Age-related macular degeneration (AMD, ARMD); (MAK-yu-lur). ...... aberrometry of the eye,â New Zealand Optics Magazine, Eye on Ophthalmology, May 2002.
Faculty of Computers and Information Computer Science Department
Artificial Intelligence Techniques for Ocular Pattern Classification Amr Ahmed Sabry Abdel‐Rahman Ghoneim Supervised by:
Professor Atef Zaki Ghalwash Professor of Computer Science, The Faculty of Computers & Information, Helwan University, Cairo, Egypt.
Associate Professor Aliaa Abdel‐Haleim Abdel‐Razik Youssif Associate Professor of Computer Science, The Faculty Of Computers & Information, Helwan University, Cairo, Egypt.
Assistant Professor Hosam El‐Dean Mahmoud Bakeir Assistant Professor of Ophthalmology, The Faculty of Medicine, Cairo University, Cairo, Egypt.
A thesis submitted to Helwan University in accordance with the requirements for the degree of Master of Science in Computer Science at the Faculty of Computers & Information, Department of Computer Science.
May 2007
This page is left blank intentionally
2
References
ﺑﺴﻢ ﺍﷲ ﺍﻟﺮﲪﻦ ﺍﻟﺮﺣﻴﻢ ﻦ ﻴ ﺑ ﺮﻕ ﹶﻔ ِﻠ ِﻪ ﻻ ﻧﺳﻭﺭ ِﺒ ِﻪﻭﻛﹸﺘ ﻼِﺋ ﹶﻜِﺘ ِﻪﻭﻣ ﻦ ﺑِﺎﻟ ﱠﻠ ِﻪ ﻣ ﻮ ﹶﻥ ﹸﻛ ﱞﻞ ﺁﺆ ِﻣﻨ ﻤ ﺍﹾﻟﺑ ِﻪ ﻭﺭ ﻦ ﻴ ِﻪ ِﻣ ﻧ ِﺰ ﹶﻝ ِﺇﹶﻟﺎ ﺃﹸﻮ ﹸﻝ ِﺑﻤﺮﺳ ﻦ ﺍﻟ ﻣ ﺁ ﺎ ﺇِﻻﻧ ﹾﻔﺴ ﻪ ﺍﻟ ﱠﻠ ﹶﻜ ﱢﻠﻒ( ﻻ ﻳ285) ﺼﲑ ِ ﻤ ﻚ ﺍﹾﻟ ﻴ ﻭِﺇﹶﻟ ﺎﺑﻨﺭ ﻚ ﻧﺍﺎ ﹸﻏ ﹾﻔﺮﻌﻨ ﻭﹶﺃ ﹶﻃ ﺎﻌﻨ ﺳ ِﻤ ﻭﻗﹶﺎﻟﹸﻮﺍ ِﻠ ِﻪﺳﻦ ﺭ ﺣ ٍﺪ ِﻣ ﹶﺃ ﺎﻴﻨ ﻋ ﹶﻠ ﺤ ِﻤ ﹾﻞ ﺗ ﻻﺎ ﻭﺑﻨﺭ ﺎﺧ ﹶﻄ ﹾﺄﻧ ﻭ ﹶﺃ ﺎ ﹶﺃﻧﺴِﻴﻨ ﺎ ِﺇ ﹾﻥﺍ ِﺧ ﹾﺬﻧﺗﺆ ﺎ ﻻﺑﻨﺭ ﺖ ﺒﺴ ﺘﺎ ﺍ ﹾﻛﺎ ﻣﻴﻬ ﻋ ﹶﻠ ﻭ ﺖ ﺒﺴ ﺎ ﹶﻛﺎ ﻣﺎ ﹶﻟﻬﻌﻬ ﺳ ﻭ ﺎﺮ ﹶﻟﻨ ﺍ ﹾﻏ ِﻔﺎ ﻭﻋﻨ ﻒ ﻋ ﺍﺎ ِﺑ ِﻪ ﻭﺎ ﻻ ﻃﹶﺎ ﹶﻗ ﹶﺔ ﹶﻟﻨﺎ ﻣﻤ ﹾﻠﻨ ﺤ ﺗ ﻻﺎ ﻭﺑﻨﺭ ﺎﺒ ِﻠﻨ ﻦ ﹶﻗ ﻦ ِﻣ ﻋﻠﹶﻰ ﺍﱠﻟﺬِﻳ ﺘﻪﻤ ﹾﻠ ﺣ ﺎﺍ ﹶﻛﻤﺻﺮ ِﺇ (286) ﻦ ﻮ ِﻡ ﺍﹾﻟﻜﹶﺎ ِﻓﺮِﻳ ﻋﻠﹶﻰ ﺍﹾﻟ ﹶﻘ ﺎﺮﻧ ﺼ ﻧﺎ ﻓﹶﺎﻻﻧﻣﻮ ﺖ ﻧﺎ ﹶﺃﻤﻨ ﺣ ﺭ ﺍﻭ
In the name of Allah, Most Gracious, Most Merciful The Messenger believeth in what hath been revealed to him from his Lord, as do the men of faith. Each one (of them) believeth in Allah, His angels, His Books, and His Messengers. ʺWe make no distinction (they say) between one and another of His Messengers.ʺ And they say: ʺWe hear, and we obey, (we seek) Thy forgiveness, our Lord, and to Thee is the end of all journeys.ʺ (285) On no soul doth Allah place a burden greater than it can bear. It gets every good that it earns, and it suffers every ill that it earns. (Pray:) ʺOur Lord! Condemn us not if we forget or fall into error; our Lord! Lay not on us a burden like that which Thou didst lay on those before us; our Lord! lay not on us a burden greater than we have strength to bear. Blot out our sins, and grant us forgiveness. Have mercy on us. Thou art our Protector; help us against those who stand against Faith.ʺ (286) The holy Quran: Chapter 2 ‐ Al‐Baqarah 285:286
:ﻋﻦ ﺃﰊ ﻫﺮﻳﺮﺓ ﺭﺿﻰ ﺍﷲ ﻋﻨﻪ ﺃﻥ ﺭﺳﻮﻝ ﺍﷲ ﺻﻠﻰ ﺍﷲ ﻋﻠﻴﻪ ﻭﺳﻠﻢ ﻗﺎﻝ . ﺃﻭ ﻭﻟﺪ ﺻﺎﱀ ﻳﺪﻋﻮ ﻟﻪ، ﺃﻭ ﻋﻠﻢ ﻳﻨﺘﻔﻊ ﺑﻪ، ﺻﺪﻗﺔ ﺟﺎﺭﻳﺔ:ﺇﺫﺍ ﻣﺎﺕ ﺍﺑﻦ ﺁﺩﻡ ﺍﻧﻘﻄﻊ ﻋﻤﻠﻪ ﺇﻻ ﻣﻦ ﺛﻼﺙ ﺭﻭﺍﻩ ﻣﺴﻠﻢ Abu Huraira (Allah be pleased with him) reported Allahʹs Messenger (May peace and blessings be upon him) as saying: When a man dies, his acts come to an end, but three, recurring charity, or knowledge (by which people) benefit, or a pious son, who prays for him (for the deceased). Sahih Muslim
This page is left blank intentionally
2
References
To the memory of my colleagues Ibraheim Arafat, Mustapha Gamal, & Khaled Abdel‐Mone’im …
And to my beloved city, Cairo, a city that never fails to make an impression, an everlasting, unique impression …
The research described in this thesis was carried out at the Faculty of Computers & Information ‐ Helwan University, Cairo, The Arab Republic of Egypt. 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 storage and retrieval system, without permission in writing from the author. Any trademarks in this publication are property of their respective owners.
Contents 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
Contents
vii
2
viii
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
Contents
3
4
5
2.7.4
Chromaticity Values Distribution . . . . . . . . . . . . . . . . . . . . . .
37
2.7.5
Clarity and Field Definition . . . . . . . . . . . . . . . . . . . . . . . . . . .
38
2.8 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
42
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
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 Two Dimensional Matched Filters . . . . . . . . . . . . . . . . . . . . . . . . . .
63
4.3.2
Existing Automatic Retinal Vasculature Segmentation Algorithms Review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
66
4.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
74
Automatic Detection of Hard Exudates
75
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
75
5.2 Properties of the Diabetic Retinopathy Lesions . . . . . . . . . . . . . . . . . . . .
76
Contents
ix
5.3 Automatic Diabetic Retinopathy Lesions Detection: A Literature Review ........................................................... 5.3.1
6
Existing Automatic Bright Lesions Detection Algorithms Review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
79
5.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
84
Comparative Studies
85
6.1
x
78
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
6.2.4
Observations and Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . .
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
Contents
6.3.4
7
Observations and Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . .
91
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
6.4.4
Observations and Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . .
96
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
7.1.4
Observations and Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . .
106
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
7.2.4
Observations and Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . .
114
Contents
xi
7.3 Hard Exudates Detection using a Large-Scale Support Vector Machine . ............................................
8
7.3.1
Purpose . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
118
7.3.2
Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
118
7.3.3
Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
121
7.3.4
Observations and Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . .
123
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
xii
118
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
Contents
B.4 Diabetic Retinopathy in Egypt . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Appendix C
Fundus Photography Datasets
139 142
C.1 DRIVE Database . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
142
C.2 STARE Project Database . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
143
Appendix D
Color Models
D.1 HSI Color Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
147 147
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
References
151 154
( اﻟﻤﻘﺪﻣﺔ واﻟﻜﻠﻤﺎت اﻟﺮﺋﻴﺴﻴﺔAbstract & Keywords in Arabic) ( اﻟﻐﻼفCover page in Arabic)
Contents
xiii
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 normalization’ and a ‘local contrast enhancement proceeded by illumination equalization’ which xiv
Abstract & Keywords
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 automated analysis 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 81 fundus 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/nonvessels 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 Abstract & Keywords
xv
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 largescale support vector machine for the RV segmentation was proposed. The method 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.
xvi
Abstract & Keywords
Acknowledgment For their tireless support, I thank Mom, Dad, my sister Ghada, and my brother Ayman. I am sure the success of this work would make them delighted. For many years of priceless advice and wisdom which shaped my development as a researcher, I thank my supervisors Prof. Atef Z. Ghalwash and Prof. Aliaa A. Youssif. I owe them much for their assistance in all aspects of my work, providing me great help on various topics, exchanging ideas on both academic and non‐academic matters, and for a wonderful friendship. Thank you very much! Special thanks are given to the consultant ophthalmologist Dr. Hosam El‐Dean Bakeir for his kind supervision of medical expertise. For taking the time to believe in me and investing their minds and hearts into teaching me how to be a better person, I thank my teachers – Lecturers & Assistants at the Faculty of Computers & Information (FCI), Helwan University – who have shaped my mind and made a thoughtful impact on my development over the years. And of 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 Computational Intelligence. And I wish also to express my gratitude to Dr. Waleed A. Yousef who supported me in Linear Algebra. I would like to thank many of my colleagues – the Teaching Assistants at both departments in my faculty – for their support and understanding. Acknowledgment
xvii
I would like to deeply thank my true friends from all grades, especially my class (the 2002 graduates), and both the 2005 and 2006 graduates, always giving me support, encouragement, and helping me keep my spirits up. And I shall never forget my students (hopefully the 2007 and 2008 graduates) for being so patient, understanding and loving. I acknowledge that without all my FCI‐Helwan friends, my life is empty. You kept me going on guys!! I am also indebted to many researchers who supported me with various resources during all stages of writing, and so Special thanks are due to: • Ayman S. Ghoneim; a Teaching Assistant at the Operations Research Dept., the Faculty of Computers & Information ‐ Cairo Univ. (now a M.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 Abdel‐Haleim & 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 Department of Electrical and Computer Engineering, University of Victoria, Victoria, British Columbia, Canada.) • Ghada Khoriba & Ayman Ezzat; both are Teaching Assistants at the Computer Science Department, the Faculty of Computers & Information ‐ Helwan Univ. (currently PhD. students at the Graduate School of Systems and Information Engineering, Tsukuba University, Tsukuba, Japan.) • Mohamed Ali; a Teaching Assistant at the Bioengineering Dept., the Faculty of Engineering ‐ Helwan University. xviii
Acknowledgment
The work of many researchers in this field made this work possible, but my direct contacts with some of them who kindly answered some technical questions pertaining to their work have influenced my thinking a great deal, and so I wish to express my sincere thanks to: • Chanjira Sinthanayothin; the National Electronics and Computer Technology Centre (NECTEC), Thailand. • Subhasis Chaudhuri; Professor & Head, Dept. of Electrical Engineering, Indian Institute of Technology (IIT), Powai, Bombay, India. • Sarah A. Barman; Senior Lecturer, Digital Imaging Research Centre, Kingston University, London, UK. • Norman Katz; CEO of the IP Consulting (a custom enterprise software company), San Diego, California, USA. • Langis Gagnon; Associate Professor, Department of Computer and Electrical Engineering, Université Laval, Quebec, Canada. • Meindert Niemeijer; Ph.D. Student, Image Sciences Institute (ISI), University Medical Center Utrecht, Utrecht, the Netherlands. • And finally, Stephen R. Aylward; An assistant professor in the Department of Radiology and an adjunct assistant professor in the Department of Computer Science, College of Arts & Sciences at the University of North Carolina at Chapel Hill, USA. He was the Associate Editor responsible for coordinating the review of my first IEEE Transactions on Medical Imaging paper.
Acknowledgment
xix
Declaration I declare that the work in this dissertation was carried out in accordance with the Regulations of Helwan University. The work is original except where indicated by special reference in the text and no part of the dissertation has been submitted for any other degree. The dissertation has not been presented to any other University for examination either in the Arab Republic of Egypt or abroad. Amr S. Ghoneim
xx
Chapter 1
Introduction
List of Publications In peer reviewed journals: [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,” IEEE Transactions 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.
In international conference proceedings: [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 Cairo International Biomedical Engineering Conf. (CIBEC’06), 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.
List of Publications
xxi
List of Figures
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 1.2
Simplified 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 to the 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 × 40 window. (d) Illumination equalized version of 'b' using Eq. 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 each R, G, B band separately. (c) Histogram equalization applied to the Intensity component 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' using Eq. 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
xxii
List of Figures
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) Color normalized version of 'b' using the histogram of 'a' by applying histogram specification separately using each R, G, B band of both images. (d) Histogram specification applied using the Intensity component of both images. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
31
Figure 2.11
Scatter 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 2.13
Chromaticity 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 3.2
A fundus diagnosed of having high severity retinal/sub-retinal exudates. [27] . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
45
Figure 3.3
A 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 intensity image 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
List of Figures
xxiii
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 the fundus image in Figure 3.3. (a) – (e) Image at the first, second, third, fourth and fifth level correspondingly. . . . . . . . . . . . . . . . . . . . . . . . .
51
Figure 3.8
Closing 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 4.1
One 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 Kirsch’s 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 manual RV segmentation by Valentina Kouznetsova. (d) The results of applying the piecewise 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 RV segmentation 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 in ‘2.5(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
xxiv
List of Figures
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 5.2
Manual 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 the DRIVE. 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 oversmoothed version. (h) Background subtraction of retinal blood vessels. (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 the images 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 7.1
The proposed “vessels’ direction at the OD vicinity” matched filter. . .......................................................
103
List of Figures
xxv
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/nonvessel 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, righthand 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 7.4
Results 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-based enhancement. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
113
Figure 7.6
(1st row) 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 manual segmentation. (3rd row) Another manual segmentation available for the DRIVE images. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
115
Figure 7.7
ROC curve for classification on the DRIVE dataset using the LSSVM 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 in the 2nd row. (4th row) The effect of the morphological closing when applied to inverted versions of the images in the 2nd row. . . . . . . . . .
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)’. (2nd row) 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 ‘Figure 6.14 (1st row)’. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
121
xxvi
List of Figures
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 brightlesions other than hard exudates, and indicated as having bright lesions 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 visibility of 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 eye‘s 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.1
Diabetic retinopathy effect on the vision (a) Without retinopathy. (b) With retinopathy. [1] . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
138
Figure B.2
Diabetic retinopathy syndromes as categorized by [1] (the light greyblocks 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
List of Figures
xxvii
List of Tables 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 6.2
Area under curve (AUC) measured per contrast enhancement method. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
94
Table 7.1
OD 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 segmentation methods (for the DRIVE dataset). . . . . . . . . . . . . . . . . . . . . . . . . . . . .
117
Table 7.3
Results of detecting Hard Exudates using the STARE. . . . . . . . . . . .
122
Table B.1
Projected counts and prevalence of diabetes in Egypt (population ≥ 20 years), 1995 to 2025. [8] . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
137
Table B.2
Distribution 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
xxviii
List of Tables
Awards November 2007
The 1st Place Winning Postgraduate in the MIA “Made-In-the-Arabworld” Competition – Organized by the Arab League & the Arab Academy for Science and Technology (ASTF) – Cairo, Egypt.
July 2007
The 1st Place Winning Postgraduate in the MIE “Made-In-Egypt” Competition – Organized by the IEEE-Egypt Gold Section – Cairo, Egypt.
Awards
xxix
Medical Glossary Age‐related macular degeneration (AMD, ARMD); (MAK‐yu‐lur). Group of conditions that include deterioration of the macula, resulting in loss of sharp central vision. Two general types: ʺdry,ʺ which is more common, and ʺwet,ʺ in which abnormal new blood vessels grow under the retina and leak fluid and blood (neovascularization), further disturbing macular function. Most common cause of decreased vision after age 60. Anterior chamber; Portion of the eye between the cornea and the iris. Blind spot; Sightless area within the visual field of a normal eye. Caused by absence of light sensitive photoreceptors where the optic nerve enters the eye. Choroid; The vascular middle coat of the eye located behind the retina and in front of the sclera. Ciliary body; The portion of the uveal tract between the iris and the choroid; composed of ciliary muscle and processes. Cone; Light‐sensitive retinal receptor cell that provides sharp visual acuity and color discrimination. Cornea; Transparent portion of the outer coat of the eyeball forming the anterior wall of the anterior chamber. Cotton wool spots; Infarction of the optic nerve fiber layer of the retina, as in hypertension. Diabetic retinopathy; (ret‐in‐AHP‐uh‐thee). Spectrum of retinal changes accompanying long‐standing diabetes mellitus. Early stage is background retinopathy. May advance to proliferative retinopathy, which includes the growth of abnormal new blood vessels (neovascularization) and fibrous tissue. Dilated pupil; Enlarged pupil, resulting from contraction of the dilator muscle or relaxation of the iris sphincter. Occurs normally in dim illumination, or may be produced by certain drugs (mydriatics, cycloplegics) or result from blunt trauma. Drusen; (DRU‐zin). Tiny, white hyaline deposits on Bruchʹs membrane (of the retinal pigment epithelium). Common after age 60; sometimes an early sign of macular degeneration. Fluorescein angiography; (FLOR‐uh‐seen an‐jee‐AH‐gruh‐fee). Technique used for visualizing and recording location and size of blood vessels and any eye problems affecting them; fluorescein dye is injected into an arm vein, then rapid, sequential photographs are taken of the eye as the dye circulates. Fovea; The thinned centre of the macula, responsible for fine acuity. Fundus; Interior posterior surface of the eyeball; includes retina, optic disc, macula, posterior pole. Can be seen with an ophthalmoscope. Glaucoma; Progressive optic neuropathy with characteristic nerve and visual field changes. Intraocular pressure (IOP); Pressure within the globe (Nr range = 8 ‐ 21 mm Hg). Iris; Pigmented tissue lying behind the cornea that gives color to the eye (e.g., blue eyes) and controls amount of light entering the eye by varying the size of the pupillary opening. Macula; The small avascular area of the retina surrounding the fovea. Mydriasis; Dilation of the pupil. Neovascularization; (nee‐oh‐VAS‐kyu‐lur‐ih‐ZAY‐shun). Abnormal formation of new blood vessels, usually in or under the retina or on the iris surface. May develop in diabetic retinopathy, blockage of the central retinal vein, or macular degeneration. Ophthalmologist; (ahf‐thal‐MAH‐loh‐jist). Physician (MD) specializing in diagnosis and treatment of refractive, medical and surgical problems related to eye diseases and disorders. Ophthalmoscope; (ahf‐ THAL‐muh‐skohp). Illuminated instrument for visualizing the interior of the eye (especially the fundus). Optic disc, Optic nerve head; Ocular end of the optic nerve. Denotes the exit of retinal nerve fibers from the eye and entrance of blood vessels to the eye. Optic nerve; Largest sensory nerve of the eye; carries impulses for sight from the retina to the brain. Peripheral vision; Side vision; vision elicited by stimuli falling on retinal areas distant from the macula. Pupil; Variable‐sized black circular opening in the center of the iris that regulates the amount of light that enters the eye. Retina; The innermost layer of the eye comprised of ten layers.
xxx
Medical Glossary
Chapter
1
Introduction 1.1
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 most attractive 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. Chapter 1
Introduction
1
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 part of 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 2
Chapter 1
Introduction
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 to analyze 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. Chapter 1
Introduction
3
1.3
Thesis Overview
This thesis mainly presents a literature review and a comparative study for 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 1 continues 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. 4
Chapter 1
Introduction
Data Acquisition A typical digitized retinal fundus image acquired for medical diagnoses through a non-mydriatic fundus camera.
Pre-Processing Color Normalization
Mask Generation
Contrast Enhancement
Illumination Equalization
Analyzing Retinal Landmarks
Fovea Localization
Optic Disc Localization Retinal Blood Vessels Segmentation Type-of-Eye (left/right) Detection Boundary Detection
Distinguishing the Central Retinal Artery and Vein
Detecting Retinopathies Diabetic Retinopathy Glaucoma
Bright Lesions (mainly Hard Exudates)
Differentiate Exudates from Cottonwool Spots
Dark (Red) Lesions Differentiate Haemorrhages from Microaneurysms
… …
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.) Chapter 1
Introduction
5
Chapters 3 and 4 provide 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 8 ends 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 of blood 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 6
Chapter 1
Introduction
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 pigment of the eye, whereas the back contains a black pigment.
