for the 1st row images. (3rd row) The

binarization hard-responses of the RV segmentation outputs inthe 2ndrow. (4throw) The effect of the morphological closing when

applied to inverted versions of the images in the 2ndrow. . . . . . . . . .

120

Figure 7.9

The final steps while detecting Exudates. (1st row) The watersheds

segmentation results of the images in Figure 6.14 (4th row). (2ndrow) A binary mask showing the regions finally selected as exudates

according to their properties. (3rd row) The identified exudates

shown in blue superimposed on the original images of Figure6.14 (1strow). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

121


27/200

List of Figures xxvii

Figure 7.10

STARE images diagnosed manually as not containing any form of

bright-lesions, and indicated as free of any bright-lesions by our

proposed approach (a message indicating so is superimposed on the

upper-left most corner). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

123

Figure 7.11

STARE images manually diagnosed as having forms of bright-

lesions other than hard exudates, and indicated as having brightlesions by our proposed approach. The identified bright lesions are

shown in blue (follow the white arrows), and superimposed on the

original images. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

123

Figure A.1 Front view of a healthy retina. [97] . . . . . . . . . . . . . . . . . . . . . . . . . . . 129

Figure A.2

(a) An example fundus image shows "Puckering the macula" a

macular disease were an opaque membrane obscures the visibilityof the macula and drags the para-macular vessels. (b) The dragged

vessels are shown better in fluorescein angiogram. [98] . . . . . . . . . . .

130

Figure A.3

(a) Typical Hartmann-Shack spot pattern (inverted) from a human

eyes measurement. (b) Hartmann-Shack wavefront sensor with

micro-lens array and image sensor in the focal plane. [101] . . . . . . . .

132

Figure A.4 A typical iris image. [105] . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133

Figure B.1Diabetic retinopathy effect on the vision (a) Without retinopathy. (b)

With retinopathy. [1] . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .138

Figure B.2Diabetic retinopathy syndromes as categorized by [1] (the light grey-

blocks point out DR forms that are not detected by this work.). . . . .140

Figure D.1

Chromaticity Diagram, by the CIE (Commission Internationale de

l'Eclairage the international Commission on Illumination). [112] .

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

150


28/200

List of Tables

xxviii

L i s t o f T a b l e s

Table 2.1 Image-Clarity grading scheme. [41] . . . . . . . . . . . . . . . . . . . . . . . . . . 40

Table 2.2 Field definition grading scheme. [41] . . . . . . . . . . . . . . . . . . . . . . . . . 41

Table 2.3 The sensitivity and specificity of inadequate detection. [41] . . . . . . . 41

Table 6.1 Results of comparing Mask Generation methods using the DRIVE. . 86

Table 6.2Area under curve (AUC) measured per contrast enhancementmethod. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

94

Table 7.1OD detection results for the proposed and literature reviewed

methods. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .107

Table 7.2

Results for the performance evaluation experiments made for our

presented method, and compared to different literature segmentationmethods (for the DRIVE dataset). . . . . . . . . . . . . . . . . . . . . . . . . . . . .

117

Table 7.3Results of detecting Hard Exudates using the STARE. . . . . . . . . . . . 122

Table B.1Projected counts and prevalence of diabetes in Egypt (population 20 years),1995 to 2025. [8] . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

137

Table B.2Distribution of the 300 diabetic patients by their seeking of medical

care. [110] . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .141

Table C.1

The total STARE publicly available 82 images used by [52]

and/or [18], together with the ground truth diagnoses of each image.

[27] . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

146


29/200

Awards xxix

A w a r d s

November

2007

The 1stPlace Winning Postgraduate in the MIA Made-In-the-Arab-

worldCompetition Organized by the Arab League & the Arab

Academy for Science and Technology (ASTF) Cairo, Egypt.

July 2007

The 1stPlace Winning Postgraduate in the MIE Made-In-Egypt

Competition Organized by the IEEE-Egypt Gold Section Cairo,

Egypt.