Figure 1.2
Simplified diagram of a horizontal cross section of the human eye. [4]
Retina is the innermost membrane, which lines the inside of the wall’s 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]. Chapter 1
Introduction
7
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 spot or 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.
a b 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 8
Chapter 1
Introduction
1.6
Fundus Photography and Eye Diseases
Currently, the majority of screenings are carried out by fundal examination 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 (especially those 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]. Chapter 1
Introduction
9
1.6.1
Diabetic Retinopathies
Diabetes is a disease that affects about 5.5% of the population worldwide, a number that can be expected to increase significantly in the coming years [7]. In Egypt, Diabetes mellitus is a major emerging clinical and public health problem, while recent surveys from Oman and Pakistan suggest that this may be a regional phenomenon [8]. By the year 2025, nearly 9 million Egyptians (over 13% of the population ≥ 20 years of age) will have diabetes [8], which is considered one of the causes of sudden loss in vision [6]. Diabetic retinopathy is recognized as one of the most prevalent complications of diabetes (both type 1 and type 2) that affects the blood vessels of the retina, and is one of the leading causes of blindness especially in economically advanced countries [9]. It is estimated that people with diabetes have a 25 times greater risk of going blind than non-diabetic population [1]. Everyone who has diabetes is at risk for developing diabetic retinopathy, but not everyone develops it. About 10% of all diabetic patients have diabetic retinopathy, which is the primary cause of blindness in the Western World [1, 7, 10-12], and this is likely to be true in Hong Kong [13]. The personal and social costs of blindness in terms of higher liability to dependence, potential loss of earning capacity, and increased likelihood of greater social support needs, are significant for individuals, for the caring services and for society [11]. See Appendix B for more information about diabetes, diabetic retinopathy, and their prevalence in Egypt. Retinal photography is significantly more effective than direct ophthalmoscopy in detecting diabetic retinopathy [13]. Diabetic retinopathy screening may reduce the risk of blindness in these patients by 50% and can provide considerable cost savings to public health systems [5]. Digital fundus images are more appropriate for automatic screening systems since 10
Chapter 1
Introduction
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].
a b Figure 1.4 (a) A normal optic disc. (b) Glaucomatous optic disc. [15] Chapter 1
Introduction
11
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 of retinopathy [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 somehow large 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 see Appendix C).
12
Chapter 1
Introduction
Chapter
2
Preprocessing 2.1
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 decrease the 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. Chapter 2
Preprocessing
13
2.2
Fundamentals of Retinal Digital Image Representation
Digital images in general may be defined as a two-dimensional light intensity function, f( x, y), where x and y are spatial (plane) coordinates, and the amplitude (value) of f at 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' image – or 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 35º-FOV show a smaller area of the retina compared to 45º-FOV cameras, 35º-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. 14
Chapter 2
Preprocessing
a b Figure 2.1 (a) The area of the retina captured in the photograph with respect to the
different FOV's. (b) The computation of scale according the FOV-geometry. [19]
2.3
Mask Generation
Mask generation aims to label the pixels belonging to the (semi) circular retinal fundus Region of Interest (ROI) in the entire image (Figure 2.2), and exclude the pixels outside the ROI belonging to the dark surrounding region (the background of the image). Those pixels are not strictly dark (0 intensity value) and the need to discard them from the following processing stages is essential [22, 23]. Figure 2.2
A typical mask for a fundus image. [24]
Masks can be generated automatically by simple thresholding of the green color channel followed by 5×5 median filtering [22]. In [23], thresholding was again used but for the three channels (the R, G, and B color bands separately) to generate a binary image for each band. The threshold Chapter 2
Preprocessing
15
(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, t = 35 ), and then the morphological operators – opening, closing, and erosion – were applied respectively (to the result of the preceding step) using a 3×3 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]. 16
Chapter 2
Preprocessing
a b c d 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 × 40 window. (d) Illumination equalized version of 'b' using Eq. 2.1.
In order to overcome the non-uniform illumination, illumination equalization is applied to the image, where each pixel I (r , c) is adjusted
using the following equation [18, 19]: I eq (r , c) = I (r , c) + m − I W (r , c)
(Eq. 2.1)
where m is the desired average intensity (128 in an 8-bit grayscale image) and I W (r , c) is the mean intensity value (i.e. local average intensity). The Chapter 2
Preprocessing
17
mean intensity value is computed independently for each pixel as the average intensity of the pixels within a window W of size N × N. The local average intensities are smoothed using the same windowing (Fig. 2.3). The window size N applied 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 a spatially 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 significant intensities. 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 18
Chapter 2
Preprocessing
– 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].
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]
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]. Chapter 2
Preprocessing
19
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 with L total possible intensity levels in the range [ 0, L − 1] is defined as the discrete function: h(rk ) = nk
k = 0,1, 2,K, L − 1
(Eq. 2.2)
where rk is the kth intensity level in the given interval and nk is the number of pixels in the image whose intensity level is rk [31]. The probability of occurrence of gray level rk in an image (i.e. the probability density function 'PDF') is approximated by: p r (rk ) = nk / n k = 0,1, 2,K, L − 1
(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 rk in the input image to a corresponding pixel with level s k in the output image using the following equation (based on the cumulative distribution function 'CDF'): k
k
j =0
j =0
s k = T (rk ) = ∑ p r (r j ) = ∑ n j / n
k = 0,1, 2, K , L − 1
(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]. 20
Chapter 2
Preprocessing
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 M × M , and centered on the pixel ( i, j ) , the mean of the intensity within W can be defined as: < f >W (i , j ) = 1 / M ∗
∑ f (k , l )
(Eq. 2.5)
( k ,l )∈W ( i , j )
while the standard deviation of the intensity within W is:
σ
2
W
= 1/ M 2 ∗
∑ ( f (k , l ) − < f
( k ,l )∈W ( i , j )
>W ) 2
(Eq. 2.6)
Suppose that f min and f max , are the minimum and maximum intensities in the whole image, then the adaptive local contrast enhancement transformation is defined as: ⎡ [ΨW ( f ) − ΨW ( f min )] ⎤ f (i, j ) → g (i, j ) = 255 ∗ ⎢ ⎥ ⎣ [ΨW ( f max ) − ΨW ( f min )]⎦
(Eq. 2.7)
where the sigmoidal function ΨW ( f ) is: ⎡ ⎛ < f > W − f ⎞⎤ ⎟⎟⎥ ΨW ( f ) = ⎢1 + exp⎜⎜ σW ⎝ ⎠⎦ ⎣
−1
(Eq. 2.8)
The local contrast enhancement is also used by [10], since the exponential produces significant enhancement when the contrast is low (i.e. the local intensity variation is small, and consequently σ W is small), while it provides little enhancement if the contrast is high ( σ W is large). Chapter 2
Preprocessing
21
Figure 2.5
(a) Typical retinal image. [24] (b) Contrast enhanced version of 'a' by applying histogram equalization to each R, G, B band separately. (c) Histogram equalization applied to the Intensity component 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'.
a b c d e
The local contrast enhancement of a colored image is applied to the intensity component of the HSI representation (see Appendix D) of that image. The HSI model separates the intensity component from the color carrying components, yielding significantly better results than using the 22
Chapter 2
Preprocessing
typical RGB model (Figure 2.5(c) and (d)). Finally, the local contrast enhancement increases the noise within the image, so it's recommended to apply a 2D-Gaussian/median filter to the image before enhancement [32]. 2.5.4
Adaptive Histogram Equalization
Contrast enhancement is an essential step in order to detect retinal blood vessels in general, and particularly important while considering the detection of small blood vessels characterized by low contrast levels and intensities that decline significantly with the reduction of the vessels' width. For that reason, adaptive histogram equalization (AHE) was applied by [33] for contrast enhancement, and was found to be more effective than the classical histogram equalization. In order to apply the AHE to an intensity image I, each pixel p in the image is adapted according to the following equation:
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' using Eq. 2.9.
a b c
Chapter 2
Preprocessing
23
r
⎛ ⎞ I AHE ( p ) = ⎜⎜ ∑ s ( I ( p ) − I ( p ' )) / h 2 ⎟⎟ * M ⎝ p '∈R ( p ) ⎠
(Eq. 2.9)
where M = 255 , R ( p ) denotes the pixel p's neighborhood specified by a square window with length h, s (d ) = 1 if d > 0 and s (d ) = 0 otherwise (i.e. if d ≤ 0 ). The values of h and r where empirically chosen to be 81 and 8 respectively. In [33], the AHE was applied to the illumination equalized inverted green band of the image (Figure 2.6). 2.5.5
Background Subtraction of Retinal Blood Vessels (BSRBV)
It is a method proposed by Lin and Zheng [34] to enhance the RV by subtracting the non-stationary background. The fundus image was modeled as having two components: the blood vessels (foreground) and the background which was estimated by averaging the neighborhood (local) intensity to suppress the vessels. The estimated background component was subsequently smoothed by threshold averaging (using the minimum of local average values from the images’ four corner). Then finally, it was subtracted from the original image resulting in an enhanced image (Figure 2.7(a)), free from background interference, thus suitable for later segmentation. 2.5.6
Estimation of Background Luminosity and Contrast Variability (EBLCV)
In order to normalize the illumination and contrast in retinal images, Foracchia et al. [35] used the same foreground-background model proposed by [34]. The non-uniform contrast and luminosity imposed by an acquisition model f (⋅) to the original image I °, resulting in the observed image I can be described as follows: I ( x, y ) = f ( I o ( x, y )) = C ( x, y ) I o ( x, y ) + L( x, y )
(Eq. 2.10)
where C and L are respectively the contrast and luminosity drift factors. 24
Chapter 2
Preprocessing
Therefore, an estimate Iˆ o of the original image was recovered using the following equation: I ( x, y ) − Lˆ ( x, y ) Iˆ o ( x, y ) = Cˆ ( x, y )
(Eq. 2.11)
where Cˆ and Lˆ are the estimation of C and L. In order to extract the background pixels, the mean and standard deviation were calculated in the neighborhood N of each pixel, and then the Mahalanobis distance from the mean was used to mark pixels as background using a threshold value set to 1. The luminosity and contrast drifts were then estimated by calculating the mean and standard deviation for the background pixels in N. Finally, given the observed image, the original image (Figure 2.7(b)) can be recovered using (Eq. 2.11).
a b 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)'.
2.6
Color Normalization
Currently, most retinal screening systems analyze only gray-level images, discarding color information despite the usage of high resolution Chapter 2
Preprocessing
25
color cameras. Color information is potentially useful in classification given that color measurements of lesions in retinal images showed significant differences [22]. Therefore, in order to consistently point out lesions and abnormal colored objects, color must relate directly to the inherent properties of the imaged objects and be independent of imaging conditions. In General, imaging conditions such as lighting geometry (scene illumination) and the imaging device (illuminant color) scales pixels' values in an image, though it doesn't affect the human perception of color. In other words, the assumption that the RGB response of a camera depends only on light reflected from the surface of an object, and thus properly measuring the characteristics of that object, is not correct. The camera's response depends also on the light incident on the object, and the properties of the recording device. Moreover, retina's color (pigmentation) across the population and different patients is variable being strongly correlated to skin pigmentation (amount of melanin) and iris color, thus affecting the classification based on the relatively small color variation between the different retinal lesions. 2.6.1
Gray-World Normalization
Gray-World normalization [22, 36] aims to eliminate the effects due to illumination color (Figure 2.8). Given two pixels viewed under one color of light ( r1 , g 1 , b1 ) and ( r2 , g 2 , b2 ) , the values of the same two pixels but viewed under another different color of light is ( αr1 , β g1 , γb1 ) and ( αr2 , β g 2 , γb2 ) . Then the difference between the two colors of light scaled
the RGB pixels by the factors ( α , β , γ ) respectively. Therefore, and in order to cancel the effect of the scaling factors, we can use the following algebraic equation [36] to calculate the new values for the two pixels before scaling: ⎛ 2r1 2 g1 2b1 ⎜⎜ , , ⎝ r1 + r2 g1 + g 2 b1 + b2 26
⎞ ⎛ 2r2 2g 2 2b2 ⎟⎟, ⎜⎜ , , ⎠ ⎝ r1 + r2 g 1 + g 2 b1 + b2
⎞ ⎟⎟ ⎠
Chapter 2
(Eq. 2.12) Preprocessing
And thus, after scaling, the new values for the two pixels will be: ⎛ 2αr1 2βg1 2γb1 ⎜⎜ , , ⎝ αr1 + αr2 β g1 + βg 2 γb1 + γb2
⎞ ⎛ 2αr2 2βg 2 2γb2 ⎟⎟, ⎜⎜ , , ⎠ ⎝ αr1 + αr2 β g1 + βg 2 γb1 + γb2
⎞ ⎟⎟ ⎠
(Eq. 2.13) Clearly, by canceling the scaling factors from both the numerator and denominator terms (Eq. 2.13), normalization is achieved. The two-pixel example is then extended to the general case using the following equation in order to normalize N-pixels:
r new =
N ∗r N∗g , g new = , and RTotal GTotal
b new =
N ∗b BTotal
(Eq. 2.14)
where ( r new , g new , b new ) are the new values for any pixel ( r , g , b), N is the total number of pixels in each band, and RTotal , GTotal , and BTotal represents the total summation of all the pixels in each of the R, G, and B bands respectively. Dividing (Eq. 2.14) by N gives the final form for the gray-level normalization: r new =
r R Avg
, g new =
g , and G Avg
b new =
b B Avg
(Eq. 2.15)
where R Avg , G Avg , and B Avg represent the average (mean) of all the pixels in each of the R, G, and B bands respectively. It's clear that the mentioned equations effectively canceled the effect of the scaling factors. 2.6.2
Comprehensive Normalization
As stated before, images are captured in conditions varying in the lightning geometry and the illuminant color. And as shown in the previous section, gray-world normalization can successfully remove the effect of different illuminant colors. As for lightning geometry, it can be effectively Chapter 2
Preprocessing
27
removed by representing the image via its chromaticity values (see Appendix D). Practically, both variations don't occur separately, and thus normalization is needed to remove changes arose by both variations at once effectively. Applying color normalization or chromaticity will not eliminate the changes completely, producing unacceptable results [36].
a b c d 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.
Therefore, comprehensive normalization (Figure 2.8) was developed in order to eliminate the image's variation due to both lightning geometry and illuminant color together at the same time [36]. Comprehensive normalization simply applies the chromaticity normalization followed by gray-world normalization (i.e. Eq. D.15 followed by Eq. 2.15), this process is repeated until the change in values is less than a certain termination amount (i.e. there is no sensible change occurring due to the normalization 28
Chapter 2
Preprocessing
iterations). Typically, convergence takes about 4 or 5 iterations [36]. The comprehensive normalization technique was found convergent; termination is guaranteed for all images. And it was found also to be unique; the resultant comprehensively normalized convergent images are the same – unique – for any scene viewed under any lightning geometry and under any illumination color. 2.6.3
Histogram Equalization
Applying histogram equalization (see section 2.5.2) individually to the 3 RGB color bands of an image in order to enhance its contrast affects the color perceived. Consequently, histogram equalization can be used for the color normalization of the retinal fundus images [22]. On the other hand, this will extremely increase the influence of the blue channel in the output retinal images which normally reflect little blue light (Figure 2.5(b)). 2.6.4
Histogram Specification
Histogram specification (matching) is a technique employed to generate an image that has a specified histogram [4]. Thus, enhancement methods can be based on a histogram with a desired specified shape instead of the original histogram or a uniform histogram that results from histogram equalization. The first preprocessing step in [10] was to normalize the retinal images since the retina’s color in different patients is variable being strongly correlated to skin pigmentation and iris color (Figure 2.10). A retinal image was selected as a reference, and then histogram specification was used to modify the values of each image in the database such that its frequency histogram matched the reference image distribution (i.e. the image's histogram). Histogram specification can be applied through the following summarized procedure (Figure 2.9) given by [4]. Given the input image with gray-levels rk , the first step is to obtain its histogram h(rk ) using Eq. 2.2, Chapter 2
Preprocessing
29
and then compute the histogram equalized gray-levels s k for the given input gray-levels rk using Eq. 2.4. The Given Input image: Gray-levels
Histogram Equalization
Histogram Equalization
T (rk )
G( z k ) s k = vk
rk & Histogram
zk
Inverse Transformation
Inverse Transformation
T −1 ( s k )
G −1 (v k )
h(rk )
The reference image: Gray-levels & Histogram
h( z k )
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.
Given the reference image – the one having the desired histogram – with gray-levels z k , we also obtain its histogram and its histogram equalized gray-levels v k , and then we equalize the histogram equalized gray-levels of both the given and reference images (i.e. v k = s k ) giving the following equation: k
k
j =0
j =0
vk = G ( z k ) = ∑ p z ( z j ) = ∑ n j / n = sk
k = 0,1, 2, K , L − 1
(Eq. 2.16) Since v k = s k , and based on Eq. 2.4 and Eq. 2.16, an equation relating the original gray-levels of both the given and reference images – rk & z k – can be given as follows: z k = G −1 [T (rk )] = G −1 ( s k )
k = 0,1, 2,K , L − 1
(Eq. 2.17)
where G −1 is the inverse transformation function of G. In addition, we can also conclude that: 30
Chapter 2
Preprocessing
G( z k ) − sk = vk − sk = 0
k = 0,1, 2,K, L − 1
(Eq. 2.18)
Thus in order to find the new gray-level values for the given image specified by the gray-level values of the reference image, each rk given will be assigned to the smallest integer gray-level z k in the interval [ 0, L − 1] satisfying the following condition: G( z k ) − sk ≥ 0
k = 0,1, 2,K, L − 1
(Eq. 2.19)
a b c d Figure 2.10
(a) Reference image. [24] (b) Typical retinal image. (c) Color normalized version of 'b' using the histogram of 'a' by applying histogram specification separately using each R, G, B band of both images. (d) Histogram specification applied using the Intensity component of both images.
2.7
Image Quality Assessment
Automatic image quality assessment (IQA) is a major concern in medical imaging as it decides whether a particular image is suitable for diagnose Chapter 2
Preprocessing
31
purposes [37]. Besides, it's of significant importance considering automatic screening systems and telemedicine operations where images are captured by non-medical staff, and that's why images need to be evaluated automatically. IQA methods aid in revealing a significant percentage of poor quality fundus images that hinders further analysis due to many factors described in 2.1, also it helps in evaluating different preprocessing methods. Unfortunately, not many researches are dedicated to the study of retinal IQA methods. Overall, IQA methods depend on features extracted from ideal images. 2.7.1
Convolution of Global Intensity Histograms
To the best of our knowledge, the earliest published retinal IQA method is by Lee and Wang [38]. The method used the average intensity histogram for selected set of 20 excellent quality images as a template of the desired histogram. Intensity histograms can measure the effect of brightness, contrast, and signal-to-noise ratio, which are considered the three essential image quality parameters. The intensity histogram can be approximated as a Gaussian curve using the following equation: f (i ) = A * exp(− (i − M ) 2 / 2σ 2 )
(Eq. 2.20)
where i is the pixel intensity value (0:255), A is the peak value of the Gaussian curve, M is the desired over-all intensity mean value, and finally σ is the standard deviation (almost one sixth of the range of the retinal band R, i.e. σ = R / 6 ). Thus, and assuming that A = 1, kernels of the template can be given as follows: K (i ) = exp(− 18 (i − M ) 2 / R 2 )
(Eq. 2.21)
To measure the quality of a given retinal image, its histogram H is convolved with the template histogram K:
32
Chapter 2
Preprocessing
255
C = K ⋅ H = ∑ K (i ) * H (i )
(Eq. 2.22)
i =1
The convolution C is then normalized, giving the quality index Q, which has a maximum value of 1 when the retinal image is of an excellent quality, and 0 indicating that the image is of a very poor quality. The Q wasn't used only to automatically assess the quality of retinal images; it was also used to evaluate different enhancement methods used to improve the over-all quality of the retinal images [38]. In addition, the authors state that the method can be extended and thus applicable for other medical images. 2.7.2
Edge Magnitude and Local Pixel Intensity Distributions
In [37], Lalonde et al. described an alternative method based on edge magnitude and local intensity distributions in order to assess the quality of low resolution retinal images, discriminating them into good, fair and bad images. The new method is similar to that of [38] as it models a good image using a set of high-quality images. The first criterion used by [37] is the edge magnitude distribution which can be thought of as a measure of focus. Since well-focused images have clear and sharp borders, consequently high magnitude edges, it was found that good retinal images have a distribution (histogram) of the edge magnitudes similar to a Rayleigh distribution (Figure 2.12) but with a gentler drop as the intensity increases. On the other hand, bad images have edge distributions that fall rapidly. The difference between the edge magnitude histogram of a target image T and a typical edge magnitude histogram R constructed from the edge maps of a set of highquality images is evaluated using the following equation: d edge (T , R) =
# of bins
∑ i
Chapter 2
( Ri − Ti ) 2 , ∀i | Ri + Ti ≠ 0 Ri + Ti
Preprocessing
(Eq. 2.23)
33
The second criterion used by [37] is the local intensity distribution. All the images in the high-quality set are combined to form an ideal gray-scale model I Mean . The target image I T arg et is then segmented using a standard segmentation algorithm into uniform regions, and the histogram is computed ( I T arg et )
for each region 'i ' in both I T arg et and I Mean giving HRi
and HRi( I Mean ) .
The dissimilarity between the two images can be computed using the following equation: d int ensity ( I T arg et , I Mean ) =
# of regions
∑ i =1
( I T arg et )
Sizei ⋅ W ( HRi
, HRi( I Mean ) )
(Eq. 2.24) where Sizei is the size of the ith region. Finally, the function W is defined using the means of both histograms µh1 − µh2 by: ⎡ µh1 − µh2 ⎤ W (h1 , h2 ) = ⎢ ⎥ ⎣ min( µh1 , µh2 ) ⎦
2
(Eq. 2.25)
Lalonde et al. [37] applied their method to a dataset of 40 images, classified the images into three classes (Figure 2.11) and compared their results against those of an expert ophthalmologist. The experiments showed that both features are appropriate for evaluating the quality of an image, especially the edge magnitude distribution. On the other hand, local intensity distributions where found a little inconclusive [37], and thus it wasn't considered in the next approach described by [39]. 2.7.3
Asymmetry of Histograms Derived from Edge Maps
Toniappa et al. [39] described another approach related to that of Lalonde et al. [37] in that the shape of the edge maps histograms are indicative of image quality, discriminating them into good and poor quality images. This automatic method was considered especially for assessing the 34
Chapter 2
Preprocessing
image quality of retinal images from infants born prematurely in order to monitor Retinopathy-of-Prematurity (ROP). The quality of infant retinal images in general is poorer and varying wider compared to the typical adult retinal images [39]. Therefore, instead of using high quality images to construct a template image or histogram, the measurement of the asymmetry of histograms generated from edge maps is used to assess the image quality.
Figure 2.11 Scatter plot showing the separability of the three classes "Good image",
"Fair image" and "Bad image". [37]
For testing the approach, a training set of 73 infant 648×480 images taken with a CCD color 'RetCam 120' camera were used. The images were categorized as red-based or blue-based images according to the overall color of the image which is dependent on the ethnicity of the infant. Then all images where transformed from the RGB format to the HSI format (see Appendix D), and the intensity images were used to produce the edge maps. Chapter 2
Preprocessing
35
The edge maps were generated by applying a Gaussian matched filter to detect the strengths of retinal blood vessels in the image [30], for more details see Chapter 4. After the edge map for an image is constructed, it's thresholded by a threshold value of 1% of the maximum intensity value in order to discard the noise of the image, and then the histogram of the thresholded edge map is obtained. For a high-quality image, where the vessels are well focused, there will be a wide range of edge strengths present in the edge map, and hence – and as we have mentioned before – its histogram will have an asymmetric shape similar to a Rayleigh distribution with a gentle drop as the intensity increases [37, 39]. In contrast, narrow ranges of edge strengths will be obtained from poor quality ill-focused images, resulting in histograms with a symmetric Gaussian structure. 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]
Skew s(x) is a number that characterizes the shape of a distribution, it's a measure of symmetry, or more precisely, the lack of symmetry. A distribution is symmetric if it looks the same to the left and right of the 36
Chapter 2
Preprocessing
center point (e.g. the skewness for a normal distribution is zero), therefore skew is used in order to directly measure the asymmetry of the edge map histogram. Besides, the standard deviation of the edge map histogram is used to validate the skew measure, given that the standard deviation is on average nine times lower than the skew value measured for an infant retinal image [39]. The skew value is measured using the following equation: n s ( x) = (n − 1)(n − 2)
⎛ xi − x ⎞ ⎟ ⎜ ∑ s ⎠ i =1 ⎝ n
3
(Eq. 2.26)
where n is the number of data points ( x1 , x 2 , x3 , K, x n ), while x is the mean, and s is the variance. The results (using 21 images) showed that the skew values (representing the histogram's asymmetry) of low-quality images tend to be smaller than skew values for higher-quality ones. Even though some degree of overlap is detected between skew values representing the two classes (Good quality and Poor quality), skew values tend to be effective in categorizing retinal images into the two groups. 2.7.4
Chromaticity Values Distribution
Most of the automated screening systems dealing with fundus images depend only on the intensity of the images (i.e. the green intensity, or the intensity component of the HSI model). Discarding the color information by most systems is due to the large natural variation in retinal pigmentation – mainly for ethnic reasons – which consequently masks the naturally slight variations between the different retinal lesion types [10, 22]. Therefore, Most of the previous quality assessment measurements evaluated only the intensity information of the retinal images. In [22], Goatman et al. used the chromaticity values distribution (described in Appendix D) as an IQA technique for evaluating color Chapter 2
Preprocessing
37
normalization methods. The Chromaticity values were used to evaluate the effects of applying three color normalization methods described in 2.6 (i.e. Grey-world normalization, Histogram equalization, and Histogram Specification) to retinal images in order to discriminate different retinal lesions. The clustering of the chromaticity values representing four different retinal lesions (i.e. Cotton wool spots, Hard exudates, Blot Haemorrhages, and Drusen) was measured before applying any of the normalization methods, and after applying each of the normalization methods independently (Figure 2.13). It was found that histogram specification applied to each of the RGB bands separately (Figure 2.10(c)) had the best results in separating the four different lesions using the chromaticity space. 2.7.5
Clarity and Field Definition
To the best of our knowledge, the most recent retinal IQA method is the one proposed by Fleming et al. [41] for automatic diabetic retinopathy grading systems. These systems need to automatically determine that the retinal images are of sufficient quality to be used in grading; for example, blurred images can mask retinal lesions leading to diseased retinas mistakenly graded as normal. The developed method used two aspects to define the quality of a retinal image; image clarity and field definition. Considering the first aspect, image clarity aims to define images with sufficient details for an automatic grading system. For that reason, the visibility of the narrow macular blood vessels is used to measure the clarity of the image (Figure 2.14(a)), since these vessels become easily invisible with any degradation. In addition, macular vessels have width and color similar to those of microaneurysms (the earliest sign of diabetic retinopathy) which must be detected by any diabetic retinopathy grading system [41]. 38
Chapter 2
Preprocessing
a b Figure 2.13 Chromaticity space (r and g values) plots. (a) No normalization (b)
Histogram specification. [22] Chapter 2
Preprocessing
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.
a b
Alternatively, field definition (Figure 2.14(b)) defines images that show the FOV desired for diabetic retinopathy grading systems (i.e. images of full 45° FOV, containing the OD, and centered on the macular region with at least 2 OD diameters of visible retina around the fovea) [41]. Grading Category Excellent Good Fair
Inadequate
Description Small vessels are clearly visible and sharp within one optic disc diameter around the macula. The nerve fiber layer is visible. Either small vessels are clearly visible but not sharp within one optic disc diameter around the macula or the nerve fiber layer is not visible. Small vessels are not clearly visible within one optic disc diameter around the macula but are of sufficient clarity to identify thirdgeneration branches within one optic disc diameter around the macula. Third-generation branches within one optic disc diameter around the macula cannot be identified.
Table 2.1 Image-Clarity grading scheme. [41]
40
Chapter 2
Preprocessing
Grading Category Excellent Good
Inadequate
Description The entire macula and optic disc are visible. The macula is centered horizontally and vertically in the image. The entire macula and optic disc are visible. The macula is not centered horizontally and vertically in the image, but both main temporal arcades are completely visible and the macula is complete (i.e. the fovea is at least two optic disc diameters away from the temporal edge of the image). Either a small-pupil artifact is present (a ring obscuring the outer portion of the image that appears green in the original color image), or at least one of the macula, optic disc, superior temporal arcade, or inferior temporal arcade is incomplete.
Table 2.2 Field definition grading scheme. [41]
Sensitivity
Specificity
Inadequate clarity detection
100%
90.9%
Inadequate field detection
95.3%
96.4%
Inadequate overall quality detection
99.1%
89.4%
Sensitivity: how good the method was at finding true positives (i.e. number of detected true positives / all the true positives). Specificity: how good the method was at excluding true negatives (i.e. number of detected true negatives / all the true negatives). Table 2.3 The sensitivity and specificity of inadequate detection. [41]
The developed method used 395 images for training, and it was tested using 1039 images taken from 977 eyes of 489 patients attending a diabetic retinopathy screening program. An image was labeled inadequate if either the image clarity or field definition was inadequate (Table 2.1 and 2.2); otherwise, the image was graded as having adequate overall quality. Table 2.3 summarizes the sensitivity and specificity of inadequate detection. Chapter 2
Preprocessing
41
2.8
Summary
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. Preprocessing steps label the pixels belonging to the retinal fundus ROI, wilt or even remove artifacts from retinal images, and are a necessary pre-requisite to obtain images with normalized values for luminosity, contrast, and color. In addition, IQA methods are critical for evaluating enhancement methods, and deciding whether a particular image is suitable for further analysis. State-of-the-art studies still struggle with the issue of normalizing contrast, illumination, and color in retinal images, mainly due to the lack of literature reviews and comparative studies. Furthermore, though [35] is a published study that compares the performance of various normalization methods, in general, the available normalization and IQA methods are not being evaluated on large benchmark publicly-available datasets such as [24] and [27]. This chapter presented a literature review of automatic mask generation methods, followed by various methods for normalizing a fundus image, and finally the methods used for automatically assessing the quality of retinal images. Comparative studies accomplished among a number of the presented methods, using two publicly-available fundus datasets, can be found in Chapter 6.
42
Chapter 2
Preprocessing
Chapter
3
Automatic Localization of the Optic Disc 3.1
Introduction
The optic disc (OD) is defined as one of the main features of a fundus retinal image (Figure 1.3), where methods are described for its automatic recognition and localization [17, 32]. Segmenting the OD is a key preprocessing element in many algorithms designed for the automatic extraction of retinal anatomical structures and lesions [23], thus, an associated component of most retinopathy screening systems. Recognizing and localizing the OD is essential, since it often serves as a landmark for other retinal fundus features; for example, the relatively constant distance between the OD and the macula center (fovea) can be used as a priori knowledge to help estimating the location of the macula [23, 32]. The OD was also used as a starting point for retinal blood vessels tracking methods [23, 42]; the large vessels located in the surrounding area of the OD can serve as seeds for vessel tracking procedures. In addition, the OD-rim (boundary) causes false responses for linear blood vessel filters [43]. The change in the shape, color or depth of OD is an indicator of various ophthalmic pathologies (Figure 3.1) especially for glaucoma [44], thus the OD dimensions are used to measure abnormal features due to certain retinopathies, such as glaucoma and diabetic retinopathies [15, 42]. Chapter 3
Automatic Localization of the Optic Disc
43
Glaucoma damages the optic nerve head (OD) due to an increase in pressure within the eye as a result of blockage of the flow of aqueous humour, a watery fluid produced by the ciliary body (Figure 1.2), this damage is visually observed as a change in the relative areas of the optic disc and the cup within the disc [15] (Figure 1.4). Furthermore, the OD can initially be recognized as 'one or more' candidate exudates regions – one of the most occurring lesions in diabetic retinopathies [44] – due to its similarity of its color to the yellowish exudates (Fig. 3.2). Identifying and removing the OD improves the classification of exudates regions [10].
a b c d 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]
After stating the motivations of localizing the OD, this chapter continues on firstly by defining the OD, and highlighting the difference between disc 44
Chapter 3
Automatic Localization of the Optic Disc
localization and disc boundary detection, then it explores most of the available methods for automatic OD localization, and finally the chapter summarizes the OD-detection methods presented and discusses their results.
Figure 3.2
A fundus diagnosed of having high severity retinal/sub-retinal exudates. [27]
3.2
Properties of the Optic Disc
The OD is the brightest feature of the healthy fundus, and it has approximately a circular or a vertically slightly oval (elliptical) shape. In colored fundus images, the OD appears as a bright yellowish or white region (Figure 3.3).
Figure 3.3
A typical healthy fundus image, it shows the properties of a normal optic disc. [24]
The OD is considered the entrance region of the blood vessels and the optic nerves to the retina, also characterized by a pale view owing to the lack of an underlying pigmented layer [17, 42]. The blood vessels – the central Chapter 3
Automatic Localization of the Optic Disc
45
retinal artery and central retinal vein – supplies the upper layers of the retina with blood, while the optic nerve serves as a conduit for the flow of information from the eye to the brain [18]. The OD measures ≈1800µm in diameter [17]. Measured relatively to the retinal fundus image, it occupies about one seventh of the entire image [32]. Alternatively, according to [42], the OD size varies from one person to another, having a diameter in the range of 65~100 in a retinal image of about 512×512 pixels, or about one tenth to one fifth of the image.
3.3
Automatic Optic Disc Localization: A Literature Review
This section begins by defining two processes; OD localization and disc boundary detection. Then it reviews most of the main methods used to localize the OD automatically. 3.3.1
Optic Disc Localization versus Disc Boundary Detection
The process of automatically detecting/localizing the OD aims only to correctly recognize and locate the centroid (centre point) of the OD. On the other hand, disc boundary detection aims to correctly segment the OD by detecting the boundary between the retina and the nerve head (neuroretinal rim). Some methods estimated the contour (boundary) of the OD as a circle or an ellipse (e.g. [15, 23, 32, and 45]). As the obtained circle or ellipse will not exactly match the disc boundary, the exact contour of the OD has been studied, and 'snakes' was proposed due to its ability to bridge discontinuities of the edges [44]. 3.3.2
Existing Automatic OD-Localization Algorithms Review
Although the OD has well-defined features and characteristics, localizing the OD automatically and in a robust manner is not a straight-forward process, since the appearance of the OD may vary significantly due to retinal pathologies. In consequence, in order to effectively detect the OD, the 46
Chapter 3
Automatic Localization of the Optic Disc
various methods developed must consider the variation in appearance, size, and location among different images [19]. This section highlights a number of different techniques employed to automatically localize the OD. a b c d Figure 3.4
(a) Adaptive local contrast enhancement applied to the intensity image 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.
The appearance of the yellowish OD region was characterized by a relatively rapidly variation in intensity because of the "dark" blood-filled vessels were beside the "bright" nerve fibers. This variation in intensity of the adjacent pixels was employed by Sinthanayothin et al. [5, 32] to locate the OD; by identifying the area with the highest variation using a window size equal to that of the OD. The images were preprocessed using the adaptive local contrast enhancement (see 2.5.3) which was applied to the HSI representation (see Appendix D) of the images using a 49×49 running Chapter 3
Automatic Localization of the Optic Disc
47
window. In the 112 TIF fundus images used, the OD occupied about one seventh of the entire 570×550 pixels images. Thus – to locate the OD – the mean, variance, and the average variance through the image was calculated using a running 80×80 window. The index of the highest average variance is considered the center of the OD (Figure 3.4). The OD was located correctly in 111 out of the 112 images used in [5, 32], achieving a success rate of 99.1%. Conversely, it located the OD correctly in 34 out of the 81 images contained in the STARE-dataset [27], achieving a success rate of almost 42.0% only [19]. Instead of using the mean variance, Walter and Klein [46] approximated the OD center as the center of the largest brightest connected object in a fundus image. Their morphological based algorithm initially converted the colored RGB image into its HSI representation. Then a binary image including all the bright regions in the image was obtained by simply thresholding the intensity image. Assuming that the bright appearing retinopathies such as exudates are not very big, and thus far from reaching the OD size [46], the largest connected object in the binary image corresponds to a large portion of the OD. Furthermore, the center of the largest object can be considered the center of the OD. Walter and Klein tested their method on 30 color images of size 640×480 containing various pathologies, locating the OD successfully in all the 30 images [46]. Their method was tested on the STARE-dataset and they achieved a success rate of 58% (i.e. 47 successes out of 81) [19]. In [47], Chrástek et al. applied a 31×31 averaging filter to the green band image, and located the OD roughly at the point of the highest average intensity. The OD localization was 97.3% (254 of 261 images) successful. Again, the brightness was used by Huiqi Li and Opas Chutatape [42, 44] in order to find the OD candidate regions for their model-based approach. 48
Chapter 3
Automatic Localization of the Optic Disc
The pixels with the highest 1% gray-levels in the intensity image are selected; obviously these pixels are mainly from areas in the OD or bright lesions. The selected pixels are then clustered, and clusters having pixels less than 0.04% of the total pixel number in the whole image (i.e. less than ≈100 pixels) are discarded. Then a square with sides of 1.4 times the average OD diameter (i.e. ≈120×120 pixels) is defined for each cluster, and centered on the cluster's centroid. A disc-space (OD model) is created by applying PCA 'Principle Component Analysis' to a training set of 10 intensity normalized (i.e. 0~255) square N×N sub-images manually cropped around the OD (Figure 3.5), then obtaining the six eigenvectors corresponding to the largest six eigenvalues. For each pixel in the candidate regions, the PCA transform is then applied through an N×N sub-image with different scales (0.8~1.2). The OD can be localized as the region with the smallest Euclidian distance to its projection onto the disc-space. Using 89 color images resized to 512×512 this method achieved a success rate of 99% (i.e. 88 successes out of the 89) [44].
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.
Another model-based (template matching) approach was employed by Alireza Osareh et al. [1, 10, and 48] in order to approximately locate the OD. Chapter 3
Automatic Localization of the Optic Disc
49
Initially, the images were normalized by applying histogram specification, and then the OD region from 25 color-normalized images was averaged to produce a 110×110 pixels gray-level template (Figure 3.6). The normalized correlation coefficient was then used to find the most perfect match between the template and all the candidate pixels in the given image. Evaluated on 75 images, the template image could always provide a good approximation of the OD location [1, 48]. Evaluated on the STARE-dataset, the template matching approach achieved a success rate of 58% [19].
Figure 3.6
The OD template image used by Alireza Osareh [1, 10].
One more template matching approach is the Hausdorff-based template matching used together with pyramidal decomposition (Figure 3.7) and confidence assignment by Lalonde et al. [45]. In the beginning, multiresolution processing is employed through pyramidal decomposition which allows large scale object tracking. The small bright retinal lesions – e.g. exudates –vanish at lower resolutions, facilitating the search for the OD region with few false candidates. A simple confidence value is calculated for all the OD candidate regions, and it represents the ratio between the average pixel intensity inside the candidate region, and the average intensity in its neighborhood. The Canny edge detector and a Rayleigh-based threshold were then applied to the green-band image regions corresponding to the candidate regions, constructing a binary edge map. Finally, the edge map regions are matched to a circular template with different radii using the 50
Chapter 3
Automatic Localization of the Optic Disc
Hausdorff distance. Another confidence value is calculated for all the regions having a Hausdorff distance between the edge map and the template less than a certain threshold value. This new confidence value represents a percentage; which is the number of overlapped template pixels divided by the total number of template pixels. The region having the highest total confidence value is considered the OD. This approach was tested on 40 images, and in all cases the OD region was detected correctly [23, 45]. Again evaluated on the STARE-dataset, this approach achieved a success rate of 71.6% [19].
a b c d e Figure 3.7 Five-level pyramidal decomposition applied to the green band of the
fundus image in Figure 3.3. (a) – (e) Image at the first, second, third, fourth and fifth level correspondingly.
Pyramidal decomposition aids significantly in detecting large areas of bright pixels that probably coincides with the OD, but it can be simply fooled by large areas of bright pixels that may occur near the image's borders Chapter 3
Automatic Localization of the Optic Disc
51
due to uneven illumination. For that reason, Frank ter Haar [19] applied illumination equalization (Eq. 2.1) to the green band of the image, and then a resolution pyramid using a simple Haar-based discrete wavelet transform. Finally, the OD area is chosen to correspond with the brightest pixel at the fifth level of the resolution pyramid, achieving a 70.4% success rate using the STARE-dataset. An alternative to the later method is based on the pyramidal decomposition of both the vasculature and the green plan, achieving only 54.3% [19]. The fifth level of the resolution pyramid for both the illumination equalized green band and the binary vessel segmentation are summed, with the highest value corresponding to the OD centre.
Figure 3.8
Closing applied to the green band of the fundus image in Figure 3.3 in order to suppress the blood vessels.
A different method employed to detect the OD is the Hough transform; a technique capable to finding geometric shapes within an image. In [15], Abdel-Ghafar et al. employed the CHT 'Circular Hough Transform' to detect the OD which has a roughly circular shape, and consequently, estimate the centre of the OD. Only the green band of the image was used, and then the retinal blood vasculature in the green image was suppressed using the closing morphological operator (Figure 3.8). The Sobel operator and a simple threshold were then used to extract the edges in the image. CHT was then applied to the edge points, and the largest circle was found consistently 52
Chapter 3
Automatic Localization of the Optic Disc
to correspond to the OD [15], but quantitative results were not provided. Barrett et al. [49] also proposed using the Hough transform to localize the OD, the method was implemented by [19] in two different ways achieving a success rate of 59.3% and 67.9% when tested using the STARE-dataset. Frank ter Haar [19] explored the Hough transform using a couple of methods. In the first one, Hough transform was applied only to pixels on or close to the retinal vasculature in a binary image (thresholded at t = 0.5) of the vessel segmentation obtained by [50], the binary vasculature was dilated using a 5×5 squared kernel in order to increase the possible OD locations. This method attained a success rate of 71.6% when tested using the STAREdataset. Alternatively, Frank ter Haar applied Hough transform once again but only to the brightest 0.35 percent of the fuzzy convergence [18, 51] image achieving only 65.4%, once more dilation was applied to the convergence image using a 6×6 squared kernel to overcome the gaps created by small vessels [19]. The shape, color, and size of the OD showed large variance especially in the presence of retinopathies, and therefore, detection methods based on these properties are shown to be weak, and impractical. An alternative property to examine is the retinal vasculature. The OD is the entrance point for both the optic nerve, and the few main blood vessels which split into many smaller vessels that spread around the retina. As a result, most of the recently proposed techniques try to utilize the information provided by the retinal vasculature. Fuzzy convergence [18, 51] is a novel voting type algorithm developed by Hoover and Goldbaum in order to determine the origination of the retinal vasculature (convergence point), and thus, localize the OD. The convergence point is considered the only consistently visible property of the OD [18]. The inputs to the fuzzy convergence algorithm are six binary vessel Chapter 3
Automatic Localization of the Optic Disc
53
segmentations – each at a different scale – obtained from the green band image. The segmentations are created by applying six threshold values (t = 0.3K 0.8) to the vessel probability map generated by “piecewise threshold probing of a matched filter response” [52]. Each vessel is modeled by a fuzzy segment, which contributes to a cumulative voting image; – a convergence image (Figure 3.9) – where each pixel is assigned a value corresponding to the amount of fuzzy segments the pixel is contained in [19]. Finally, the convergence image is smoothed and thresholded to determine the strongest point(s) of convergence. If the final result was inconclusive, the green band image is illumination equalized (using Eq. 2.1), and Fisher's linear discriminant is applied to regions containing the 1500 brightest pixels in order to localize the OD. This method which is based on multiple fuzzy convergences along with illumination equalization achieved 89% correct detection when tested using the STARE-dataset [18].
Figure 3.9
The fuzzy convergence image of a retinal blood vasculature. [18]
Frank ter Haar [19] constructed a vessel-branch network from a binary vessel image, and then searched for the branch with the most vessels so as to locate the OD. Unfortunately, this proposed method attained a success rate of only 38.3%. Yet again, and using the previously constructed vesselbranch network, a method based on path convergence with Hough transform was investigated, achieving 59.3% correct detections [19]. In this later method, the vessel-branch network was searched for all paths, and since the 54
Chapter 3
Automatic Localization of the Optic Disc
end points of all paths represents a degree of convergence, therefore the area were more path-endings are located probably coincides with the OD location. The Hough transform was then applied to these areas (having the highest paths convergence) in order to detect the OD.
Figure 3.10
A schematic drawing of the vessel orientations. [19]
One last method employed by Frank ter Haar [19] is based on fitting the vasculature orientations on a directional model. Since starting at the OD the retinal vasculature follows approximately the same divergence pattern in all retinal images (Figure 3.10), any segmented vasculature can be fitted on the directional model representing the common divergence pattern in order to locate the OD. The directional model (DM) was created using the vessel segmentations of 80 training images; the method by Niemeijer et al. [50] was used to automatically segment the vessels. For each of the 80 vessel segmentations, the segmentation is improved manually, binary vessel segmentation is obtained (t = 0.5) , the output binary segmentation is then
thinned, the branch-points erased, the vessel segments detected, and all small vessel segments (< 10 pixels) removed. For all the pixels in each remaining vessel, the orientation is calculated to form a directional vessel map. The 80 directional maps are then manually aligned, and the DM is created by averaging – at each pixel – all the corresponding orientation values in the 80 Chapter 3
Automatic Localization of the Optic Disc
55
directional maps. The – right eye – DM created is reflected in order to create a left eye DM. Finally, each pixel in an input vasculature is aligned to the OD centre in both DMs, and then the distance between the fitted input vasculature and both DMs is measured. The pixel having the minimal distance to both DMs is selected as the OD location. This method achieved the highest success rate among the 15 automatic OD-localization methods evaluated by Frank ter Haar [19] (93.8% when tested on the STARE-dataset).