30/200

Medical Glossaryxxx

M e d i c a l G l o s s a r y

Agerelated macular degeneration (AMD, ARMD); (MAKyulur).Group of conditions that include

deteriorationofthemacula,resultinginlossofsharpcentralvision.Twogeneraltypes: dry, whichis

morecommon,and wet, inwhichabnormalnewbloodvesselsgrowundertheretinaand leakfluid

andblood(neovascularization),furtherdisturbingmacularfunction.Mostcommoncauseofdecreased

visionafterage60.Anteriorchamber;Portionoftheeyebetweenthecorneaandtheiris.

Blindspot;Sightlessareawithinthevisualfieldofanormaleye.Causedbyabsenceof lightsensitive

photoreceptorswheretheopticnerveenterstheeye.

Choroid;Thevascularmiddlecoatoftheeyelocatedbehindtheretinaandinfrontofthesclera.Ciliary

body;Theportionoftheuvealtractbetweenthe irisand thechoroid;composedofciliarymuscleand

processes. Cone; Lightsensitive retinal receptor cell that provides sharp visual acuity and color

discrimination.Cornea;Transparentportionoftheoutercoatoftheeyeballformingtheanteriorwallof

the anterior chamber. Cottonwool

spots; Infarction of the opticnerve fiber layer of the retina, as in

hypertension.

Diabetic retinopathy; (retinAHPuhthee).Spectrumof retinal changesaccompanying longstanding

diabetesmellitus.Early stage is background retinopathy.May advance toproliferative retinopathy,whichincludes the growth of abnormal newblood vessels (neovascularization) and fibrous tissue. Dilated

pupil;Enlargedpupil,resultingfromcontractionofthedilatormuscleorrelaxationoftheirissphincter.

Occursnormallyindimillumination,ormaybeproducedbycertaindrugs(mydriatics,cycloplegics)or

resultfromblunttrauma.Drusen;(DRUzin).Tiny,whitehyalinedepositsonBruchsmembrane(ofthe

retinalpigmentepithelium).Commonafterage60;sometimesanearlysignofmaculardegeneration.

Fluorescein angiography; (FLORuhseen anjeeAHgruhfee). Technique used for visualizing andrecording location and sizeofbloodvessels and any eyeproblems affecting them; fluoresceindye is

injectedinto

an

arm

vein,

then

rapid,

sequential

photographs

are

taken

of

the

eye

as

the

dye

circulates.

Fovea;Thethinnedcentreofthemacula,responsibleforfineacuity.Fundus;Interiorposteriorsurfaceof

theeyeball;includesretina,opticdisc,macula,posteriorpole.Canbeseenwithanophthalmoscope.

Glaucoma;Progressiveopticneuropathywithcharacteristicnerveandvisualfieldchanges.

Intraocularpressure(IOP);Pressurewithintheglobe(Nrrange=8 21mmHg).Iris;Pigmentedtissue

lyingbehindthecorneathatgivescolortotheeye(e.g.,blueeyes)andcontrolsamountoflightentering

theeyebyvaryingthesizeofthepupillaryopening.

Macula;Thesmallavascularareaoftheretinasurroundingthefovea.Mydriasis;Dilationofthepupil.

Neovascularization; (neeohVASkyulurihZAYshun). Abnormal formation of new blood vessels,

usuallyinorundertheretinaorontheirissurface.Maydevelopindiabeticretinopathy,blockageofthe

centralretinalvein,ormaculardegeneration.

Ophthalmologist; (ahfthalMAHlohjist).Physician (MD) specializing indiagnosis and treatment of

refractive,medicalandsurgicalproblemsrelatedtoeyediseasesanddisorders.Ophthalmoscope;(ahfTHALmuhskohp).Illuminatedinstrumentforvisualizingtheinterioroftheeye(especiallythefundus).

Opticdisc,Opticnervehead;Ocularendoftheopticnerve.Denotestheexitofretinalnervefibersfrom

theeyeandentranceofbloodvesselstotheeye.Opticnerve;Largestsensorynerveoftheeye;carries

impulsesforsightfromtheretinatothebrain.