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]
Closely related to vasculature fitting on a directional model, Foracchia et al. [53] identified the position of the OD using a geometrical model of the vessel structure. Once again, this method is based on the fact that in all images, the retinal vasculature originates from the OD following a similar directional pattern. In this method, the main vessels originating from the OD is geometrically modeled using two parabolas. Consequently, the OD position can be located as the common vertex of the two parabolas (i.e. the convergence point for the retinal vasculature). After the complete model of vessels direction was created (Figure 3.11), the vasculature of the input image is extracted, and the difference between the model directions and the extracted vessels directions is minimized using the weighted residual sum of 56
Chapter 3
Automatic Localization of the Optic Disc
squares (RSS) and a simulated annealing (SA) optimization algorithm. Tested on the STARE-dataset, this method achieved a success rate of almost 98% (i.e. 79 successes out of the 81) [53]. A final recent method that depends on the retinal vasculature in order to detect the OD is the one proposed by Fleming et al. [54], where the temporal arcade (i.e. the paths taken by both superior and inferior blood vessels- see figure 1.3) is used to localize the OD. The arcade represents the major vessels forming approximately semi-elliptical paths that enclose the fovea. The semi-elliptical shape of the arcades was detected by applying a generalized Hough transform (GHT). Then finally, a circular Hough transform – guided by the arcade – is performed in order to determine the OD position. Evaluated on 1056 images, this approach achieved a success rate of 98.4% [54]. In [55], the strengths of three OD properties are combined. The blood vasculature convergence towards the OD, the OD's appearance as a bright disc (with a diameter ≈ 1500µm), and the large vessels entering the OD from above and below are used jointly to locate the OD. No quantitative results were provided. Figure 3.12
The OD localization filter (template) used by [43].
Using most of the OD properties previously mentioned, Lowell et al. [43] carefully designed an OD template (Figure 3.12) which was then correlated to the intensity component of the fundus image using the full Pearson-R correlation. This localization filter (template) consists of a Laplacian of Gaussian with a vertical channel carved out of the middle Chapter 3
Automatic Localization of the Optic Disc
57
corresponding to the main blood vessels departing the OD vertically. The OD localization was almost 99% (95 of 96 images) successful. Instead, Tobin et al. [56] proposed a method that mainly relied on vasculature-related OD properties. A Bayesian classifier (trained using 50 images) was used to classify each pixel in red-free images as OD or Not-OD, using probability distributions describing the luminance across the retina and the density, average thickness, and average orientation of the vasculature. Abràmoff and Niemeijer [57] generally used the same features for OD detection, but the features were measured in a special fashion. Then a kNN regression (trained using 100 images) was used to estimate the OD location. To the best of our knowledge, the latter two methods are the only supervised methods proposed in literature for OD detection. 3.3.3
Alternative OD-Localization Algorithms
The OD localization methods reviewed in the previous section paid attention only to fully automated methods that processed retinal fundus images. Therefore, several alternative algorithms for detecting the OD were not included in the preceding review, since they are out of the scope of this work. For example, a semi-automatic algorithm is described in [58], where the centroid of the OD was localized by applying Hough transform to all the points within a manually – user defined – window. The OD was detected in order to avoid sensitive parts of the retina from being harmed during the laser treatment (several thousands of laser shots) targeting the retinal areas affected by choroidal neovascularization “CNV” (a diabetic retinopathy considered among the most important blinding conditions today) [58]. Besides, several OD localization algorithms don't depend on the retinal fundus images. In [59], 3-dimensional optic nerve images are captured using the Confocal Scanning Laser Tomography (CSLT), which is a modern eye imaging technique. The CSLT images were automatically analyzed to 58
Chapter 3
Automatic Localization of the Optic Disc
discriminate between healthy and glaucomatous optic nerves based on shape information. In [60], the OD was detected using three images obtained using different modalities of the scanning laser ophthalmoscope (SLO). The edges in the infrared reflexion (IRC), static scotometry (SK), and argon-blue (ARC) images were used together with the Hough transform to localize the OD for an automatic retinal map generation that helps while diagnosing and treating age-related macular degeneration.
3.4
Summary
Segmenting the OD is an essential step while developing retinal screening systems. The OD location is a prerequisite for many algorithms designed for the automatic extraction of retinal anatomical structures and lesions. Up to date studies still struggle with the issue of localizing the OD, especially when evaluating the method using a harsh dataset full of retinopathies’ manifestations. In addition, a significant percentage of the published methods are not evaluated against publicly-available benchmark datasets, and Frank ter Haar [19] is almost the only comprehensive comparative study evaluating various literature methods using the STARE. This chapter presented a literature review of automatic OD localization methods. A novel technique inspired by the work of Foracchia et al. [53] and Frank ter Haar [19] is presented in Chapter 7. The new method is evaluated using two publicly-available datasets and the results are compared with the other OD detection methods. In general, most of the explored methodologies can be roughly considered utilizing one or more of the following approaches: •
Searching for the highest intensity variation.
•
Searching for the largest connected area with high average intensity values. Chapter 3
Automatic Localization of the Optic Disc
59
•
Searching for the brightest region using multiresolution pyramidal decomposition.
•
Searching for the region most resembling an OD-model using template matching.
•
Searching for the OD geometrical shape (i.e. the region having an ovoid contour), using Hough transform, or using Hausdorff-based matching between detected edges and a circular template.
•
Searching for the region most resembling an OD-model using template matching.
•
Exploring information provided by the retinal vasculature; searching for the area with highest vessel convergences, the area with most vessels endings, the area most resembling the directional model of the blood vessels at the OD vicinity, or the area where a geometrical model of the main vessels (an arcade or a parabola) most fits.
•
Using supervised classifiers (e.g. kNN and Bayesian) trained using various OD properties to classify pixels.
Generally, methods localizing the OD using the intensity variance, high intensity values, and shape are not robust, especially in diseased retinas. On the other hand, methods exploiting the properties of the vasculature tend to achieve improved performances even when applied to abnormal images, and hence they outperform other OD detection methods. Tabulated results comparing various OD localization methods can be found in Chapter 7, or in the comparative study accomplished by Frank ter Haar [19] for fifteen different automatic OD detection methods.
60
Chapter 3
Automatic Localization of the Optic Disc
Chapter
4
Automatic Segmentation of the Retinal Vasculature 4.1
Introduction
Segmentation of the retinal vasculature (RV) in digital fundus images is a key step in automatic screening systems for diabetic retinopathy (DR); one of the most prevalent complications of diabetes [9], and the leading cause of blindness in the Western World [7]. The RV serves as a landmark for other fundus features, such as the optic disc ([18], [19], and [53]). In addition, RV segmentation is required to stop the curvilinear shapes from interfering with later analysis of DR lesions [17], and to assist ophthalmologists by parameters such as vessel tortuosity. The RV can possibly be used as an indicator of the state of the cerebral vasculature, assisting the detection of cerebral microvascular changes in aging, as well as diseases such as vascular dementia and stroke [61]. Besides, measuring the RV geometry at bifurcations (e.g. vessel diameters, branch angles and branch lengths) can be used in obtaining cardiovascular related characteristics, such as volumetric flow, velocity of flow, velocity profile, and vessel-wall shear stress for circulatory and respiratory systems [62]. Finally, since a segmentation of the RV maps the whole retina, does not move except in some diseases, and contains enough information for the localization of some anchor points, RV segmentation is used in the registration of retinal images [28]. Chapter 4
Automatic Segmentation of the Retinal Vasculature
61
After stating the motivations of segmenting the RV, this chapter continues on firstly by defining the RV, and then it explores some of the available methods for automatic RV segmentation. Finally, the chapter summarizes the RV-segmentation methods presented and discusses their results.
4.2
Properties of the Retinal Vasculature
The retinal vasculature consists of the superior/inferior temporal/nasal blood vessels (Figure 1.3); these vessels can be further classified in arteries and veins. The main vessels (arcades) originate from the OD and follow a path that can be geometrically modeled as two parabolas having a common vertex inside the OD (Figure 3.11) [53]. In normal retinas, the vasculature grows fainter as it approaches the macular region (centered on the fovea) where the cones (the photoreceptors highly sensitive to color) are primarily located. Although large and small blood vessels are physically connected, there are significant differences in their sizes and contrast levels; the contrast of the blood vessels fades as the vessels move away from the OD and the main arcades. The blood-filled vasculature appears darker than the retinal background, except in angiograms (Figure A.2) where Fluorescein (a liquid visible in ultra-violet light) is injected by ophthalmologists in order to make the diffusion of blood into the retinal vascular system more visible.
4.3 Automatic Retinal Vasculature Segmentation: A Literature Review Segmentation of the main large blood vessels is rather a straightforward process in noise-free images, but the process turn out to be much more complicated in pathological images full of lesions (e.g. DR manifestations) and/or images of poor quality. In addition, the segmentation of small vessels (especially in the peripheral region) is difficult because of their low contrast 62
Chapter 4
Automatic Segmentation of the Retinal Vasculature
levels. Yet, slight changes in the small blood vessels should be recorded since they may have significant implications concerning serious diseases. RV segmentation methodologies are roughly divided into four major categories: 1) Matched filters, 2) Vasculature profile tracking, 3) Morphological operators, and 4) Pixel classification (Supervised approaches). Through this section, we will review literature methods dealing with digital fundus images only, discarding those applied to fluorescein angiograms, red-free images, and others. 4.3.1
Detection of Blood Vessels in Retinal Images using TwoDimensional Matched Filters
‘A matched filter describes the expected appearance of a desired signal, for purposes of comparative modeling’ [52]; ‘Matched filters are based upon similarity measures between an ideal signal and a measured signal’ [17]. This subsection presents the benchmark classical method proposed by Chaudhuri et al. [30] for segmenting the RV using 2D matched filters owing to its importance. Chaudhuri et al. used an edge fitting method to model the RV. Edge fitting methods are one of the two basic approaches to edge detection which includes also the enhancement/thresholding methods. The later methods enhance the discontinuities in a gray-level image using neighborhood operators (e.g. Robert, Prewitt, and Sobel operators [4]), while the former edge fitting methods minimize the distance between the noisy original data and a predefined edge model. Therefore, the edge fitting methods' results are optimal for the chosen edge model and less sensitive to noise, but they usually involve more computations. Another motivation for using such kernel-based (template matching) methods is the ease of implementation on specialized high speed hardware in addition to the possibility for parallel computing, yet it is computationally expensive in the absence of dedicated hardware. Chapter 4
Automatic Segmentation of the Retinal Vasculature
63
Chaudhuri et al. [30] observed three remarkable properties of the RV, on which they built their method. First of all, the blood vessels usually have small curvatures, hence, may be modeled/approximated using piecewise linear segments. The second observation was that the blood vessels appear much darker compared to the retinal background. These vessels almost never have ideal step edges, and the gray-level (brightness) profiles of the crosssections of retinal vessels have an intensity profile which can be approximated by a Gaussian curve. The last observation is that, although the width of a vessel decreases as it travels away from the optic disc, such a change in vessel quality is a gradual one. Twelve different ‘15×15’ two-dimensional (templates) were generated to model the retinal vasculature along all different orientations with an angular resolution of 15°. The twelve kernels were enough since vessel segments lying within ±7.5° of the direction of the chosen kernel were found to respond well. The original image was smoothed by a ‘5×5’ mean filter to reduce the effect of noise. Then the set of 12 kernels (Figure 4.1) is applied to a fundus image and at each pixel only the maximum of their responses is retained (Figure 4.2(a)). To design the kernels, let p = [ x y ] a point in the
discrete kernel, while θ i is the orientation of the ith kernel. It is assumed that the kernel is centered about the origin [0 0] , hence, the rotation matrix ‘ ri ’ is given by: ⎡cos θ i ri = ⎢ ⎣ sin θ i
− sin θ i ⎤ cos θ i ⎥⎦
(Eq. 4.1)
and, the corresponding point in the rotated coordinate system ‘ pi ’ is: pi = [u v] = p ri 64
T
Chapter 4
(Eq. 4.2) Automatic Segmentation of the Retinal Vasculature
Figure 4.1
One of the 12 different kernels that have been used to detect vessel segments along the vertical direction. [30]
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
4
3
2
1
-2 -5 -6 -5 -2
1
2
3
4
0
0
4
3
2
1
-2 -5 -6 -5 -2
1
2
3
4
0
0
4
3
2
1
-2 -5 -6 -5 -2
1
2
3
4
0
0
4
3
2
1
-2 -5 -6 -5 -2
1
2
3
4
0
0
4
3
2
1
-2 -5 -6 -5 -2
1
2
3
4
0
0
4
3
2
1
-2 -5 -6 -5 -2
1
2
3
4
0
0
4
3
2
1
-2 -5 -6 -5 -2
1
2
3
4
0
0
4
3
2
1
-2 -5 -6 -5 -2
1
2
3
4
0
0
4
3
2
1
-2 -5 -6 -5 -2
1
2
3
4
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
The Gaussian curve has infinitely long double-sided trails which will be truncated into u = ±3σ . And hence, a neighborhood N can be defined as N = {(u, v) : | u |≤ 3σ , | v |≤ L / 2} , with σ and L chosen to be 2, and 9 respectively (σ is a parameter governing the caliber of the blood vessels, while L models the length of the segment assumed to have a fixed orientation). The ith convolutional mask ‘ K i' ’ is given by: K i' ( x, y ) = K i ( x, y ) − mi
∀ pi ∈ N
(Eq. 4.3)
where the initial weights of the ith kernel ‘Ki’: K i ( x, y ) = − exp(−u 2 / 2σ 2 ) ∀ pi ∈ N
(Eq. 4.4)
and the mean value of the kernel ‘mi’ is determined by: mi =
∑ K ( x, y ) / A i
(Eq. 4.5)
pi ∈N
Chapter 4
Automatic Segmentation of the Retinal Vasculature
65
where A denotes the number of points in N. The coefficients are then multiplied by a scale factor of 10 and truncated to their nearest integer. Finally, an automatic thresholding algorithm [63] was used to binarize (Figure 4.2(b)) the matched-filter response ‘MFR’ (i.e. separate the enhanced vessel segments from the background (see Appendix E)). Compared to the Sobel operator and a morphological edge detector, the matched filter shows improved results; it preserves the continuity of the vessels and performs the segmentation process effectively even when the local contrast is quite low. According to [50], this algorithm achieved an area under the ROC curve (AUC) of 0.7878 when applied to the test set of the DRIVE dataset [24].
a b 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].
4.3.2
Existing Automatic Retinal Vasculature Segmentation Algorithms Review
Other ideal vessel profiles (models) are available and are based on assumptions about the anatomy and the practical observation of the RV. For example, instead of using the vessel profile proposed by Chaudhuri et al. 66
Chapter 4
Automatic Segmentation of the Retinal Vasculature
[30], Li and Chutatape [12] used Kirsch’s eight template impulse response arrays (Figure 4.3) to enhance and detect the RV. The eight arrays were convolved with the image to obtain the gradient at different orientations, and as in [30], only the maximum response from all eight directions is retained. Again, a threshold is applied to decide whether a pixel belongs to the RV or not. No quantitative results were reported by Li and Chutatape [12].
Figure 4.3
The eight impulse response arrays of Kirsch’s method. [12]
5
-3 -3
-3 -3
5
-3 -3 -3
5
5
5
5
0
-3
0
5
-3
0
-3
-3
0
-3
5
-3 -3
-3 -3
5
5
5
5
-3 -3 -3
-3
East
West
North
South
-3 -3 -3
-3
5
5
-3 -3 -3
5
5
-3
5
0
-3
-3
0
5
-3
0
5
5
0
-3
5
5
-3
-3 -3 -3
-3
5
5
-3 -3 -3
Northeast
Southwest
Northwest
Southeast
Rather than proposing different filters to match the RV profile, Hoover et al. [52] used the 12 kernels presented in 4.3.1, but presented a novel thresholding method that does not depend on a single global threshold. Thresholding using global attributes (e.g. Otsu [63]) does not provide adequate classification due to the unfortunately strong responses obtained from lesions in the retinal image. In order to segment the RV, Hoover et al. [52] probed the area of the MFR were a potential piece of the RV is present, iteratively decreasing the threshold. And at each iteration, local (regionbased) properties are examined to decide whether to carry on probing, and then finally whether the piece is a vessel or not [52]. The results of this method (Figure 4.4) are available online among the STARE dataset [27], on which it achieved an AUC of 0.7590 [7]. ‘Tracking saves computation time by avoiding scanning the entire image, whereas matched filters are swept over its entirety’[17]. On the other hand, Chapter 4
Automatic Segmentation of the Retinal Vasculature
67
tracking has to begin at a priori known position, and is insufficient to provide a complete RV segmentation in the case where a vessel fades away in the middle part then emerges again in its extended direction [64].
a b c d Figure 4.4
(a) A typical fundus image from the STARE dataset. (b) The first manual RV segmentation by Adam Hoover. (c) The second manual RV segmentation by Valentina Kouznetsova. (d) The results of applying the piecewise threshold probing of a MFR. [27]
Chutatape et al. [64] employed both matched filters and tacking in order to segment the RV completely and computational efficient. Chutatape et al. [64] started by applying a second order derivative 2D-Gaussian ‘25×13’ filter to locate the center point and width of a vessel in its cross-sectional profile, then they used an extended Kalman filter (used for object tracking in control systems) as an estimator to track the RV. 68
Chapter 4
Automatic Segmentation of the Retinal Vasculature
Alternatively, Gagnon et al. [23] used a tracking only to detect the RV, starting at the border of the OD, thus localizing the OD is an initial step for Gagnon’s tracking algorithm. The tracking algorithm uses the edge map computed by Canny edge operator [4], the algorithm tracks a vessels border (edge line) while monitoring the presence of its twin border (i.e. a dual edge tracking). Both Chutatape et al. [64] and Gagnon et al. [23] did not provide any quantitative results for their tracking algorithms. The third category of algorithms used while segmenting the RV is the one using morphological operators. An example of this approach is the method proposed by Zana and Klein [65] where morphological filters and cross-curvature evaluation are used to detect vessel-like structures. The sum of top-hats transformation (based on morphological closing, see [4]) is used to improve the contrast of all linear parts, and then the Laplacian is used for computing the curvature. The algorithm was mainly evaluated using angiograms, beside green images, and inverted green-band digital images (Figure 4.5(d)). According to [50], this algorithm proposed by Zana and Klein [65] achieved an AUC of 0.8984 when applied to the test set of the DRIVE dataset [24]. Supervised methods proved to be more improved and accurate, recording higher performances in terms of receiver operating characteristic (ROC) analysis [7], [50], and [66]. In [50], Niemeijer et al. proposed a supervised pixel classification method for RV segmentation. The pixel features consisted of the green-band intensity value, in addition to the Gaussian and its derivatives up to order 2 at scales 1, 2, 4, 8, and 16. Niemeijer et al. compared three classifiers (kNN, linear, and quadratic classifiers), the kNNclassifier was superior. Moreover, this method’s classification (Figure 4.5(e)) achieved the best performance when compared to four other methods, reporting an AUC of 0.9294 [50] when evaluated using the DRIVE. Chapter 4
Automatic Segmentation of the Retinal Vasculature
69
a b c d e f 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 RV segmentation methods by Chaudhuri et al. [30] (c), Zana and Klein [65] (d), Niemeijer et al. [50] (e), and Staal et al. [7] (f). 70
Chapter 4
Automatic Segmentation of the Retinal Vasculature
The kNN-classifier is employed again by Staal et al. [7], who presented another supervised method that achieved an AUC of 0.9520 using the DRIVE (Figure 4.5(f)) and 0.9614 using the STARE. In order to extract features for each pixel, the image ridges (i.e. crests coinciding with the vessels’ centerlines) are extracted and used to compose primitive line segments. The image is then patched by assigning each pixel to the nearby primitive line segment, and thus the features vector for a pixel contains properties of the patch and line segment it belongs to. Once again, the initial experiments were done using three classifiers (kNN, linear, and quadratic classifiers), and the kNN-classifier was superior. It takes about 15 minutes to segment an image in the DRIVE database on a Pentium-III PC, running at 1.0 GHz with 1-GB memory [7]. Soares et al. [66] presented a method for the RV segmentation using the 2D-Gabor wavelet as the main feature, and a Gaussian mixture model ‘GMM’ classifier (i.e. a Bayesian classifier with class-conditional probability density functions described as a linear combination of Gaussian functions). The set of features for classification included the inverted green channel and its maximum 2D-Gabor transform response for scales 'a' = 2, 3, 4, and 5 pixels, and θ spanning from 0° up to 170° at steps of 10° (Figure 4.6). These scales 'a' were chosen to detect vessels of all possible widths. The wavelet transform was implemented using the fast Fourier transform, and the implementation is publicly available as open source MATLAB® scripts [67]. The classifier was trained using a set of one million labeled pixels obtained from the DRIVE training-set manual segmentations. The method’s performance is evaluated on both DRIVE and STARE datasets, achieving an AUC 0.9614 and 0.9671 respectively. It takes about 3 minutes to generate the features from an image on an AMD Athlon XP PC (2167 MHz clock) with 1-GB memory [66], while the classification of its pixels takes less than 10 seconds. Chapter 4
Automatic Segmentation of the Retinal Vasculature
71
a b c d 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 in ‘2.5(a)’.
Condurache and Aach [68] used four methods (vessel maps) to enhance the separability between vessels and non-vessels, these vessel maps are the results of 1) a Top-hat transform (see [4]), 2) the multiscale analysis of the Top-hat Hessian matrix’s eigenvalues, 3) a Laplace pyramid multi-resolution analysis, and 4) a nonlinear high-pass filter. The Top-hat vessel map (Figure 4.7(a)) was the most separable single vessel map, followed by the Top-hat Hessian-based enhancement (Figure 4.7(b)) [68]. A supervised 72
Chapter 4
Automatic Segmentation of the Retinal Vasculature
multidimensional hysteresis classifier is used for the RV classification, and their vessel segmentation tool is made available at [69].
a b Figure 4.7 (a) Top-hat and (b) Top-hat Hessian-based enhancements, both
applied to the retinal image in ‘2.5(a)’.