Peripheralvision;Sidevision;visionelicitedbystimulifallingonretinalareasdistantfromthemacula.

Pupil;Variablesizedblackcircularopening in thecenterof the iris thatregulates theamountof light

thatenterstheeye.

Retina;Theinnermostlayeroftheeyecomprisedoftenlayers.


31/200

Chapter 1 Introduction 1

C h a p t e r 1

Introduction1.1 Motivation

The classification of various ocular/ophthalmic eye related patterns is

an essential step in many fields. Whether in medical diagnosis medical

image processing or security, analyzing and classifying ocular images can

aid significantly in automating and improving real-time systems.

Consequently, that leads to a practical and sensible reduction in time and

costs, and prevents in case of medical diagnosis people from suffering

due to various forms of pathologies, with blindness at the forefront.

Medical image processing can be considered recently as one the mostattractive research areas due to the considerable achievements that

significantly improved the type of medical care available to patients,

although it's a multidisciplinary that requires comprehensive knowledge in

many areas such as medicine, pattern recognition, machine learning, and

image processing. Medical image analysis can highly assist physicians in

diagnosing, treating, and monitoring changes of various diseases; hence, a

physician can obtain decision support [1].

Diabetes severe progression is one of the greatest immediate challenges

to the current worldwide health system [1]. Diabetic retinopathy (DR),

beside others, is a common complication of diabetes, and a leading cause of

blindness in Egypt, the Middle-East, and in the working-age population of

western countries. Glaucoma is also a leading cause of blindness worldwide.


32/200

Chapter 1Introduction

2

In general, threats to vision and blinding complications of DR and glaucoma

give little or no warning, but can be moderated if detected early enough for

treatment. Thus, annual screening(s) that employs direct examination with an

ophthalmoscope especially for diabetic patients is highly recommended.

Automatic screening has been shown to be cost effective compared to the

high cost of conventional examination. Insufficient ophthalmologists

especially in rural areas also hinders patients from obtaining regular

examinations. Thus, an automatic system for analyzing the retinal fundus

images would be more practical, efficient, and cost effective.

Artificial Intelligence (AI) may be defined as the branch of computer

science that is concerned with the automation of intelligent behavior. It is

still a young discipline, and its structure, concerns, and methods are less

clearly defined than those of a more mature science such as physics [2],

although it has always been more concerned with expanding the capabilities

of computer science than with defining its limits. AI can be broadly

classified into two major directions:

Logic-Based Traditional AI which includes symbolic processing,

And, Computational Intelligence (CI) which is relatively new and

encompasses approaches primarily based on the bio-inspired artificial

neural networks and evolutionary algorithms, besides fuzzy logic rules,

support vector machines, and also hybrid approaches.

CI approaches can be trained to learn patterns, a property that must be a partof any system that would claim to possess general intelligence [2], and hence

the so-called Machine Learning is one major branch of AI. CI techniques

are increasingly being used in biomedical areas because of the complexity of

the biological systems as well as the limitations of the existing quantitative

techniques in modeling [3]. Even though just stated, performed researches

concerning the analysis and classification of ocular images are mainly based


33/200


on other approaches; for example, statistical methods, geometrical models,

and convolutional kernels.

Finally, many of the researches and methods conducted, especially for

medical diagnosis, lack the evaluation on large benchmark datasets. As a

result, carrying out comparative studies is durable and inflexible.

1.2 Objectives

The main objective of this research is to aid in developing automatic

screening systems for retinopathies (especially DR). Such systems will

significantly help ophthalmologists while diagnosing and treating patients.

Automated screening systems promise with more efficient and costless

medical services, in addition to delivering health-care services to rural areas.

Over and above, automated screening systems will lend a hand to hold back

the personal and social costs of DR as one of the most prevalent

complications of diabetes and one of the leading causes of blindness.

To develop any automatic ocular screening system, firstly we have toanalyze the anatomical structure of the retinal fundus image, which consists

mainly of the Retinal Blood Vessels (RBV), Optic Disc (OD), and Macula.