The final method is this literature review is proposed by Sinthanayothin et al. [32] where artificial neural networks (ANN) are used for supervised classification. The ANN is scanned over the whole image, and the response of the network is proportional to the likelihood of a vessel being present at the pixel of interest. Features for each pixel consisted of the first component of the principal components analysis (PCA) of the three spectral bands (i.e. the R, G, & B) in a ‘10×10’ sub-window localized on the pixel being classified. In addition, the Canny edge detector response when applied (using the same sub-window) to the first principal component image is added to the set of features to further enhance the vessels/non-vessels separability. A three layers ANN (200 input, 20 hidden, and 2 output nodes respectively) was trained using five-sixths of a 25094 examples dataset, with about twothirds of the set comprising non-vessels data. To enhance the classification results, Sinthanayothin et al. [32] applied some post-processing steps using Chapter 4
Automatic Segmentation of the Retinal Vasculature
73
three criteria (shape, size, and compactness [70]) calculated for all classified-as-vessels regions.
4.4
Summary
The RV segmentation is a cornerstone module in any retinopathy screening system, especially for DR screening. The RV acts as a main landmark for further retinal image analysis, it assists in monitoring cardiovascular and cerebral vasculature properties, and is used for retinal image registration. There are various sate-of-the-art studies dedicated to RV segmentation, and employing a variety of methodologies; such as morphological operators, object tracking, matched filters and supervised classification. The later category showed to be superior in most of the recent studies, with the supervised method presented by Soares et al. [66] achieving the highest performance in terms of AUC when evaluated using both the DRIVE and STARE. Besides enhancing the vessels/non-vessels classification output using benchmark datasets, recent studies are also concerned with raising the computational efficiency, in order to reduce the overall segmentation time. Although most recent studies are evaluated using the DRIVE, a significant percentage of the published methods are not evaluated against publicly-available benchmark datasets, and Niemeijer et al. [50] is almost the only comprehensive comparative study evaluating various literature methods using the DRIVE. This chapter presented a literature review of automatic RV segmentation methods. A novel technique inspired by the work of both Soares et al. [66] and Condurache and Aach [68] is presented in Chapter 7. The new method is evaluated using the publicly-available DRIVE dataset. Tabulated results comparing various RV segmentation methods can be found in Chapter 7, or in the comparative study accomplished by Niemeijer et al. [50] for five different automatic RV segmentation methods. 74
Chapter 4
Automatic Segmentation of the Retinal Vasculature
Chapter
5
Automatic Detection of Hard Exudates 5.1
Introduction Diabetic retinopathies (DR) affecting diabetic patients can usually be graded –
according to severity – into
three stages (background retinopathy, maculopathy, and proliferative retinopathy), more details can be found in Appendix B. Background retinopathy is the most common form of retinopathies, and the less threatening to vision, and that’s why most of the DR screening systems are targeting the detection of background retinopathies which probably must occur before maculopathy and proliferative retinopathy. At this stage the blood vessels in the retina are only very mildly affected, they may bulge slightly (microaneurysms) and may leak blood (haemorrhages) or fluid (exudates), and without affecting the macula area. Maculopathy and proliferative retinopathies are more advanced stages of DR, where vision is highly threatened. The macula region is significantly affected in maculopathy, while proliferative retinopathy is characterized by loss or new-formation of blood vessels especially at blood-deprived areas of the retina. Other ways by which diabetes can affect the eye in general are temporary blurring and cataracts. However, through this study, we are only concerned with the detection of some of the background DR symptoms. Chapter 5
Automatic Detection of Hard Exudates
75
Most sight-loss or sight-threatening problems caused by DR can be prevented or managed by laser treatment (photocoagulation) if it is diagnosed early enough; therefore regular annual eye examinations are extremely important. Considering the identification and classification of DR, retinal non-mydriatic digital photography and its interpretation by an ophthalmologist is proved to be significantly more sensitive than direct ophthalmoscopy as performed by general physicians [13], [71]. Therefore, studies are carried out to justify replacing existing procedures with systematic
screening for DR in terms of cost effectiveness [72], to propose decision support frameworks for automated screening of DR [73], and to address the effect of image resolution on automatic screening routines [74]. After stating the motivations of identifying and classifying DR lesions, this chapter continues on firstly by defining the different lesions, and then it explores some of the available methods for automatic bright DR-lesions detection (mainly Exudates). Finally, the chapter summarizes the lesions detection methods presented and discusses their results.
5.2
Properties of the Diabetic Retinopathy Lesions
Early symptoms of DR (i.e. background retinopathy) can be roughly categorized into two sets; bright lesions, and dark ‘red’ lesions (Figure 5.1). Dark lesions include microaneurysms, which is the earliest clinically recognizable DR warning. Microaneurysms appear as small round reddish dots between the larger vessels in the fundus image, and they occur because of the weakening in the small vessels’ and capillaries’ walls, which lead to blood leakage from these weakened walls. Microaneurysms may rupture leading to the emergence of haemorrhages; another form of DR dark lesions. These haemorrhages have irregular shapes (sometimes flame-shaped), they are generally larger than microaneurysms, and they get larger as the retinal vasculature (RV) gets further damaged. The total number of microaneurysms 76
Chapter 5
Automatic Detection of Hard Exudates
Haemorrhages
Exudates
Microaneurysms Cottonwool spot
Drusen
a b c d e
Figure 5.1 (a)–(d) STARE images showing different symptoms of DR. (e) Drusen,
a macular degeneration disease usually confused with bright DR lesions. Chapter 5
Automatic Detection of Hard Exudates
77
is known to be dynamic; i.e. new microaneurysms are continually forming while existing ones disappear, therefore, microaneurysms turnover needs to be automatically measured [75] since the number of microaneurysms may indicate the probability of DR’s future progression into a more severe stage. And as the RV’s damage progresses, the vessels will start leaking lipids and proteins leading to the appearance of the so-called exudates as small bright (white or yellowish-white) dots in the retinal image. Exudates can be classified to soft-exudates (i.e. oval ill-defined pale white/yellowish-white shapes), and hard-exudates (i.e. small sharp-edged white/yellowish-white deposits). Moreover, and due to the RV leakage, the blood supplies to some areas will be impaired (i.e. ischemic), and thus cottonwool spots (another form of bright DR lesions) will emerge as fluffy whitish patches (blobs) with no well-defined margins if compared to the relatively distinct exudates’ margins. Cottonwool spots are swellings at the surface of the retina due to the deprived-of-blood collapsing nerve fibers. The presence of ischemic areas due to RV leakage and blockage will stimulate the retina to generate new vessels (neovascularisations) that may further burst into larger haemorrhages. Bright DR lesions can be confused with drusen; deep yellowish-white dots with blurred edges seen in age related macular degeneration.
5.3 Automatic Diabetic Retinopathy Lesions Detection: A Literature Review Automatic DR lesions detection is carried out using various methodologies; supervised classification, edge detection, model matching, fuzzy clustering, morphological operators, and many other techniques are stated in literature. Usually, and before applying any of the lesion detection techniques, methods for color/contrast normalizations are applied to reduce the inter-image variation (see chapter 2). Methods employing color 78
Chapter 5
Automatic Detection of Hard Exudates
information for lesion detection are the most affected with inter-image color variance due to pigmentation, especially when dealing with multi-ethnic communities or minority natives such as the Australian Aborigine, the New Zealand Maori, and the Canadian Inui [76]. This section aims to review most of the available literature concerning the detection of bright DR-lesions automatically, and exudates specifically. 5.3.1
Existing Automatic Bright Lesions Detection Algorithms Review
Li and Chutatape [77] detected the exudates from fundus images using region growing and edge detection, reporting a sensitivity and specificity of 100% and 71% respectively. Because the illumination is not normalized through out the image, Li and Chutatape divided the image into 64 subimages, and dealt with each sub-image alone in the Luv color-space. Per subimage, the Canny edge-detector and a region growing method are applied to extract exudates. Region growing was employed also by Sinthanayothin et al. [14] for the detection of exudates. A recursive region growing segmentation (RRGS) algorithm identified similar pixels within a cluster to determine its boundary. The approach was tested using 30 images (of which 21 contained exudates), the sensitivity reported was 88.5% and the specificity was 99.7%. In [1], Alireza Osareh investigated two approaches in order to effectively extract exudates, finding that the Luv is the most suitable color space since it shows the prevalent separability between both exudates/non-exudates classes. The first approach “pixel-level exudates recognition” used manually labeled (Figure 5.2) exudate and non-exudate pixels (50001 and 50437 pixels respectively). Principal Components Analysis (PCA) is then applied using various window sizes (1×1, 3×3, 5×5, and 7×7), and three classifiers were used in the supervised training process (k-nearest neighbor ‘kNN’, Gaussian quadratic ‘GQ’, and Gaussian mixture model ‘GMM’ classifiers). Chapter 5
Automatic Detection of Hard Exudates
79
The best result was the 5×5 window size; using validation sets of 12500 and 12609 pixels representing exudates and non-exudates respectively, the GMM achieved an overall classification accuracy of 96.49%, followed by the kNN with 90.26%, and finally the GQ with 88.24% [1]. The second approach “region-level exudates recognition” segmented the image using a fuzzy cmean (FCM) clustering algorithm, and then features extracted from each cluster are used to discriminate the regions into exudates or non-exudates. Two classifiers (a back-propagation neural network ‘BP-NN’, and a support vector machine ‘SVM’) were investigated, and the BP-NN achieved an accuracy of 93.3%, while the SVM achieved 90.4%. In terms of the area under the receiver operating characteristic curve (AUC), the classifiers reported 0.966 and 0.924 respectively.
a b c Figure 5.2 Manual image segmentation. (a) An abnormal image. (b) Manually segmented exudates (in green). (c) Close-up view of exudates. [1]
Nearly the same approach is used by Zhang and Chutatape [78], [79] for detecting bright lesions in general (i.e. exudates, and cottonwool spots), and then discriminate between them. Zhang and Chutatape used an improved fuzzy c-mean (IFCM) to cluster potential exudate regions, and again in the Luv color space. And then, in order to obtain the final regions, the SVM (trained using region-based features) is employed to classify three kinds of candidate bright lesion regions: non-lesion areas, exudates, and cottonwool 80
Chapter 5
Automatic Detection of Hard Exudates
spots. The classification between bright lesions and bright non-lesions reported a sensitivity of 97% and a specificity of 96%, while classification between exudates and cottonwool spots reported 88% and 84% respectively. Beside the Luv color space, Luo et al. [80] used the Lab color space to compute an object-based color difference image that highlights bright and dark objects and suppresses the undesired background. The difference image is obtained ‘according to the 2D histogram distribution on L-u plane, and then watershed transform is performed on the color difference image to extract object candidates’. No quantitative results were given to this approach. To extract potential bright objects, Goldbaum et al. [55] convolved a flat circular template at different scales with green-band regions having intensity values between 255 and 1.2 times the mean blood vessel intensity. Potential regions’ properties (e.g. region’s color, border’s color, texture measure …etc.) are then evaluated using the logit classifier to determine the final exudate regions, reporting an accuracy of 89%. Another template based approach for exudate detection is proposed by Li and Chutatape [12]; Kirsch’s method (Figure 4.3) is applied to both the red and green bands independently. The exudates and retinal vasculature will both be detected by Kirsch’s templates in the green image, but only exudates will give response in the red image. Therefore, the exudates can be easily segmented using morphological operators. Morphological closing and local variation are used by Walter et al. [81], [82] to locate potential exudate candidates. Initially, morphological closing of a size larger than the thickness of the thickest vessel is used to fill all the vessels. Then using local variation, all regions with high intensity variations (including the OD) are selected as possible exudate candidates. Finally, the exudates are obtained by setting the candidates to zero, applying dilation and Chapter 5
Automatic Detection of Hard Exudates
81
then subtracting the dilated image from the original one to find the contour of the exudate regions. This algorithm achieved a sensitivity of 93% when compared to a manual segmentation. Instead of using statistical measurements such as the variance, Wang et al. [83] employed a statistical classifier named minimum distance discriminant (MDD) which is derived from the Bayes rule. The feature space consists of the spherical coordinates of the typical R, G, and B values which are the θ, φ, and L calculated using the following equations: L = ( R 2 + G 2 + B 2 )1 / 2 , θ = Arc tan(G / R), and ϕ = Arc cos( B / L) (Eq. 5.1) The MDD output is further enhanced by a non-uniform illumination adjustment and a local window-based verification strategy. This approach recorded an accuracy of 100% for identifying images with exudates, and 70% for identifying normal images. Another statistical based classifier (k-nearest neighbor) is used by Meindert Niemeijer [84] to detect and differentiate bright lesions (exudates, cottonwool spots, and drusen). The system is trained using 93067 exudate pixels, 33959 cottonwool spots pixels, and 287186 drusen pixels, while 300 images were selected as a testing set. The steps of the approach as stated by Meindert Niemeijer starts by classifying each pixel according to pixel features (e.g. pixel color, cluster size … etc.) giving the probability of whether the pixel is a part of a bright lesion. Then pixels with high probabilities are grouped together into clusters with probabilities indicating the likelihood that the cluster is a true bright lesion. Finally, clusters with high probabilities are differentiated into exudates, cottonwool spots, or drusen. For detecting the bright lesions, the system reported a sensitivity of 0.95% and a specificity of 0.88%. As for the differentiation between the 82
Chapter 5
Automatic Detection of Hard Exudates
different forms of bright lesions, the system achieved ‘sensitivity / specificity pairs 0.95 / 0.86, 0.70 / 0.93 and 0.77 / 0.88 for the detection of exudates, cottonwool spots and drusen, respectively’ [84]. A totally different approach is by analyzing the interaction of light (i.e. scattering and absorption) with the different structures (retina, choroid and the sclera) of the fundus (Figure 5.3). Since the underlying structures of the fundus responds interacts differently with propagating light, Preece and Claridge [85] proposed to calculate the portion of light remitted from the fundus tissue. Plotting the RGB values (representing the range of the tissue color) on orthogonal axes, healthy tissues showed a small predicted volume within the three-dimensional space. On the other hand, abnormalities due to retinopathies vary the color of the fundus image due to changes in the pigmentation, therefore, the three dimensional plot shows a difference compared to the predicted healthy model.
Figure 5.3 Tissue layers within the ocular fundus. [85] Chapter 5
Automatic Detection of Hard Exudates
83
5.4
Summary
Detecting the bright lesions (exudates, cottonwool spots, drusen … etc.) in retinal fundus image is the one of the essential targets for any automatic screening system, and has a deep impact on the society (especially for diabetic patients). Bright lesions are generally whitish-yellow regions (having nearly the same color of the Optic Disc), and they are clear symptoms for the presence of diabetic retinopathies. There are various sateof-the-art studies dedicated to bright lesions segmentation, and employing a variety of methodologies; such as morphological operators, statistical measurements and classification, matched filters, region growing, clustering algorithms, and supervised classification. Although a significant percentage of the published methods reported their achievements, they are not evaluated against publicly-available benchmark datasets. This chapter presented a literature review of automatic bright lesions detection methods, and focused on particularly hard exudates. A novel technique for bright lesions detection is presented in Chapter 7. The new method is evaluated using the publicly-available STARE dataset. Additional approaches for exudate segmentation can be found in [17].
84
Chapter 5
Automatic Detection of Hard Exudates
Chapter
6
Comparative Studies Besides the literature reviews we presented, this chapter presents the contribution of our study in the form of comparative studies. Each section presents an experiment for comparing some of the available preprocessing literature methods. For all experiments, results are shown using the STARE and/or DRIVE publicly available datasets (see Appendix C). A list of the published papers covering the experiments can be found at the beginning of the document.
6.1
A Comparative Study of Mask Generation Methods
6.1.1
Purpose
The purpose of this experiment is to compare the mask generation methods available in literature. And to determine which method(s) are able to generate binary masks that coincide precisely with the ROI in the digital retinal fundus images. Efficient mask generation will significantly affect the results of the remaining preprocessing methods, and it will certainly affect the accuracy of retinal landmarks/pathologies detection algorithms. 6.1.2
Method
To compare the performance of the 3 automatic mask generation methods described in literature (see section 2.3), we used the manually thresholded FOV included in the DRIVE dataset as a Gold standard. We applied the 3 methods as described in literature. For the simple threshold Chapter 6
Comparative Studies
85
used by Goatman et al. [22] we used the algorithm proposed by Nobuyuki Otsu [63] that selects a global image threshold using only the gray-level histogram (see Appendix E). As for Gagnon et al. [23], we used the 'mean + 4*std. deviation' as a threshold value. 6.1.3
Results
The average sensitivity (number of detected true positives) and specificity (number of detected true negatives) of the 3 methods when applied to the DRIVE are almost equal (Table 6.1), ranging from 98.3 to 100 and 99.3 to 99.95 respectively, with the second method [23] having rather best results. Conversely, when applying the 3 methods to the STARE dataset (Figure 6.1), the method proposed by Frank ter Haar [19] achieved extensively best results, followed by the method in [23]. No concrete results can be shown to the later comparison since no manual masks – to serve as a gold standard – are included with the STARE images. Mask Generation Methods
Sensitivity a
Specificity b
Frank ter Haar [19]
98.987
99.936
Goatman et al. [22]
98.282
99.953
Gagnon et al. [23]
100
99.3
a
Sensitivity: how good the method was at finding true positives (i.e. number of detected true positives / all the true positives). b Specificity: how good the method was at excluding true negatives (i.e. number of detected true negatives / all the true negatives).
Table 6.1 Results of comparing Mask Generation methods using the DRIVE.
6.1.4
Observations and Conclusion
From the mask generation results reported, it is clear that the method presented by Frank ter Haar [19] is superior, since it employed morphological operators to the thresholded image, instead of depending on thresholds alone which can be simply deceived by the uneven illumination. 86
Chapter 6
Comparative Studies
However, thresholding the three color bands of an image as proposed by Gagnon et al. [23] achieves better results than using the green-band solely.
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. Chapter 6
Comparative Studies
87
6.2
A Comparative Study of Illumination Equalization Methods
6.2.1
Purpose
The purpose of this experiment is to compare the illumination equalization methods available in literature, and to conclude whether they can eliminate vignetting and other forms of uneven illumination effectively from digital retinal fundus images. Efficient illumination equalization mask will significantly affect the accuracy of retinal landmarks/pathologies detection algorithms. This section will prove the effect of illumination equalization on optic disc (OD) detection methods. In addition, section 6.3 will prove the effect of illumination equalization on segmenting the retinal vasculature (RV). 6.2.2
Method
As mentioned before, one of the main drawbacks of uneven illumination (vignetting in particular) is the inability to simply analyze the OD. Simple thresholding of the high intensity pixels in a standard fundus image should help in localizing the OD, or at least serve as successful candidates for the OD location. Thus, to measure the effect of both equalization methods proposed by Hoover and Goldbaum [18] and Yang et al. [25], we investigate whether any of the highest 2% intensity pixels lie within the OD region or not. This approach for roughly localizing the OD was applied to the red band intensities by Li and Chutatape [12]. In this experiment, we investigate the highest 2% intensity pixels in the green band after applying both equalization methods. For evaluating the equalization method of Hoover and Goldbaum [18], we used the implementation proposed by Frank ter Haar [19] (see section 2.4); i.e. a running window of only one size (40 × 40) and pixels inside the ROI were used to calculate the mean intensity value. 88
Chapter 6
Comparative Studies
6.2.3
Results
Using the 121 images in both the DRIVE and STARE datasets, the highest 2% intensity pixels in the green band successfully fell in the OD region of 102 images, with a success rate of about 84.3%. After applying both illumination equalizations proposed by [19] and [25], the success rates noticeably rose to 98.35% (i.e. failed in 2 images only), and 97.52% (i.e. failed in 3 images), respectively. 6.2.4
Observations and Conclusion
It is clear that applying illumination equalization to fundus images considerably improves further analysis tasks. Both equalization methods found in literature have nearly the same influence (Figure 6.2).
(a)
(b)
(c)
Figure 6.2 Results of comparing Illumination Equalization methods using the DRIVE. 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. Chapter 6 Comparative Studies 89
6.3
A Comparative Study of Contrast Enhancement Methods
6.3.1
Purpose
The purpose of this experiment is to compare some of the contrast enhancement methods available in literature, and to conclude whether they can affect the process of segmenting the retinal vasculature from digital retinal fundus images. According to available studies, efficient contrast enhancement methods will significantly affect the accuracy of RV segmentation algorithms, and thus affecting the processes of detecting other retinal landmarks (e.g. the OD, and macula) and retinopathies. This section will compare and evaluate the effect of various contrast enhancement methods, especially while segmenting the RV. 6.3.2
Method
Instead of applying the adaptive local contrast enhancement to the intensity image as presented by [5], we propose a hybrid method by applying the local contrast enhancement to the illumination equalized image obtained by [25] using division by an over-smoothed version. Thus, in order to compare the eight reviewed contrast normalization methods (Figure 6.3 (b)– (i)), and the proposed hybrid method (Figure 6.3 (j)), a vessel segmentation algorithm [30] is applied to the 60 images (before and after applying each of the normalization methods), (Figure 6.4). The publicly available manual RV segmentations are used as our gold standard. The intensity image (Figure 6.3 (a)) represents a normal gray-level image before enhancement. The adaptive histogram equalization, and the BSRBV methods were applied to an inverted green band image, thus both methods used an inverted version of the 2D matched filters used for segmentation [30].
90
Chapter 6
Comparative Studies
The EBLCV was applied using 201×201 tessellated windows [35], and then full images were approximated using bicubic interpolation. For evaluating the performance of the normalization methods, we used the receiver operating characteristic (ROC) curve [86] (Figure 6.5). The area under the ROC curve (AUC) was calculated using a non-parametric estimate [87] to measure the overall performance as shown in Table 6.2. 6.3.3
Results
Considering the literature reviewed methods, the adaptive histogram equalization was found to be the most effective method as it obviously improved the segmentation algorithm, thus it attained the largest AUC. The adaptive local contrast enhancement and the typical histogram equalization methods recorded respectively the second and third best results, noticeably improving the RV segmentation algorithm. In general, the three mentioned methods reasonably improved and normalized the contrast and illumination throughout the images, while the remaining methods attained a poor performance although they all have almost increased the AUC. Applying the adaptive local contrast enhancement to an illumination equalized image further enhanced the results, and achieved better results than being applied to the intensity component of an image as proposed in [5]. 6.3.4
Observations and Conclusion
This experiment compared nine different methods affecting the contrast of retinal fundus images. Two publicly available datasets of total 60 images along with a simple vasculature segmentation algorithm were used for the test purpose. Employing methods for illumination equalization solely attains poor results. However, since the hybrid proposed method achieved a slightly higher AUC than other available methods, we may conclude that combining methods for both illumination and contrast normalization achieves promising results. Chapter 6
Comparative Studies
91
a b c d e f g h i j 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 blood vessels. (i) Estimation of background luminosity and contrast variability. (j) Adaptive local contrast enhancement applied to ‘g’ instead of ‘a’.
92
Chapter 6
Comparative Studies
a d g j
b c e f h i k
Figure 6.4 (a)–(j) are the results of applying a RV segmentation method to the images Fig. 6.3 (a)–(j) respectively. (k) A manual segmentation of Fig. 6.3(a) (used as a gold-standard). Chapter 6
Comparative Studies
93
Figure 6.5 ROC curves of the compared contrast enhancements methods. Normalization Methods Intensity image Green band image Histogram equalization Adaptive local contrast enhancement Adaptive histogram equalization Desired average intensity Division by an over-smoothed version Background subtraction of retinal blood vessels Estimation of background luminosity and contrast variability Adaptive local contrast enhancement applied to the illumination equalized image obtained using division by an over-smoothed version
AUC 0.629 0.641 0.787 0.833 0.874 0.646 0.676 0.677 0.650 0.876
Table 6.2 Area under curve (AUC) measured per contrast enhancement method.
94
Chapter 6
Comparative Studies
6.4
A Comparative Study of Color Normalization Methods
6.4.1
Purpose
The purpose of this experiment is to compare some of the color normalization methods available in literature, and to conclude whether they can affect the process of segmenting the retinal vasculature from digital retinal fundus images. Color normalization methods have to effectively eliminate the variances in color among fundus datasets in order to reconsider color as useful information while detecting retinal landmarks and retinopathies. This section will compare and evaluate the effect of various color normalization methods while segmenting the RV. 6.4.2
Method
As we described before in section 2.7.4, Goatman et al. [22] used the chromaticity values (see Appendix D) distribution for evaluating color normalization methods. The Chromaticity coordinates were used to evaluate the effects of applying the three color normalization methods (Grey-world normalization, Histogram equalization, and Histogram Specification) to retinal images in order to discriminate different retinal lesions. The clustering of the chromaticity values representing four different retinal lesions was measured before and after applying each of the normalization methods (Figure 2.13). Instead of discriminating retinal lesions, we use the chromaticity values to discriminate vessels from non-vessels so as to evaluate the effect of the color normalization methods on automatic RV segmentation methods. The clustering of the chromaticity values representing vessels/non-vessels was measured before applying any of the normalization methods, and after applying each of the four normalization methods independently. The average – vessels/non-vessels – chromaticity values per each image were plotted, and Chapter 6
Comparative Studies
95
an ellipse was drawn centered on the mean value of each cluster as proposed in [22]. Principal components analysis (Hotelling Transform) is used to find the direction of maximum variance with which the major axis is aligned. The semi-major axis equals to two standard deviations in the major axis's direction, while the semi-minor axis equals to a standard deviation in the orthogonal direction. 6.4.3
Results
Figure 6.6 shows that before applying any of the normalization methods, there was a noticeable overlap with elements scattered in both clusters. After applying the color normalization methods the overlap persists, however, the elements in each cluster (especially the non-vessels) become more intense. A clear separation between clusters was found only after applying histogram equalization. Comprehensive normalization showed more intense clusters, and a narrower overlap compared to the gray-world normalization. Histogram specification recorded the worst results with a comprehensive overlap, although it showed the clearest separation of four retinal lesions clusters in [22]. 6.4.4
Observations and Conclusion
This experiment compared four different methods affecting the color of retinal fundus images. Two publicly available datasets of total 60 images along with the simple chromaticity color model were used for the test purpose. Histogram equalization showed the best results, though it enhances the blue component of the retinal image. Therefore, more experiments need to be done to measure accurately the effect of the histogram equalization method on analyzing the retinal fundus image. Finally, the experiment showed that ‘in general’ color normalization methods may possibly affect and/or improve the RV segmentation process. 96
Chapter 6
Comparative Studies
a b Figure 6.6 Color normalization chromaticity plots. (a) Before applying any normalization methods. (b) Gray-world normalization. Red ellipse and elements plots represent the non-vessels cluster, while the blue represent the vessels cluster. Chapter 6
Comparative Studies
97
c d Figure 6.6 Color normalization chromaticity plots (Continued). (c) Comprehensive normalization. (d) Histogram equalization. Red ellipse and elements plots represent the non-vessels cluster, while the blue represent the vessels cluster.
98
Chapter 6
Comparative Studies
e Figure 6.6 Color normalization chromaticity plots (Continued). (e) Histogram specification (matching). Red ellipse and elements plots represent the non-vessels cluster, while the blue represent the vessels cluster.
Chapter 6
Comparative Studies
99
Chapter
7
The Developed Automatic DR Screening System Components Following the literature reviews, and the comparative studies conducted, we present in this chapter our contribution of novel methods. Each section presents an experiment for applying a novel method to develop one of the components in our proposed automatic DR screening system (Figure 1.1). For all experiments, results are shown using the STARE and/or DRIVE publicly available datasets (see Appendix C). A list of the published papers covering the experiments can be found at the beginning of the document.
7.1
Optic Disc Localization by Means of a Vessels’ Direction Matched Filter
7.1.1
Purpose
Optic Disc (OD) detection is a main step while developing automated retinal screening systems, especially for detecting DR. We present in this section a novel method to automatically detect the position of the OD in digital retinal fundus images. The method starts by normalizing luminosity and contrast through out the image using illumination equalization and adaptive histogram equalization methods respectively. The OD detection algorithm is based on matching the expected directional pattern of the retinal blood vessels. At the end of this experiment, the proposed method is going to be evaluated using both the DRIVE and STARE. 100
Chapter 7
The Developed Automatic DR Screening Sys. Components
7.1.2
Method
This experiment, inspired by the work of Frank ter Haar [11], Hoover and Goldbaum [19], and Foracchia et al. [21], presents a method for the automatic detection of the OD. The proposed method comprises several steps. Initially a binary mask is generated. Then the illumination and contrast through out the image are equalized. Finally, the retinal vasculature is segmented, and the directions of the vessels are matched to the proposed filter which represents the expected vessels’ directions in the OD vicinity. Initially, and for mask generation, we used the method proposed by Frank ter Haar [19], who thresholded the image’s red-band, and then followed by three morphological operators to give the final ROI mask (see section 2.3). The method of Frank ter Haar was chosen since it was superior compared to other mask generation methods as shown in Experiment 1 (section 6.1). Then, to equalize the illumination across the image, the illumination equlaization presented by Hoover and Goldbaum [18] is applied using a running window of only one size (40×40) as proposed by [19]. As proposed by [19] once more, the ROI mask of the retinal images is shrunk by five pixels to discard the pixels near the border. We applied the illumination equalization to the green-band image since it displays the best vessels/background contrast. The final step in preprocessing the image, is applying the adaptive histogram equalization (AHE) proposed by Wu et al. [33] to enhance and normalize the contrast (see section 2.5.4). Before applying the AHE, the illumination equalized green-band image obtained from the previous step was inverted. AHE was chosen for contrast enhancement since it achieved the highest AUC when compared with other literature available methods as shown in Experiment 3 (section 6.3). To segment the retinal blood vessels, we used the simple and standard edge fitting algorithm proposed by Chaudhuri et al. [30], where the Chapter 7
The Developed Automatic DR Screening Sys. Components
101
similarity between a predefined 2D Gaussian template and the fundus image is maximized (see 4.3.1). And in order to generate a binary vessels/nonvessels image, the maximum responses are thresholded using the global threshold selection algorithm proposed by Otsu [63] (see Appendix E). Instead of applying the 12 templates to an averaged green-band image as presented by [30], applying them to the adaptively histogram equalized image significantly improves the segmentation algorithm as shown in Experiment 3, as it increases the AUC from 0.641 to 0.874. A vessels direction map (VDM) can be obtained from the segmentation algorithm by recording the direction of the template that achieved the maximum response at each pixel. Then, for all the pixels labeled as non-vessel, the corresponding values in the VDM can be assigned to ‘-1’ or not-a-number (NAN) in order to exclude them from further processing. ‘A matched filter describes the expected appearance of a desired signal, for purposes of comparative modeling’ [52]. Thus, in order to detect the OD, a simple vessels’ direction matched filter is proposed to roughly match the direction of the vessels at the OD vicinity (Figure 7.1). The 9×9 template is resized using bilinear interpolation to sizes 241×81, 361×121, 481×161, and 601×201 to match the structure of the vessels at different scales. These sizes are specially tuned for the STARE and DRIVE, but they can be easily adjusted to other datasets. The difference between all four templates (in the single given direction) and a VDM is calculated, and the pixel having the least accumulated difference is selected as the OD center (Figure 7.2(i)). To reduce the computational burden, matched filters are applied only to candidate pixels picked from the fundus image. The binary vessel/non-vessel image is thinned (Figure 7.2(f)). Hence, reducing the amount of pixels labeled as vessels into the vessels’ centerline. All remaining vessel-labeled pixels that are not within a 41×41 square centered on each of the highest 4% intensity pixels (Figure 7.2(g)) in the illumination equalized image are re102
Chapter 7
The Developed Automatic DR Screening Sys. Components
labeled as non-vessel pixels (Figure 7.2(h)). This final step aims only to reduce the number of OD candidates, and thus altering the size of the square or the amount of highest intensity pixels simply has no significant effect. The remaining vessel-labeled pixels are potential OD centers, thus selected as candidates for applying the four sizes of the matched filter. 135 150 165 165 0 15 15 30 45
120 135 150 165 0 15 30 45 60
105 120 135 150 0 30 45 60 75
105 105 120 135 0 45 60 75 75
90 90 90 90 90 90 90 90 90
75 75 60 45 0 135 120 105 105
75 60 45 30 0 150 135 120 105
60 45 30 15 0 165 150 135 120
45 30 15 15 0 165 165 150 135
Figure 7.1 The proposed “vessels’ direction at the OD vicinity” matched filter.
7.1.3
Results
The proposed method achieved a success rate of 98.77% (i.e. the OD was detected correctly in 80 out of the 81 images contained in the STARE dataset). The estimated OD center is considered correct if it was positioned within 60 pixels of the manually identified center, as proposed in [18] and [53]. The average distance (for the 80 successful images) between the estimated OD center and the manually identified center was 26 pixels. The only case in which the OD was not correctly detected (Figure 7.3(a)), was due to uneven crescent-shaped illumination near the border that biased the OD candidates and affected the vessels segmentation algorithm. OD detection results for the proposed and literature reviewed methods are summarized in Table 7.1. Besides, and for comparison, Figure 7.3(b)–(h) shows the OD detection results of the proposed method on the images shown in [53]. Additionally, the OD was detected correctly in all of the 40 DRIVE dataset images (a 100% success rate) using the proposed method (Figure Chapter 7
The Developed Automatic DR Screening Sys. Components
103
7.4). The average distance between the estimated and the manually identified OD centers was 17 pixels.
(a)
(b)
(c)
(d)
(e)
(f)
(g)
(h)
(i)
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
Chapter 7
The Developed Automatic DR Screening Sys. Components
a b c d e f g h 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].
A MATLAB® prototype was implemented for the procedure applying the matched filters, where runs needed on average 3.5 minutes for each image on a Laptop (2-MHz Intel® Centrino® 1.7 CPU and 512 Mb RAM). Though using 2D matched filters for retinal vessels segmentation and OD detection involves more computation than other literature reviewed methods, the use of such template-based methods eases the implementation on specialized high speed hardware and/or parallel processors. Chapter 7
The Developed Automatic DR Screening Sys. Components
105
Figure 7.4 Results of the proposed method using the DRIVE dataset (white cross
represents the estimated OD center).
7.1.4
Observations and Conclusion
Characterized as the brightest anatomical structure in a retinal image, selecting the highest 1% intensity pixels should contain areas in the OD [44], unfortunately, 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]. Illumination equalization significantly normalized the luminosity across the fundus image, increasing the number of OD candidates within the OD vicinity. Selecting the highest 4% intensity pixels guaranteed the presence of OD-candidates within the OD area. Moreover, using high intensity pixels to filter segmented vessels implicitly searches for areas of high variance, as proposed in [32]. However, employing intensity to 106
Chapter 7
The Developed Automatic DR Screening Sys. Components
the segmented vessels as proposed is much more robust compared to OD localization methods based on intensity variation or just on intensity values. Later methods will not be straightforward once the OD loses its distinct appearance due to the non-uniform illumination or pathologies. OD Detection Methods Sinthanayothin et al. [5], [32] – Highest average variation Walter and Klein [46] – Largest brightest connected object Chrástek et al. [47] – Highest average intensity Li and Chutatape [42], [44] – Brightness guided, PCA modelbased Osareh et al. [1], [10], & [48] – Averaged OD-images modelbased Lalonde et al. [45] – Hausdorff-based template matching, pyramidal decomposition & confidence assignment Frank ter Haar [19] – Resolution pyramid using a simple Haarbased discrete wavelet transform Frank ter Haar [19] – Pyramidal decomposition of both the vasculature & green plan Abdel-Ghafar et al. [15] – Circular Hough Transform Barrett et al. [49] – Hough transform as implemented by Frank ter Haar[19] (A) Barrett et al. [49] – Hough transform as implemented by Frank ter Haar[19] (B) Frank ter Haar [19] – Hough transform applied only to pixels on or close to the retinal vasculature Frank ter Haar [19] – Hough transform applied to the fuzzy convergence image Hoover and Goldbaum [18], [51] – Fuzzy convergence Frank ter Haar [19] – Searching for the branch with the most vessels Frank ter Haar [19] – Searching for the area with more pathendings Frank ter Haar [19] – Fitting the vasculature orientations on a directional model Foracchia et al. [53] – A geometrical model of the vessel structure using two parabolas Goldbaum et al. [55] – Blood vasculature convergence, OD's appearance, & the large vessels entering the OD Lowell et al. [43] – OD Laplacian of Gaussian template & full Pearson-R correlation Chapter 7
Local Dataseta 99.1% 100% 97.3%
STARE Datasetb 42.0% 58.0% -
99%
-
100%
58.0%
100%
71.6%
89%
70.4%
93.2%
54.3%
No quantitative results reported -
59.3%
-
67.9%
96.3%
71.6%
97.4%
65.4%
-
89.0%
78%
38.3%
88.5%
59.3%
99.5%
93.8%
-
97.5%
No quantitative results reported 99%
The Developed Automatic DR Screening Sys. Components
-
107
Tobin et al. [56] – Vasculature-related OD properties & a Bayesian classifier Abràmoff and Niemeijer [57] – Vasculature-related OD properties & a kNN regression Our Proposed Methodc – Vessels’ Direction Matched Filter
81%
-
99.9%
-
100%
98.8%
a
Local dataset: The success rates (percentages) shown for local dataset are taken from the corresponding references. b STARE dataset: The success rates (percentages) shown for the STARE dataset for reviewed methods are all taken from the comparative study accomplished by Frank ter Haar [19], except for those corresponding to the references [18], [51], and [53] (taken from the corresponding reference). c Our Proposed method result is shown in the last row (the highest four success percentages using the STARE are shaded in light gray). Table 7.1 OD detection results for the proposed and literature reviewed methods.
AHE transformed the image into a more appropriate appearance for the application of the segmentation algorithm. Though the 2D matched filter achieved poor results compared to other retinal vessels segmentation methods [50], applying it to the AHE image significantly improved its performance. A clear advantage of using the 2D matched filter for vessels segmentation is the ability to obtain the VDM implicitly while segmentation, without any additional algorithm as proposed in [56] and [57]. The proposed 2D vessels’ direction matched filter was successful and robust in representing the directional model of the retinal vessels surrounding the OD. Resizing the filter into four sizes aimed to capture the vertical longitudinal structure of the retinal vessels, as proposed by Tobin et al. [56], where the convolution mask was roughly 1-OD diameter wide and 3-OD diameters tall. The proposed filter is obviously simpler than the directional model proposed by [19], the parabolic geometrical model proposed by [53], and the fuzzy convergence approach proposed by [18]. In addition, using the STARE dataset that is full of tough pathological situations, the proposed method achieved better results compared to the reviewed methods (Table 7.1). 108
Chapter 7
The Developed Automatic DR Screening Sys. Components
7.2
Retinal Vasculature Segmentation using a Large-Scale Support Vector Machine
7.2.1
Purpose
Retinal vasculature (RV) segmentation is the basic foundation while developing retinal screening systems, since the RV acts as the main landmark from which further analysis proceeds. Recently, supervised classification proved to be more efficient and accurate when employed for the segmentation process. Also, novel features have been used in literature methods showing high separability between both vessels/non-vessels classes. This experiment, inspired by the work of Soares et al. [66], presents a method for the segmentation of the RV using the 2D-Gabor wavelet as the main feature. The pixel features are further improved by adding both Top-hat and Hessian-based enhancements. And finally, a large-scale support vector machine (LS-SVM) is used as the supervised classifier. More to the point, this experiment is concerned with developing fast (computationally efficient) methods while achieving high accuracy. The publicly available benchmark DRIVE dataset will be used for evaluating the performance of the proposed method. 7.2.2
Method
Our proposed method starts with preprocessing the retinal image by simply discarding the red and blue components in order to enhance the contrast. We do not need to apply any other normalization or enhancement method since our features extraction approaches highly enhances the vessels. The green band is then inverted, so that the vasculature appears brighter than the background. The images are finally padded (Figure 7.5(b)) in order to reduce the false detection of the ROI border. Padding is simply done by determining the background pixels near the exterior ROI border, and Chapter 7
The Developed Automatic DR Screening Sys. Components
109
replacing each one with the mean value of its neighbors [66]. This process is done repeatedly to expand the ROI to the required extent. A typical way to find a suitable representation (features) of data that will assist further analysis is to transform the signal using a set of basis functions. Wavelets process data at different scales (resolutions); therefore they have advantages over traditional Fourier transforms in analyzing signals. Soares et al. [66] employed a continuous wavelet transform Tψ(b, θ, a), where b, θ, and a are respectively the time-shift (translation), rotation, and dilation (scale) parameters. The mother analyzing wavelet ψb,θ,a was chosen to be a 2D-Gabor wavelet defined as:
1 2
ψ ( x) = exp( jko x) exp(− | Ax |2 )
(Eq. 7.1)
where A = diag[Є-1/2, 1], ε ≥ 1 is a 2×2 matrix defining the elongation in any direction. The value of the ε is set to 4, while ko = [0, 3]. And we have mentioned before in section 4.3.2, the wavelet transform was implemented using the fast Fourier transform, and the implementation is publicly available as open source MATLAB® scripts [67]. The final set of features used by [66] for classification included the inverted green channel and its maximum 2DGabor transform response for scales 'a' = 2, 3, 4, and 5 pixels, and θ spanning from 0° up to 170° at steps of 10° (Figure 7.5(c)–(f)). These scales 'a' were chosen to detect vessels of all possible widths. In [68], and in order to segment the vasculature, 5 methods (vessel maps) were used to enhance the separability between vessels and non-vessels. The Top-hat filter simply defined as the difference between an image and its morphologically closed version was used to equalize the background. Therefore, it decreased both the intra-background and the intra-vessel classes' variances. The Top-hat vessel map (Figure 7.5(g)) was the most 110
Chapter 7
The Developed Automatic DR Screening Sys. Components
separable single vessel map, followed by the Top-hat Hessian-based enhancement [68]. The Hessian matrix is the matrix of second partial derivatives of a function, thus if applied to an image, it describes the second order structure of the intensity variations in the neighborhood of each pixel. Analyzing the Hessian matrix can be used in detecting line structures; dark lengthy tubular structures on 2D bright backgrounds (i.e. vessels) are reflected by a Hessian Matrix having a large positive eigenvalue and a small eigenvalue of whatever sign [88], [89]. Again, to decrease the variance within the vessels class, Condurache and Aach [68] enhanced all vessels regardless to their size by calculating the highest eigenvalue of the Hessian matrix at 3 different scales of the Top-hat vessel map (Figure 7.5(h)). Condurache and Aach [68] made their vessel segmentation tool available at [69]; the tool can generate both Top-hat and Hessian-based vessel maps. Instead of applying the Hessian enhancement to the Top-hat output, we propose applying it directly to the green-band in order to achieve smaller intra-vessels class variance (Figure 7.5(i)). SVMs are statistical learning tools that rely on mapping the data from two categories to a sufficiently high dimension where they can always be separated by a hyperplane [90]. ‘While there may exist many hyperplanes that separate the two classes, the SVM classifier is based on the hyperplane that maximizes the separating margin between the two classes’[91]. The training examples defining this hyperplane (i.e. the borderline training examples) are called the support vectors; they are the most difficult patterns to classify, yet they are the most informative in the classification task [90]. Though the dimensionality of the transformed space in a SVM is typically much higher than the original feature space, the complexity of the classifier is mainly described by the number of support vectors. In addition, SVMChapter 7
The Developed Automatic DR Screening Sys. Components
111
based approaches are able to significantly outperform competing methods in many applications [91]–[93]. Training a SVM (i.e. the process of finding the optimal hyperplane) leads to a quadratic optimization problem which depends on the number of training examples. Therefore, and in order to deal with large training sets, LS-SVMs decompose the problem into a series of smaller tasks so as to have a tractable solution [94]. In our work, we used the publicly available MATLAB® toolbox for SVM and kernel methods [95]; the LS-SVM class was used with a Gaussian mapping function. 7.2.3
Results
In order to evaluate the performance of the presented method, we trained the LS-SVM using sets of pixels randomly selected from the 20 images in the DRIVE training set. The available manual segmentation served as our ground truth, and the available FOV masks were shrunk by 3 pixels in order to discard any border effect. Initially, for computational reasons, only 50 thousands pixels from the 20 images were randomly selected for training the LS-SVM. For the selected pixels, four experiments we done using different feature sets in order to determine the final features vector. The LS-SVM was trained initially using the inverted green-band and the 2D-Gabor responses at four scales (i.e. the training set used by [66]). Then the Top-hat enhancement was added to the previous set. In the third experiment, we added the Top-hat Hessian-based enhancement (i.e. the Hessian enhancement applied to the Top-hat response as proposed by [68]). Finally, beside the inverted green-band, 2D-Gabor responses, and the Top-Hat response, we added the Hessian-based enhancement applied to the greenband image. The final features set reported the highest area under the ROC curve (AUC) as shown in Table 7.2. 112
Chapter 7
The Developed Automatic DR Screening Sys. Components
(a)
(b)
(c)
(d)
(e)
(f)
(g)
(h)
(i)
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 2DGabor wavelet response for scales ‘a’ = 2, 3, 4, and 5 pixels respectively. (g) Tophat enhancement. (h) Top-hat Hessian-based enhancement. (i) Green-band Hessian-based enhancement. Chapter 7 The Developed Automatic DR Screening Sys. Components 113
Another set of experiments were made to determine the most effective size for the training set (i.e. number of pixels) in terms of computational time and accuracy. The initial 50 thousand pixels set is divided into five smaller sets (each of 10 thousands), then each of the new five sets is further divided into 5 new sets (each of only 2 thousands pixels). Each of the new 30 sets is used individually to train the LS-SVM, and the average AUC and computational time among equal sets is calculated. Table 7.2 shows a dramatic drop in both training and classification (Figure 7.6) time as the training set decreased, while the AUC slightly increased (Figure 7.7). The experiments were carried out using a Laptop (2-MHz Intel® Centrino® 1.7 CPU and 512 Mb RAM), while the results of Staal et al. [7] and Soares et al. [66] are obtained using “a Pentium-III PC, running at 1.0 GHz with 1-GB memory”, and “an AMD Athlon XP 2700+ PC (2167 MHz clock) with 1-GB memory” respectively. For all experiments, each training set is equally divided between both vessels/non-vessels classes. 7.2.4
Observations and Conclusion
Supervised segmentation is recently the method of choice when segmenting the RV, since it can simply adapt by using a suitable pre-labeled training set, and it also achieves higher performances [50]. Yet, a disadvantage of supervised methods is the need for manual segmentations that may take 2 hours in average for a single image [7]. Though the method proposed by Soares et al. [66] reported slightly higher AUC than our presented method, Table 7.2 shows that LS-SVM takes significantly lower time for training (about 4 sec. compared to the 9 hours). In addition, LSSVMs can be trained using only 2000 training pixels, which are again significantly below the 1 million pixels used by Soares et al. [66]. Therefore, our presented method is much more suitable when it comes to adapting with new datasets. LS-SVMs trained using the 2000 pixels’ sets are not so simple 114
Chapter 7
The Developed Automatic DR Screening Sys. Components
that it can not explain the differences between both vessels/non-vessels classes, yet they are not so complex to give poor classification results on novel patterns (the case of overfitting). As a result, LS-SVMs tend to perform well (i.e. generalize) when applied to data outside the training set.