Detecting the previously mentioned structures will further help in detecting

and quantizing several retinopathies. Then, to detect the presence of DR, we

have to detect a DR manifestation such as exudates, or microaneurysms.

The presence of AI approaches has to be investigated through this

research, since these approaches have proven great effectiveness in pattern

classification and image processing. Employing CI approaches while dealing

with retinal fundus images may highly improve the results achieved. Finally,

all the methods that will be selected for a comparative study or for being

employed in the final proposed system should be compared against the

appropriate benchmark dataset(s) to realize a practical evaluation.


34/200


4

1.3 Thesis Overview

This thesis mainly presents a literature review and a comparative studyfor some of the basic tasks employed while developing an automated ocular

screening system for detecting DR. These main tasks and the overall scope

of this work are shown in Figure 1.1.

The thesis starts with exploring various phases used for preprocessing a

retinal fundus image; which includes mask generation, color normalization,

contrast enhancement, and illumination equalization. Preprocessing a fundus

image is vital step that prepares the image to be effectively analyzed.

Selected preprocessing methods are compared and evaluated using a

standard dataset. The thesis then moves on by exploring various approaches

used for detecting retinal fundus images' landmarks (retinal blood vessels,

optic disc)which are considered the most important anatomical structures in

fundus images. Then new methods are proposed, compared and evaluated on

somehow large, benchmark, and publicly available datasets.

The thesis continues by surveying some researches that aim to detect and

somehow aid the quantification of bright forms manifestations of DR,

and especially the hard exudates. The later researches basically depend on

using the retinal structures previously segmented. Finally, the thesis

summarizes the contribution achieved, and brings together all different

modules of the automated screening system which will be able at this point

to identify diabetic patients who need further examination.

1.4 Thesis Outline

Chapter 1continues on by furnishing basics required to understand the

presented work, and introduces the medical background. Chapter 2

describes different methodologies used to preprocess fundus images and

prepare them for further analysis.


35/200


Figure 1.1A typical generic automatic eye screening system, the figure highlights

the modules that will be included in our research (the light grey-blocks point out

modules that are out of the scope of this work.)

Diabetic Retinopathy

Bright Lesions (mainlyHard Exudates)

Dark (Red)Lesions

Pre-Processing

Mask GenerationColor Normalization

Contrast Enhancement Illumination Equalization

Analyzing Retinal Landmarks

Optic Disc Localization

Retinal Blood Vessels

Segmentation

Fovea Localization

Detecting Retinopathies

Glaucoma

Data Acquisition

Distinguishing the Central

Retinal Artery and Vein

A typical digitized retinal fundus image

acquired for medical diagnoses through a

non-mydriatic fundus camera.

Differentiate Exudates

from Cottonwool Spots

Boundary Detection

Type-of-Eye (left/right)

Detection

Differentiate

Haemorrhages from

Microaneurysms


36/200


6

Chapters 3 and 4provide an overview of previous work, and describe in

some details the implementation for automatically segmenting the optic disc,

and the retinal blood vessels respectively. Chapter 5 describes the

approaches used to automatically extract Exudates, as main bright DR

manifestation of non-proliferative background diabetic retinopathy.

Chapters 6 and 7 present our prototyped system for preprocessing

fundal images, automatically segmenting retinal landmarks, and

automatically detecting exudates. Both chapters also include the comparative

studies carried out to evaluate the system, and provide the results of

experiments that test the capabilities/limitations of the presented methods.

Chapter 8ends this thesis with a summary of the main achievements of the

research, a discussion of possible improvements, proposing possible areas

for future research, as well as concluding remarks.

1.5 Eye Anatomy

Developing a basic understanding of the human eye anatomy (Figure 1.2)

is an appropriate step before going on with this thesis. The eye is nearly a

sphere, with an average diameter of approximately 20 mm [4], enclosed by

three membranes: the cornea and sclera compose the outer cover, the

choroid, and the retina. The cornea is a tough, transparent tissue that covers

the frontal surface, and continuous with it, the sclera is an opaque membrane

that encloses the remainder of the optic globe.

The choroid lies directly below the sclera and it contains a network ofblood vessels that serve as the major source of nutrition. Even superficial

injury to the choroid, often not deemed serious, can lead to severe eye

damage as a result of inflammation that restricts blood flow [4]. The choroid

coat is heavily pigmented and so helps to reduce the amount of extraneous

light entering the eye and the backscatter within the optical globe. At its

anterior extreme, the choroid is divided into the ciliary body and the iris


37/200


diaphragm. The central opening of the iris (the pupil) varies in diameter

from approximately 2 to 8 mm. The front of the iris contains the visible pig-

ment of the eye, whereas the back contains a black pigment.

Retina is the innermost membrane, which lines the inside of the walls

entire posterior portion. When the eye is properly focused, light from an

object outside the eye is imaged on the retina. Pattern vision is afforded by

the distribution of discrete light receptors photoreceptors over the

surface of the retina. These receptors are responsible for receiving light

beams, exchanging them into electrical impulses and then transmitting these

impulses information to the brain where they are turned into images [1].

Figure 1.2

Simplified diagram

of a horizontal cross

section of the human

eye. [4]


38/200


8

There are two classes of receptors: cones and rods. The cones in each eye

number between 6 and 7 million, are highly sensitive to color, and are

located primarily in the central portion of the retina, called the fovea a

circular indentation of 1.5 mm in diameter.Rods are distributed over the

retinal surface and number from 75 to 150 million. The absence of receptors

in the region of emergence of the optic nerve fibers and the blood vessels

from the eye results in the so-called blind spotor the optic disc [4]. Detailed

description of the optic disc, retinal blood vessels, and the fovea are found in

sections 3.2, 4.2 and D.2 respectively, while figure 1.3 shows an example of

a right fundus image including the main anatomical structures.

Figure 1.3(a) A typical retinal image from the right eye. (b) Diagram of the

retina. [1]

The main retinal components numbered in Figure 1.3 are as follows:1- Superior temporal blood vessels

2- Superior nasal blood vessels

3- Fovea / Macula

4- Optic nerve head / Optic disc

5- Inferior temporal blood vessels

6- Inferior nasal blood vessels

a b


39/200


1.6 Fundus Photography and Eye Diseases

Currently, the majority of screenings are carried out by fundalexamination performed by medical staff, which is expensive, and has been

shown to be inaccurate [5]. Using digital fundus photography provides us

with digitized data that could be exploited for computerized detection of

diseases. Fully automated approaches involving fundus image analysis by a

computer could provide an immediate classification of retinopathy without

the need for specialist opinions. Thus, it is more cost-effective, ideal for

those who are unable or unwilling to travel to hospital clinics (especiallythose who live in rural areas), and greatly facilitating the management of

certain diseases.

Automated retinopathy screening systems lately depend on fundus

images taken by a non-mydriatic fundus camera which mainly does not

require pupillary dilatation, and its operator does not need to be skilled at

ophthalmoscopy [5]. In addition, fundus photography surpassed fluorescein

angiography and infrared fluorescence spectrum since no dye need to be

injected into the bloodstream. See Appendix A for more details about the

various forms of eye-related images.

The objective of an automatic screening of fundus images is to improve

the image appearance, interpretation of the main retinal components and

analysis of the image in order to detect and quantify retinopathies such as

microaneurysms, haemorrhages and exudates [5]. In this thesis, we are

mainly concerned with diabetic retinopathies, and somehow glaucoma,due

to their impact on society, and were both should be included among the

avoidable major causes of blindness as some forms of treatment are

available. Among patients presenting to the Alexandria Specialized Medical

Committee for Eye Diseases (Alexandria, Egypt), glaucoma was responsible

for 19.7% of blindness, and diabetic retinopathy for 9% [6].