Figure 7.6 (1st row) 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 manual segmentation. (3rd row) Another manual segmentation available for the DRIVE images. Chapter 7
The Developed Automatic DR Screening Sys. Components
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.
The results show that the method presented in this experiment is more computationally efficient in classification time than the method proposed by Staal et al. [7], and comparable to Soares et al. [66]. Regarding the features set, Table 7.2 shows that adding both the Top-hat and Hessian-based enhancements to the set of features selected by Soares et al. [66] improved the performance, thus the AUC from 0.9329 to 0.9449. In addition, applying 116
Chapter 7
The Developed Automatic DR Screening Sys. Components
the Hessian-based enhancement to the Top-hat response as proposed by [68] reported no improvement in classification results. Conversely, applying the Hessian-based enhancement to the green-band images enhanced the AUC from 0.9374 to 0.9449. Size of training set Segmentation ( number methods of pixels for training)
LS-SVM (our presented method)
Pixels features vector
Inverted green50 band, & 2D-Gabor thousands responses
1
-
0.9329
Inverted greenband, 2D-Gabor 50 thousands responses, & Tophat response
1
-
0.9374
Inverted greenband, 2D-Gabor 50 responses, Top-hat thousands response, & Tophat Hessian response 50 thousands 10 thousands
Inverted greenband, 2D-Gabor responses, Top-hat response, & Green-band 2 thousands Hessian response Staal et al. [7] Soares et al. [66]
Average Average Number classification Average Max. training of sets time (per AUC AUC time (runs) image)
Convex set regions based on ridge detection Inverted green1 million band, & 2D-Gabor responses -
1
-
0.9374
1
-
0.9449
2 hours
10 min.
5
0.9481 0.9503
1.75 min.
2.39 min.
25
0.9505 0.9537
3.75 sec.
30 sec.
-
≈15 min.
-
-
0.9520
-
-
0.9614 9 hours
10 sec.
Table 7.2 Results for the performance evaluation experiments made for our
presented method, and compared to different literature segmentation methods (for the DRIVE dataset). Chapter 7
The Developed Automatic DR Screening Sys. Components
117
7.3
Hard Exudates Detection using a Large-Scale Support Vector Machine
7.3.1
Purpose
Hard exudates (as one of the main bright lesions) detection is the main target of any retinal screening system; since those lesions are main manifestations of diabetic retinopathies, and can be sight-threatening if not treated early. Recently, various approaches have been employed for bright lesions detection in general, including statistical learning and computational intelligence techniques. In addition, supervised classification proved to be efficient in detecting and differentiating different forms of bright lesions. In order to take advantage of the efficient RV segmentation algorithm presented in the previous section, this experiment aims to detect hard exudates based on the output of the RV segmentation algorithm. The publicly available benchmark STARE dataset will be used for evaluating the performance of the proposed method. 7.3.2
Method
When examining the results of the LS-SVM for the RV segmentation (Figure 7.6(1st row)), we found that it shows a somehow significant response to areas of bright lesions and OD (the responses of the bright lesions and OD regions tend to be very smooth ‘without the tiny scratches covering the rest of the retinal background’, and very dark ‘almost black’). Therefore, we used the LS-SVM response for RV segmentation as our basis for detecting the presence of hard exudates. The first step while detecting hard exudates is to obtain a binary RV segmentation from the LS-SVM soft-output. This binary version can be achieved by initially performing a top-hat filtering to the soft-output using a disc structuring element with a radius of 10 pixels. Then the top-hat filtered 118
Chapter 7
The Developed Automatic DR Screening Sys. Components
soft-output is binarized using the Otsu [63] global image threshold (see Appendix E). The top-hat filtering enhances the vessels structure before applying the global thresholding method. Using this binary segmentation (Figure 7.8), and using an illumination equalized green-band image ([18, 19], See section 2.4), the average intensity of the RV per image is calculated. The second step in our approach is to segment the LS-SVM soft-response into regions that will be further examined to determine whether they constitute parts of the hard exudates or not. The watersheds algorithm [4] is used for regions segmentation; the idea of watersheds is based on viewing an image in 3-dimensions where a pixel may belong to 1) A local regional minimum, 2) A location where it will fall with certainty to a single minimum, or finally 3) A location where it may fall to more than one regional minimum. Pixels belonging to the final set constitute crest/watershed lines (dams) dividing this topographic surface into regions. Since the watersheds approach is sensitive to edges, the closing morphological operator [4] with a disc structuring element of radius 2 pixels is applied to the inverted LS-SVM soft-response before watersheds segmentation (Figure 7.8). Finally, after the soft-response image was segmented (Figure 7.9), properties of the regions were investigated in order to finally detect the exudate regions. Regions selected as hard exudates must 1) have a size more than or equal to 10 pixels, 2) more than half the region’s area coincides with the highest 5% intensity pixels from the illumination equalized green-band, and 3) the average intensity within the region must be greater than or equal to 1.35 times the average intensity of the RV segmented from the image. The average intensities for both the region and the RV are calculated from the illumination equalized green-band. Finally, all pixels within 120 pixels from the OD center are discarded even if they form a part of a selected region. Chapter 7
The Developed Automatic DR Screening Sys. Components
119
Figure 7.8 The first steps while detecting Exudates. (1st row) Typical STARE
images containing exudates. (2nd row) The LS-SVM “RV segmentation” softresponses for the 1st row images. (3rd row) The binarization “hard-responses” of the RV segmentation outputs in the 2nd row. (4th row) The effect of the morphological closing when applied to inverted versions of the images in the 2nd row. 120 Chapter 7 The Developed Automatic DR Screening Sys. Components
Figure 7.9 The final steps while detecting Exudates. (1st row) The watersheds
segmentation results of the images in ‘Figure 7.8 (4th row)’. (2nd row) 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 ‘Figure 7.8 (1st row)’.
7.3.3
Results
To evaluate the performance of the presented method, we used the 81 images available in the STARE dataset as our test set, since these images show much more manifestations of retinopathies when compared to the DRIVE dataset. But since the available benchmark datasets do not include gold standards (i.e. manual segmentations) for retinopathies, we can not Chapter 7
The Developed Automatic DR Screening Sys. Components
121
show the results of our algorithm in the form of the ROC curve. Alternatively, we will simply indicate whether an image probably includes exudates or not, and compare our indication to the diagnosis available with the STARE dataset in addition to the diagnosis of an ophthalmologist. To do so, the 81 STARE images were divided into three sets; a set of 10 images containing hard exudates (the main aim of this experiment), a set of 48 images containing no signs of hard exudates or any other form of bright lesions, and finally a set of 33 images containing different forms of bright lesions (e.g. soft exudates, cottonwool spots, drusen, Circinate retinopathy, laser marks … etc.), the final set includes also the 10 images diagnosed as containing hard exudates. The generated FOV masks were shrunk by 20 pixels in order to discard any border effect before applying the proposed algorithm. From the 10 images diagnosed as having hard exudates, the system successfully indicated so for all images, thus with an accuracy of 100% (Table 7.3). As for the images not containing any form of bright-lesions, the system indicated so in 43 of the total 48 images (Figure 7.10), thus with an accuracy of ≈ 90%. Finally, for images containing all different forms of bright-lesions, the system successfully indicated that in 21 out of the 33 images (Figure 7.11), thus with an accuracy of ≈ 64%. Ground-truth (Manual Diagnosis) Containing Hard Exudates Containing BrightLesions No Bright-Lesions Contained
System’s Output Possibly Containing Hard Exudates
No Bright-Lesions Contained
Accuracy
10
-
100%
21
12
64%
5
43
90%
Table 7.3 Results of detecting Hard Exudates using the STARE.
122
Chapter 7
The Developed Automatic DR Screening Sys. Components
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).
Figure 7.11 STARE images manually diagnosed as having forms of bright-lesions
other than hard exudates, and indicated as having bright lesions by our proposed approach. The identified bright lesions are shown in blue (follow the white arrows), and superimposed on the original images.
7.3.4
Observations and Conclusion
We proposed in this last experiment a method for automated detection of hard exudates as one of the main symptoms of DR. The method used the output of the LS-SVM trained for the RV segmentation, and thus it saves time and computational effort since it can be considered a natural consequence of the RV segmentation phase. The method is very effective in detecting hard exudates (100% accuracy), and with an acceptable false alarm rate (≈ 10%). In addition, the system shows robustness in detecting other forms of bright lesions with an accuracy of 64%, which can again be Chapter 7
The Developed Automatic DR Screening Sys. Components
123
considered a byproduct of the hard exudates detection approach, and shows a plenty of room for improvement.
124
Chapter 7
The Developed Automatic DR Screening Sys. Components
Chapter
8
Conclusions & Future Work The classification of various ocular/ophthalmic – eye related – patterns is an essential step in many fields. The work proposed in this dissertation presented the main modules of any ocular screening system, especially for screening diabetic retinopathies. In these couple of pages we will conclude and suggest further works concerning every module we tackled through this dissertation.
8.1
Preprocessing of Digital Retinal Fundus Images
The dissertation presented different categories of preprocessing methods for retinal fundus images. Comparative evaluation measures among different methods, in each category, were conducted using publicly available datasets. Initially, three mask generation methods were tested using 121 images, and the method applying morphological operators to the thresholded red band was shown superior to other methods. Illumination equalization methods noticeably improved illumination across fundus images, and thus improved the process of localizing successful OD candidates (success rate of 98.35%). Four color normalization methods were tested (using 60 images), and the histogram equalization was found to be the most effective method in clustering the vessels and non-vessels pixels. And finally, the dissertation presented and compared nine different methods for contrast enhancement of retinal fundus images. Sixty images from the two publicly available datasets along with a simple vasculature segmentation algorithm were used for the Chapter 8
Conclusions & Future Work
125
test purpose. The adaptive histogram equalization was found to be the most effective method as it obviously improved the segmentation algorithm, thus it attained the largest area under the ROC curve. The adaptive local contrast enhancement and the typical histogram equalization methods recorded respectively the second and third best results, noticeably improving the RV segmentation algorithm. In general, the three mentioned methods reasonably improved and normalized the contrast throughout the images, while the remaining methods attained a poor performance although they have almost increased the AUC. In addition, applying the adaptive local contrast enhancement to an illumination equalized image further enhanced the results, and achieved better results than being applied to the intensity component of an image as proposed in [5]. An extension for this study could be investigating other preprocessing (contrast, illumination, and color normalization) methods available in literature (e.g. [83]). In addition, more comprehensive results can be achieved by examining the effect of the presented preprocessing methods using actual retinal-anatomy detection modules, or by using image quality assessment methodologies. Besides, a more inclusive comparison can be achieved when using larger datasets of actual screening images for evaluation. As for contrast enhancement in particular, more comprehensive results can be achieved by examining the performance of the existing methods using other vasculature segmentation algorithms (e.g. [7], [26], [50] and [52]).
8.2
Retinal Landmarks Classification
The dissertation presented a simple method for OD detection using a 2D vessels’ direction matched filter. The proposed approach achieved better results compared to the results reported in literature. An extension for this study could be: 126
Chapter 8
Conclusions & Future Work
1) Enhancing the performance of the vessel segmentation algorithm which will significantly affect the performance and efficiency of the proposed method. This can be achieved by employing other preprocessing techniques, or by employing post-processing steps as proposed in [32]. 2) Using other OD properties or vascular-related OD properties beside intensity and variance to reduce the OD center candidates will further enhance the performance (e.g. vessels’ density and diameter). 3) Examining the performance of the existing OD detection methods using other large, benchmark, and publicly available datasets to achieve more comprehensive results. 4) Choosing other vessels segmentation algorithms where the VDM can be implicitly obtained (e.g. the angles of the maximum 2D-Gabor wavelet response used in [66] or by our proposed RV segmentation method can be simply used as a VDM). In this dissertation we also presented a RV segmentation method employing a large-scale SVM as a supervised classifier. The classifier was trained with only 2000 pixels instead of 1 million pixels that were previously needed. The features vector for each pixel included the inverted green-band intensity value, the 2D-Gabor wavelet response at four different scales, the Top-hat transform response, and the Hessian-based response when applied to the green-band image. Using the DRIVE dataset, the proposed classifier with the set of features achieved an AUC of 0.9537. The recorded results are comparable to those found in literature. Moreover, the presented method proved to be computationally more efficient since it recorded about 4 sec average training time. Although supervised segmentation becomes the method of choice for segmenting the RV, the need for manual segmentations takes 2 hours in average for a single image. Chapter 8
Conclusions & Future Work
127
An extension for this study could be investigating other feature sets, and testing Large-Scale SVMs with different parameters. In addition, more comprehensive results can be achieved by evaluating the performance using other datasets and evaluation criteria
8.3
Diabetic Retinopathies Classification
Finally, the last module presented in this dissertation is the work concerning the detection of hard exudates. The proposed approach was based upon the results of the Large-Scale SVM soft response on segmenting the RV. The soft-response was inverted, closed, and segmented using the watersheds into regions where the properties of each region determined whether the region was a part of a hard exudate or not. The algorithm achieved an accuracy of 100% while detecting hard exudates, and 90% while detecting images free of any bright-lesions. Moreover, when applying the algorithm to images containing various forms of bright-lesions, the proposed algorithm indicated the presence of the lesions with an accuracy of 64%. An extension for this final study could be investigating other regions’ properties upon which exudates from non-exudates regions are determined. In addition, the proposed algorithm shows a potential room-for-improvement concerning the detection of retinal bright-lesions.
128
Chapter 8
Conclusions & Future Work
Appendix
A
Eye‐Related Images This appendix presents a primary overview of the most widely used eyerelated images. The majority of these images are used in the field of medical image processing.
A.1
Fundus Images
Invention of the ophthalmoscope by Hermann von Helmholtz in 1850, gave clinicians for the first time the ability to view the retina in the living patient, flourishing knowledge and understanding of many eye conditions. Two years later, introducing the indirect ophthalmoscope by Ruete allowed a stereoscopic and wider view of the fundus. Viewing retinal periphery was now accessible to the ophthalmologist, even through slightly hazy media [96]. Periphery Optic nerve Figure A.1
Front view of a healthy retina. [97]
Macula Fovea Capillaries
Stereoscopic fundus photography (Figure A.1) was pioneered by Jackson & Weber in 1886. The prototype camera was fixed to the patient’s head, and Appendix A
Eye-Related Images
129
a 2.5-minute exposure was used. Although only the largest details of the retinal anatomy and pathology were shown, it was revolutionary for allowing photographic documentation of retinal findings [96]. Recently, for an automated diabetic retinopathy screening system, a non-mydriatic fundus camera has been proposed which does not require pupillary dilatation, its operator does not need to be skilled at ophthalmoscopy, and it can be relatively easily transported [99].
A.2
Fluorescein Angiograms
a b
Figure A.2 (a) An example fundus image shows "Puckering the macula" – a
macular disease – were an opaque membrane obscures the visibility of the macula and drags the para-macular vessels. (b) The dragged vessels are shown better in fluorescein angiogram. [98]
Fluorescein angiography is a test which allows the blood vessels in the back of the eye to be photographed as a fluorescent dye is injected into the bloodstream via your hand or arm (Figure A.2). It is particularly useful in the management of diabetic retinopathy and macular degeneration. The test is done to help the doctor confirm a diagnosis, to provide guidelines for treatment, and to keep a permanent record of the vessels at the back of the eye. Although commonly referred to as fluorescein, the dye used in angiography is fluorescein sodium – the sodium salt of fluorescein – which is 130
Appendix A
Eye-Related Images
a highly fluorescent chemical compound that absorbs blue light with fluorescence [99]. Lesions were also best visualised on a fluorescein angiogram, and so many different methods were developed to detect microaneurysms using such images [99]. However, the fluorescein angiograms are not an ideal method for an automatic screening system since they require an injection of fluorescein into the body. This disadvantage makes the use of colour fundus images, which do not require an injection of fluorescein, more suitable for automatic screening [99].
A.3
Indocyanine Green Dye
In the 1990’s, the introduction of high resolution digital imaging systems allowed the addition of contrast studies using Indocyanine Green (ICG) dye. Used for years in cardiac blood flow studies, the infrared fluorescence spectrum of the dye had previously limited its practical use in the eye. The enhanced sensitivity of the digital camera to the infrared wavelength permitted real-time observation and recording of indocyanine green dye transit through the ocular circulation. The infrared fluorescence of ICG highlights the circulation of the choroid providing enhanced view of the deeper vascular structures, images which are complementary to those produced by fluorescein [96]. Again, due to the need of injection, color fundus images are more suitable for automatic screening.
A.4
Hartmann-Shack Images
Wavefront aberrometry, an innovative technique for measuring the complete refractive status of an optical system, has recently been introduced more widely into the clinical field. Wavefront aberration can be simply understood by passing light through the pupil. Light can be though of as a bundle of arrays, or as a virtual surface perpendicular to these arrays, called Appendix A
Eye-Related Images
131
the wavefront. The shape of this surface is the "wavefront aberration function" and in an optically perfect eye takes that form of a circular plane wave of similar size to the pupil (Figure A.3 (a)). The aberrations of the eye – wavefront error – are revealed by comparing the shape of the wavefront emanating from the eye to this conceptualised ideal wavefront [100]. Figure A.3
a b
(a) Typical Hartmann-Shack spot pattern (inverted) from a human eye‘s measurement. (b) Hartmann-Shack wavefront sensor with micro-lens array and image sensor in the focal plane. [101]
In ophthalmic clinical practice the most widely technique developed to evaluate the eye's optical aberrations is the objective aberrometer based on the Hartmann-Shack principle [100]. The Hartmann- Shack (H-S) wavefront sensor (Figure A.3 (b)) was developed out of a need to solve a problem that was posed, in the late 1960s, to the Optical Sciences Center (OSC) at the University of Arizona by the US Air Force. They wanted to improve the images of satellites taken from earth [102]. 132
Appendix A
Eye-Related Images
H-S images of the human ocular globe – obtained by sensing and registering a wavefront coming from an eye that is illuminated with a thin laser beam – are to be input into a system capable of recognizing the presence of a refractive error and outputting a measurement of such an error [103]. Thus, we can develop a practical system for diagnosing and measuring ocular refractive errors in the human eye (Astigmatism, Hypermetropia and Myopia) [104]. Unfortunately, large or public H-S image databases are not available.
A.5
Iris Images
The common procedures now employed for proving one’s identity – the use of passwords, personal identification numbers, picture IDs and so forth – have given way in many situations to automated biometric analysis. One such system takes advantage of the detailed patterns within a person’s iris (Figure A.4), which make it possible to identify someone using nothing more than an infrared image of the eye and a suitably programmed computer to process the information.
Figure A.4
A typical iris image. [105]
The iris has the great mathematical advantage that its pattern variability among different persons is enormous. In addition, as an internal (yet externally visible) organ of the eye, the iris is well protected from the environment and stable over time [106], unlike retinal vein patterns (an Appendix A
Eye-Related Images
133
alternative biometric) which can change over time [107], and are vulnerable to various retinopathies. Consequently, iris patterns are employed at the United Arab Emirates (7 airports), screening of all arriving passengers against a 544,000 expellees watch-list (3,428,000 passengers screened; 840 billion iris comparisons done; 22,634 catches; no unconfirmed matches) [108].
134
Appendix A
Eye-Related Images
Appendix
B
Diabetes, Diabetic Retinopathy and their Prevalence in Egypt B.1
Diabetes Mellitus
Diabetes mellitus is a group of metabolic diseases characterized by high blood-sugar (glucose) levels, which result from defects in insulin secretion, or action, or both. Diabetes mellitus, commonly referred to as diabetes, means "sweet urine" since elevated levels of blood glucose (hyperglycemia) lead to spillage of glucose into the urine. Normally, blood glucose levels are tightly controlled by insulin – a hormone produced by the pancreas – that lowers the blood glucose level when it elevates (for example – after eating – food insulin is released from the pancreas to normalize the glucose level). In patients with diabetes mellitus, the absence or insufficient production of insulin causes hyperglycemia. Diabetes mellitus is a chronic medical condition, meaning it can last a lifetime. Over time, diabetes mellitus can lead to blindness, kidney failure, and nerve damage. And it is also an important factor in accelerating the hardening and narrowing of the arteries (atherosclerosis), leading to strokes, coronary heart diseases, and other blood vessel diseases. Diabetes is a disease that affects about 5.5% of the population worldwide, a number that can be expected to increase significantly in the coming years [7]. The most important forms of diabetes are: Appendix B
Diabetes, Diabetic Retinopathy & their Prevalence in Egypt 135
•
Type 1 diabetes (the first recognized form): A severe, chronic form of diabetes caused by insufficient production of insulin and resulting in abnormal metabolism of carbohydrates, fats, and proteins. The disease, which typically appears in childhood or adolescence, is characterized by increased sugar levels in the blood and urine, excessive thirst, frequent urination, acidosis, and wasting. Also called insulin-dependent diabetes.