40/200


41/200


they do not require the injection of fluorescein or indocyanine green dye into

the body. In general, a screening method that does not require trained

personnel would be of a great benefit to screening services by decreasing

their costs, also by decreasing the pressure on available infrastructures and

resources [14]. The pressure is due to the growing numbers of diabetic

patients, with insufficient ophthalmologists to screen them all [5],specially

while World Health Organization (WHO) advices yearly ocular screening of

patients [7].

1.6.2 Glaucoma

Glaucoma is also one of the major causes of preventable blindness. It

induces nerve damage to the optic nerve head (Figure 1.3) via increased

pressure in the ocular fluid (Figure 1.4). In most cases the damage occurs

asymptotically, i.e. before the patient notices any changes to his or her

vision. And this damage is irreversible; treatment can only reduce or prevent

further damage [15]. Age is the most constant risk factor for glaucoma, and a

family history of glaucoma is also a risk factor [6].

Figure 1.4(a) A normal optic disc. (b) Glaucomatous optic disc. [15]a b


42/200


12

Glaucoma is presently detected either by regular inspection of the retina,

measurement of the intra ocular pressure (IOP) or by a loss of vision. It has

been observed that nerve head damage precedes that latter two events and

that direct observation of the nerve head could therefore be a better method

of detecting glaucoma [16]. In [15], fundus images were used to detect and

characterize the abnormal optic disc in glaucoma.

1.6.3 Detecting Retina Landmarks

Detecting retinal landmarks (Figure 1.3) give a framework from which

automated analysis and human interpretation of the retina proceed [17].Identifying these landmarks in the retinal image will highly aid the future

detection and hence quantification of diseases in the mentioned regions. As

an example, in diabetic retinopathy, after detecting retinal main components

an image could be analyzed for sight-threatening complications such as disc

neovascularisation, vascular changes or foveal exudation. Besides just

mentioned, recognizing main components can be used for criteria that allow

the discarding of images that have a too bad quality for assessment ofretinopathy [5]. Methods for detecting the optic disc, and retinal blood

vessels are described in Chapters 3, and 4 respectively.

1.6.4 Fundus Photography Datasets

In this work we used a number of publicly available datasets of retinal

images as a benchmark to evaluate our work, and to compare the

performance of some selected methods. These datasets include a somehowlarge number of healthy retinas images and others with various diseases, it

may include also the field-of-view (FOV) mask of each image, the gold

standard (manual segmentation) used by some algorithms, and the results of

applying specific algorithms to each image (for details seeAppendix C).


43/200

Chapter 2

Preprocessing

13

C h a p t e r 2

Preprocessing2.1 Introduction

A significant percentage of fundus images are of poor quality that

hinders analysis due to many factors such as patient movement, poor focus,

bad positioning, reflections, disease opacity, or uneven/inadequate

illumination. The sphericity of the eye is a significant determinant in the

intensity of reflections from the retinal tissues, in addition, the interface

between the anterior and posterior ocular chambers may cause compound

artifacts such as circular and crescent-shaped low-frequency contrast and

intensity changes [17]. The improper focusing of light may radially decreasethe brightness of the image outward from the center, leading to sort of

uneven illumination known as vignetting [18], and which consequently

results in the optic disc appearing darker than other areas of the image [19].

These artifacts are significant enough to impede human grading in about

10% of retinal images [17], and it can reach 15% in some retinal images sets

[20]. A similar amount is assumed to be of inadequate quality for automated

analysis. Preprocessing of the fundus images can wilt or even remove the

mentioned interferences. This chapter is a literature review that starts with

describing automatic methods for mask generation, proceeds on by

discussing various methods for the preprocessing of a fundus image, and

ends up by describing methods for automatically assessing the quality of

retinal images.


44/200

Chapter 2 Preprocessing14

2.2 Fundamentals of Retinal Digital Image Representation

Digital images in general may be defined as a two-dimensional lightintensity function,f( x, y),wherexandyare spatial (plane)coordinates, and

the amplitude (value)offat any point of coordinates ( x, y)is proportional to

the intensity (brightness, or gray-level)at that point [4]. Digital images are

images whose spatial coordinates and brightness values are selected in

discrete increments, not in a continuous range. Thus a digital image is

composed of a finite number of picture elements (pixels), each of which has

a particular location and value in case of a monochrome 'grayscale' imageor values three values, usually red, green, and blue, in case of a colored

image. In a typical retinal true colored image, the value of each pixel is

usually represented by 24 bits of memory, giving 256 (28) different shades

for each of the three color bands, thus the total of approximately 16.7 million

possible colors.

Generally, the retinal images of the publicly available databases used

through this work were about 685 rows and 640 columns, for a total of

about 440,000 pixels. Although digital retinal images must be of size 2-3000

pixels (2-3 mega-pixels) to match ordinary film resolution [21], digital

retinal images are a practical alternative to ordinary filming.

The extent scope of the captured scene of the retina is called the field

of view (FOV), and is measured in degrees of arc (Figure 2.1(a)). A typical

retina has a FOV that is somewhat more than 180 degrees of arc, but it's not

all captured. The images that were used in this work have a 35 or 45 degrees

of FOV depending on the type and settings of the retinal fundus cameras

used. Since cameras using a 35-FOV show a smaller area of the retina

compared to 45-FOV cameras, 35-FOV images are comparatively

magnified and need to be subsampled before processing [19]. Figure 2.1(b)

shows the subsampling factor calculated according to the FOV-geometry.


45/200


46/200


(4-sigma threshold with a free parameter) value was automatically

calculated using pixel value statistics (mean and standard deviation)outside

the ROI for each color band. Then logical operators (AND/OR) together with

region connectivity test are used to combine the binary results of all bands in

order to identify the largest common connected mask (due to the different

color response of the camera, ROI size is not always the same for each

band)[23].

In [19], the ROI was detected by applying a threshold t to the red color

band (empirically, 35=t ), and then the morphological operators opening,

closing, and erosion were applied respectively (to the result of the

preceding step)using a 33 square kernel to give the final ROI mask. For

more details on the mentioned logical and morphological operators (i.e.

AND, OR, opening, closing, and erosion), see [4].

2.4 Illumination Equalization

The illumination in a retinal image is non-uniform (uneven) due to the

variation of the retina response or the non-uniformity of the imaging system

(e.g. vignetting, and varying the eye position relative to the camera).

Vignetting and other forms of uneven illumination make the typical analysis

of retinal images impractical and useless. For instance, the optic disc (OD) is

characterized as the brightest anatomical structure in a retinal image, hence

applying a simple threshold or grouping the high intensity pixels should

localize the OD successfully. Yet, due to uneven illumination vignetting in

particular the OD may appear darker than other retinal regions, especially

when retinal images are often captured with the fovea appearing in the

middle of the image and the OD to one side [18]. Once the OD loses its

distinct appearance due to the non-uniform illumination or pathologies,

localizing the OD won't be straightforward, especially for methods based on

intensity variation or just on intensity values [19].


47/200

Chapter 2

Preprocessing

17

Figure 2.3

(a) Typical retinal image. [24] (b) The green image (green band)of 'a'

(c) The smoothed local average intensity image of 'b'using a 40 40window.

(d) Illumination equalized version of 'b'usingEq. 2.1.

In order to overcome the non-uniform illumination, illumination

equalization is applied to the image, where each pixel ),( crI is adjusted

using the following equation [18, 19]:

),(),(),( crImcrIcrI Weq += (Eq. 2.1)

where m is the desired average intensity (128 in an 8-bit grayscale image)

and ),( crIW is the mean intensity value (i.e. local average intensity). The

a bdc


48/200


mean intensity value is computed independently for each pixel as the

average intensity of the pixels within a window Wof sizeNN. The local

average intensities are smoothed using the same windowing (Fig. 2.3). The

window sizeNapplied in [18] is variable, in order to use the same number of

pixels (between 30 and 50) every time while computing the average in the

center or near the border of the image.

In [19], a running window of only one size (40 40) and pixels inside

the ROI were used to calculate the mean intensity value, therefore the

amount of pixels used while calculating the local average intensity in the

center is more than the amount of pixels used near the border where the

running window overlaps background pixels. Although the resulting images

look very similar to those using the variable running window, the ROI of the

retinal images is shrunk by five pixels to discard the pixels near the border

where the chances of erroneous values are higher [19].

In [25], correcting the non-uniform illumination in retinal images is

achieved by dividing the image by an over-smoothed version of it using aspatially large median filter. Usually, the illumination equalization process is

applied to the green band (green image)of the retina [18, 19, 25].

2.5 Contrast Enhancement

Enhancement is processing an image so that the result is more

appropriate than the original appearance for a specific application. Contrast

enhancement refers to any process that expands the range of the significantintensities. The various possible processes differ in how the significant range

is identified and how the expansion is performed [5].

2.5.1 Green Band Processing

In order to simply enhance the contrast of the retinal fundus images,

some information is commonly discarded before processing, such as the red


49/200

Chapter 2

Preprocessing

19

so-called red-free images and blue components of the image.

Consequently, only the green band (green image)is extensively used in the

processing (Figure 2.4) as it displays the best vessels/background contrast

[26], and the greatest contrast between the optic disc and the retinal tissue

[15]. In addition, micro-aneurysms a diabetic retinopathy early symptom

are more distinguishable from the background in the green band although

they normally appear as small reddish spots on the retina [25].

Conversely, the red band tends to be highly saturated, thus it's hardly

used by any automated application that uses intensity information alone.

Besides, the blue band tends to be empty, and therefore discarded. Therefore,

many vessel detection and optic disc localization methods are based on the

green component/channel of the color fundus image [15, 25, 26, 28-30].

a b

c d

Figure 2.4

(a) A typical

RGB colored

fundus image.

(b) Red

component

image.(c) Green

component.

(d) Blue

component. [27]


50/200


2.5.2 Histogram Equalization

A typical well-known technique for contrast enhancement is the

histogram (gray-level) equalization [4], which spans the histogram of an

image to a fuller range of the gray scale. The histogram of a digital image

withLtotal possible intensity levels in the range ]1,0[ L is defined as the

discrete function:

1,,2,1,0)( == Lknrh kk K (Eq. 2.2)

where kr is the kth intensity level in the given interval and kn is the number

of pixels in the image whose intensity level is kr [31]. The probability of

occurrence of gray level kr in an image (i.e. the probability density function

'PDF')is approximated by:

1,,2,1,0/)( == Lknnrp kkr K (Eq. 2.3)

where n is the total number of pixels in the image. Thus, a histogram

equalized image is obtained by mapping each pixel with level kr in the input

image to a corresponding pixel with level ks in the output image using the

following equation (based on the cumulative distribution function 'CDF'):

1,,2,1,0/)()(00

==== ==

LknnrprTsk

j

j

k

j

jrkk K

(Eq. 2.4)

Although histogram equalization is a standard technique, it has some

drawbacks since it depends on the global statistics of the image (i.e. pixels

are modified by a transformation function based on the gray-level content of

an entire image). For example, a washed-out appearance can be seen in some

parts of the image due to over enhancement, while other parts need more

enhancing such as the peripheral region (Figure 2.5) [4, 5].


51/200

Chapter 2

Preprocessing

21

2.5.3 Local Contrast Enhancement

As a result of the histogram equalization drawbacks, a local contrast

enhancement technique was invented by Sinthanayothin et al. [5, 32] in

which it doesn't depend on the global statistics of an image, and so it's not

applied to the entire image. Instead, it's applied to local areas depending on

the mean and variance in that area. Considering a small running window or a

sub-image, W, with the size , and centered on the pixel ),( ji , the

mean of the intensity within Wcan be defined as:

=>< ),(),(),( ),(/1

jiWlk

jiW lkfMf (Eq. 2.5)

while the standard deviation of the intensity within Wis:

>


52/200


The local contrast enhancement of a colored image is applied

Artificial Intelligence Techniques for Ocular Pattern Classi...

Documents

Transcript of Artificial Intelligence Techniques for Ocular Pattern Classi...