•
Type 2 diabetes (the more common form): A mild form of diabetes that typically appears first in adulthood and is exacerbated by obesity and an inactive lifestyle. This disease often has no symptoms, is usually diagnosed by tests that indicate glucose intolerance, and is treated with changes in diet and an exercise regimen. Also called non-insulin-dependent diabetes.
B.2
Diabetes in Egypt
Major socio-demographic changes have occurred in the Eastern Mediterranean Region During the past 20 years ( the total population of the Region has almost doubled, infant and childhood mortality/crude death rates have decreased although birth rate has remained high, Life expectancy has improved dramatically, urbanization has occurred, and per capita income has increased) [8]. Associated changes in physical activity and dietary patterns have occurred and promoted the development of non-communicable diseases. The total number of persons with diagnosed and undiagnosed diabetes in Egypt will increase from 3.24 million in 1995 (9.9%) to 8.809 millions by the year 2025 (over 13% of the population ≥ 20 years of age) [8]. The estimates used conservative projections of the diabetes prevalence rates resulted from a study conducted (between 1992 and 1994) by the Egyptian Ministry of Health and Population and the United States Agency for 136
Appendix B
Diabetes, Diabetic Retinopathy & their Prevalence in Egypt
International Development (Table B.1). The study – which was conducted because clinical experience suggested that diabetes was an emerging problem in Egypt – gathered information about the prevalence of diabetes risk factors, diagnosed diabetes mellitus and previously undiagnosed diabetes in the population 20 years of age and older by age, sex, residence and socioeconomic status (SES) [8]. Year
Sex
Age (years)
Male 1995 Female
2000 Female
2025 Female
Rural
557 732 378 061 86 130
145 202 210 471 50 578
702 934 588 532 136 708
11 592 000 3 738 000 1 202 000
6.1 15.7 11.4
20-44 45-64 65+
348 628 602 904 304 279
231 084 250 663 73 832
579 712 853 587 378 111
10 817 000 3 860 000 1 446 000
5.4 22.1 26.1
2 277 734
961 830
3 239 564
32 655 000
9.9
20-44 45-64 65+
648 653 471 931 103 765
152 121 247 043 55 848
800 774 718 974 159 613
12 845 000 4 639 000 1 403 000
6.2 15.5 11.4
20-44 45-64 65+
394 847 737 147 363 687
252 667 287 760 85 768
647 514 1 024 907 449 455
12 002 000 4 686 000 1 680 000
5.4 21.9 26.8
2 720 030
1 081 207
3 801 237
37 255 000
10.2
20-44 45-64 65+
1 361 271 1 331 913 382 450
184 926 388 474 106 845
1 546 197 1 720 387 489 295
19 938 000 9 776 000 3 681 000
7.8 17.6 13.3
20-44 45-64 65+
843 414 2 070 309 1 221 145
316 573 438 577 156 214
1 159 987 2 508 886 1 377 359
19 233 000 9 388 000 4 109 000
6.0 26.7 33.5
7 210 502
1 591 609
8 802 111
66 125 000
13.3
Total Male
Urban
Diabetes prevalence
20-44 45-64 65+
Total Male
Total
Total population
Number of cases of diabetes
Total
Table B.1 Projected counts and prevalence of diabetes in Egypt (population ≥ 20
years), 1995 to 2025. [8]
Direct and indirect costs of diabetes mellitus constitute a great burden on the economy of Egypt; a national program cannot provide free care for all Appendix B
Diabetes, Diabetic Retinopathy & their Prevalence in Egypt 137
Egyptians with diabetes. Total 1990 direct costs for treatment of diabetes were approximately US$74.3 million. Indirect costs were estimated at US$11.8 million per year for the effects of absenteeism on productivity while indirect costs due to premature death were not estimated [109].
B.3
Diabetic Retinopathy
Diabetic retinopathy is recognized as one of the most prevalent complications of diabetes (both type 1 and type 2) that affects the blood vessels of the retina [9]. It is a silent disease and may only be recognized by the patient when the changes (Figure B.1) in the retina have progressed to a level, that treatment is complicated and nearly impossible [1]. In addition, it is one of the leading causes of blindness especially in economically advanced countries [9]. Functionally, blindness is defined as loss of vision sufficient to prevent an individual from being self-supporting in an occupation, making that individual dependent on other persons, agencies or devices in order to live. Efforts to treat and prevent blindness must be made by all available means. The WHO estimates that 75% of cases of blindness in developing countries are avoidable [6]. Even though, the rates of blindness in developing countries are often 10–20 times greater than those in industrialized countries, because of the absence of preventive measures, inadequacy of health services and the high prevalence of major blinding diseases.
a b Figure B.1 Diabetic retinopathy
effect on the vision (a) Without retinopathy. (b) With retinopathy. [1] 138
Appendix B
Diabetes, Diabetic Retinopathy & their Prevalence in Egypt
Diabetic retinopathy is a progressive retinal damage that takes place as a result of chemical changes caused by diabetes that affect retinal capillaries (the smallest blood vessels linking arteries to veins) [1]. Retinopathy syndromes are generally categorized as simple (non-proliferative) or proliferative (Figure B.2). Simple retinopathy may impair vision if the lesions extend into the macular region, and may lead to serious loss of vision. Proliferative retinopathy, which is the most advanced, severe, and serious complication of diabetic ophthalmology, carries a high risk of blindness. Most of the retinal screening systems (e.g. [1, 5]) are not designed to detect proliferative retinopathy which probably never occurs in the absence of non-proliferative retinopathy. In people with type 1 diabetes, the prevalence of retinopathies is higher and the severity is greater than in those with type 2. The prevalence of diabetic retinopathy shows wide variations between countries. In type 1, it ranges from 14% (India) to 80% (Finland) and in type 2 it ranges from 17% in Switzerland to 52% in the United Kingdom [9].
B.4
Diabetic Retinopathy in Egypt
Randomly chosen diabetic patients attending a diabetic clinic of the Health Insurance Organization (HIO) in Alexandria – Egypt were studied to assess their behaviours in relation to management of their disease [110]. And though the study indicates that the majority of diabetics (97.0%) periodically attended routine check-ups, the study found that the majority of patients (62.3%) exhibited poor behaviour with regard to having eye examinations and did not attend eye screening (Table B.2) although loss of vision from diabetic retinopathy can be prevented in the majority of patients provided that the condition is detected early. Moreover, a study who aimed to describe the level of knowledge of diabetic patients about the disease [111] found that a majority of diabetic Appendix B
Diabetes, Diabetic Retinopathy & their Prevalence in Egypt 139
Diabetic Retinopathy (DR) - DR is generally categorized as simple (Non-Proliferative) or Proliferative according to the presence of clinical lesions indicating leakage or occlusion, and their position in the eye. Non-Proliferative DR is further classified into Background Retinopathy, and Maculopathy.
Non-Proliferative – Simple – Retinopathy - May impair vision or lead to serious loss of vision if the lesions extend to the macular region.
Background Retinopathy - It's the form of DR that occurs due to microvascular leakage away from the macular region. It has no noticeable signs; hence, patients are not alerted until it progresses into more severe levels. - It's clinically revealed by the presence of: • Sacculations from the capillary walls (Microaneurysms), • Blood (Retinal Haemorrhages), • Lipid Exudates (Hard Exudates), and • Retinal Oedema.
Maculopathy - It's the form of DR that occurs when the previously stated background retinopathy lesions take place in the macular region. - It results in blurred vision; the most common cause of decreased vision in patients with non-proliferative DR. - Microvascular occlusion may infrequently occur at the macula; consequently, vision is harmed due to insufficient blood supply.
Proliferative Retinopathy - It's the form of DR that occurs due to the presence of wide-ranging areas of microvascular occlusion throughout the retina which depends on those vessels for its nutrition; therefore, retina releases a vasoproliferative factor stimulating the growth of new abnormal blood vessels where the normal capillaries have already closed. - It may result in bleeding into the cavity of the eye and scarring with loss of vision. Figure B.2 Diabetic retinopathy syndromes as categorized by [1] (the light grey-
blocks point out DR forms that are not detected by this work.) 140
Appendix B
Diabetes, Diabetic Retinopathy & their Prevalence in Egypt
patients (90.0%) had poor knowledge about the disease and most of the patients (83.7%) had poor knowledge about the complications associated with diabetes. Over two-thirds (77.7%) of the patients were able to name one or more complications of diabetes, and of those who named at least one, 51.0% stated retinopathy. A significant linear association was observed between knowledge about diabetic retinopathy and patient's attendance at eye screenings. Those who lacked knowledge tended to behave poorly while those who were aware of the disorder were more likely to be classified as having very good or satisfactory behavior [111]. Seeking medical care Medical check-up Poor Satisfactory Very Good
No.
1 8 291
%
No.
%
0.3 2.7 97.0
Blood sugar test Poor Satisfactory Very Good
2 10 288
0.7 3.3 96.0
Eye examination Poor Satisfactory Very Good
187 91 22
62.3 30.3 7.3
Urine sugar test Poor Satisfactory Very Good
6 11 283
2.0 3.7 94.3
Blood pressure measurement Poor Satisfactory Very Good
106 78 116
35.3 26.0 38.7
Seeking medical care
Table B.2 Distribution of the 300 diabetic patients by their seeking of medical care.
[110]
Appendix B
Diabetes, Diabetic Retinopathy & their Prevalence in Egypt 141
Appendix
C
Fundus Photography Datasets C.1
DRIVE Database
The Digital Retinal Image for Vessel Extraction (DRIVE) database has been established to facilitate comparative studies on segmentation of blood vessels in retinal images [7, 24, and 50]. The data included in this database can be used, free of charge, for research and educational purposes. The database consists of a total of 40 colour fundus photographs that have been randomly selected from a diabetic retinopathy screening population in the Netherlands (453 diabetic subjects between 31-86 years of age). Of the 40 images, 33 photographs do not show any sign of diabetic retinopathy and 7 show signs of mild early diabetic retinopathy (namely, exudates, hemorrhages and pigment epithelium changes). All of the images contained in the database were actually used for making actual clinical diagnoses, and they have been deidentified (they were stripped from all individually identifiable information and processed in such a way that this information cannot be reconstructed from the images). The images are in compressed JPEG-format, which is unfortunate for image processing but is commonly used in screening practice. They were acquired using a Canon CR5 non-mydriatic 3CCD camera with a 45 degree field-of-view (FOV). Each image was captured using 8 bits per color plane at 768×584 pixels (Figure 2.3(a)). The FOV of each image is circular with a
142
Appendix C
Fundus Photography Databases
diameter of approximately 540 pixels (for each image, the FOV mask (Figure 2.2) is included) [24, 50]. The set of 40 images has been divided into two sets; a training and a test set, both containing 20 images. For the training set, a single manual segmentation of the vasculature is available, while two manual segmentations are available for the test set (one is used as gold standard, the other one can be used to compare computer generated segmentations with those of an independent human observer). All human observers – the first and second observer of [50] and a computer science student – that manually segmented the vasculature were instructed and trained by an experienced ophthalmologist (the last author of [50]). They were asked to mark all pixels for which they were for at least 70% certain that they were vessel.
C.2
STARE Project Database
The STructured Analysis of the Retina (STARE) Project [27] was conceived and initiated in 1975 by Michael Goldbaum, M.D., at the University of California, San Diego, and has been funded continuously by the National Institutes of Health (U.S.A.) since 1986. During this time, over thirty people have contributed to the project, with backgrounds ranging from medicine to science to engineering. Images and clinical data have been provided by the Shiley Eye Center at the University of California, San Diego, and by the Veterans Administration Medical Center in San Diego. This research concerns a system to automatically diagnose diseases of the human eye. The system takes as input information observable in a retinal image. This information is formulated to mimic the findings that an ophthalmologist would note during a clinical examination. The main output of the system is a diagnosis formulated to mimic the conclusion that an ophthalmologist would reach about the health of the subject. Appendix C
Fundus Photography Databases
143
The images in the STARE database [27] are designated imXXXX where XXXX is a four digit number (available images are numbered from 0001 to 0402 while five numbers in this range are missing). Of the 397 available images, 354 images were used for experiments since 43 images have missing or incomplete annotations or ground truth diagnoses (cases were the expert physician felt that the image was impossible to diagnose alone; a physician might use verbally acquired evidence from the patient, as well as multiple images of multiple portions of the patient's eye in order to conclude a diagnosis). The images (Figure 2.4(a)) were captured using a TopCon TRV50 fundus camera at 35° field of view, and subsequently digitized at 605×700 pixels in resolution, 24-bits pixel (standard RGB) [18, 52]. Hoover et al. [52] described an automated method to locate and outline blood vessels and they made their database of segmented retinal vessel images available (including two manual segmentations). They selected twenty retinal fundus images from the STARE for testing the described method. Ten of the images are of patients with no pathology while the other ten images contain pathology that obscures or confuses the blood vessel appearance in varying portions of the image. However, there are several drawbacks to this set of images that motivated [50] to develop the DRIVE database (see section C.1). First, the images in the Hoover database are not screening images. Next, because they are scanned from film the quality of the images is lower than if they would have been acquired directly using a digital fundus camera. Finally, the variability between the manual segmentation from the first and second observer is substantial. In [18], Hoover et al. described an automated method to locate the optic nerve and again made their database that includes the retinal images and the results available. They evaluated their method using 31 images of healthy retinas and 50 images of diseased retinas, containing such diverse symptoms 144
Appendix C
Fundus Photography Databases
as tortuous vessels, choroidal neovascularisation, and hemorrhages that completely obscure the actual nerve, both sets are from the STARE project images. The nerve is visible in all 81 images, although in 14 images it is on the border of the image so that only a fraction of the nerve is visible, and in 5 of the images the nerve is completely obscured by hemorrhaging. The 81 images used in [18] include 19 of those used in [52] (Table C.1). Image № im0001 im0002 im0003 im0004 im0005 im0006 im0007 im0008 im0009 im0010 im0011 im0012 im0013 im0014 im0015 im0016 im0017 im0018 im0019 im0020 im0021 im0022 im0023 im0024 im0025 im0026 im0027 im0028 im0029 im0030 im0031 im0032 im0033 im0034 im0035 im0036 im0037
Ground Truth Diagnoses Background Diabetic Retinopathy Choroidal Neovascularisation AND Arteriosclerotic Retinopathy Drusen, large AND Geographic Atrophy RPE" Cilio-Retinal Artery Occlusion OR Central Retinal Artery Occlusion Central Retinal Artery Occlusion AND Central Retinal Vein Occlusion Drusen Choroidal Neovascularisation AND Age Related Macular Degeneration AND Arteriosclerotic Retinopathy & Hypertensive Retinopathy Choroidal Neovascularisation AND Age Related Macular Degeneration AND Arteriosclerotic Retinopathy Background Diabetic Retinopathy Histoplasmosis Choroidal Neovascularisation AND Histoplasmosis Drusen Background Diabetic Retinopathy Nevus Drusen Background Diabetic Retinopathy Choroidal Neovascularisation Central Retinal Vein Occlusion, compensated Central Retinal Vein Occlusion Central Retinal Artery Occlusion AND Central Retinal Vein Occlusion Central Retinal Vein Occlusion Arteriosclerotic Retinopathy ???? Epiretinal Membrane?? Central Retinal Vein Occlusion AND Arteriosclerotic Retinopathy Central Retinal Vein Occlusion Central Retinal Vein Occlusion Hemi-Central Retinal Vein Occlusion AND Arteriosclerotic Retinopathy Unknown Diagnosis Normal? Background Diabetic Retinopathy Normal Central Retinal Vein Occlusion Choroidal Neovascularisation AND Age Related Macular Degeneration Normal Choroidal Neovascularisation AND Age Related Macular Degeneration Choroidal Neovascularisation AND Age Related Macular Degeneration AND Hypertensive Retinopathy
Appendix C
Fundus Photography Databases
Referenced by: [52], & [18] [52], & [18] [52], & [18] [52], & [18] [52], & [18] [18] [18] [18] [18] [18] [18] [18] [18] [18] [18] [18] [18] [18] [18] [18] [18] [18] [18] [18] [18] [18] [18] [18] [18] [18] [18] [18] [18] [18] [18] [18] [18]
145
im0038 im0039 im0040 im0041 im0042 im0043 im0044 im0045 im0046 im0048 im0049 im0050 im0076 im0077 im0080 im0081 im0082 im0139 im0162 im0163 im0164 im0170 im0190 im0198 im0216 im0219 im0231 im0235 im0236 im0237 im0238 im0239 im0240 im0241 im0242 im0243 im0245 im0249 im0252 im0253 im0255 im0278 im0291 im0319 im0324
Background Diabetic Retinopathy AND Hypertensive Retinopathy Drusen, fine Background Diabetic Retinopathy OR Human Immune Deficiency Virus Drusen, fine Retinitis, Disease AND Toxoplasmosis Retinitis, Disease Retinitis, Disease Choroidal Neovascularisation AND Age Related Macular Degeneration AND Arteriosclerotic Retinopathy Drusen Hypertensive Retinopathy Coats', Not Isolated Telangiectasia Background Diabetic Retinopathy Normal Normal AND Hypertensive Retinopathy [V: How could it be?] Normal ?, possibly fine Drusen Normal Normal Background Diabetic Retinopathy Normal Normal Normal Normal Normal Normal Normal Normal Normal Normal Normal Normal Normal Normal Normal Normal Normal Normal Normal Normal Normal Normal Normal Normal Vasculitis Emboli, Non-Hollenhorst Hollenhorst Plaque
[18] [18] [18] [18] [18] [18] [52], & [18] [18] [18] [18] [18] [18] [18] [52], & [18] [18] [52], & [18] [52], & [18] [52], & [18] [52], & [18] [52], & [18] [18] [18] [18] [18] [18] [18] [18] [52], & [18] [52], & [18] [18] [18] [52], & [18] [52], & [18] [18] [18] [18] [18] [18] [18] [18] [52], & [18] [18] [52], & [18] [52], & [18] [52]
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
Appendix C
Fundus Photography Databases
Appendix
D
Color Models Color models are invented and employed in order to ease the process of defining and using colors in a standardized way, therefore each color model specifies system where each color is uniquely represented by a single point [4]. This appendix presents in details the color models needed through this research, including their conversion equations.
D.1
HSI Color Model
Although RGB "Red, Green, and Blue" color system is ideally suited for hardware implementations, and it matches nicely with the fact that the human eye is strongly sensitive to red, green, and blue primaries, it is not well suited for describing colors in terms that are practical for human interpretation [4]. Humans view a colored object as a relation between 3 attributes, hue; which is a color attribute that describes pure colors (pure yellow, green .etc.), saturation; which gives a measure of the degree to which pure color is diluted with white light, and intensity (gray-level); a representation of brightness (although brightness is a subjective descriptor that is practically impossible to measure [4]). Hence, the HSI "Hue, Saturation, and Intensity" color model decouples the intensity component from the color carrying components (hue, saturation). Transforming the intensities of the three color bands to an HSI representation allows the intensity to be processed without affecting the perceived relative color values of the pixels [32]. The proofs of the Appendix D
Color Models
147
transformation equations given below are explained in [4], and a MATLAB implementation beside the explanation can be found in [31]. D.1.1 Converting Colors from RGB to HSI Given an image in RGB color format, the following equations describe the process of converting the R, G, and B values – assumed to be normalized to the range [ 0,1] – into H, S, and I values. The hue component is given by: ⎧ θ H =⎨ ⎩360 − θ
if B ≤ G if B > G
with ⎡
⎤ ⎥ ⎢⎣[( R − G ) 2 + ( R − B )(G − B )] ⎥⎦
θ = cos −1 ⎢
1 2
[( R − G ) + ( R − B )]
(Eq. D.1)
1 2
The saturation component is given by: ⎡ ⎤ 3 S = 1− ⎢ [min( R, G, B)]⎥ ⎣ ( R + G + B) ⎦
(Eq. D.2)
Finally, the intensity component is given by:
I=
1
3
( R + G + B)
(Eq. D.3)
D.1.2 Converting Colors from HSI to RGB
Given the values of the HSI image, the following equations convert them to the corresponding RGB values. There are three sets of equations; each is applied when the H value falls in its corresponding sector. •
RG sector (0° ≤ H < 120°) : if H satisfies the condition, the RGB
components are given by the following equations: B = I (1 − S ) 148
(Eq. D.4) Appendix D
Color Models
•
⎡ S cos H ⎤ R = I ⎢1 + ⎥ ⎣ cos(60° − H ) ⎦
(Eq. D.5)
G = 3I − ( R + B )
(Eq. D.6)
GB sector (120° ≤ H < 240°) : RGB components are given by the following equations:
•
H = H − 120°
(Eq. D.7)
R = I (1 − S )
(Eq. D.8)
⎡ S cos H ⎤ G = I ⎢1 + ⎥ ⎣ cos(60° − H ) ⎦
(Eq. D.9)
B = 3I − ( R + G )
(Eq. D.10)
BR sector (240° ≤ H ≤ 360°) : RGB components are given by the following equations:
D.2
H = H − 240°
(Eq. D.11)
G = I (1 − S )
(Eq. D.12)
⎡ S cos H ⎤ B = I ⎢1 + ⎥ ⎣ cos(60° − H ) ⎦
(Eq. D.13)
R = 3I − (G + B)
(Eq. D.14)
Chromaticity
Chromaticity is a representation of digital images which is invariant to light geometry (i.e. light source direction and power) [22, 36]. Therefore, chromaticity values (Figure D.1) represent an image while discarding the intensity (brightness) information, and so can be thought of as hue and Appendix D
Color Models
149
saturation taken together [4]. The chromaticity normalized values ( r , g , b) for any pixel ( R, G, B) can be derived from the following equation: r=
R G B , g= , and b = R+G+ B R+G+ B R+G+ B
(Eq. D.15)
further, only two chromaticity values are needed to represent a color, since
r + g +b =1
(Eq. D.16)
If the light geometry affecting any pixel ( R, G, B) was scaled by the factor s, then the captured image is also scaled by the same factor, altering the pixel values to ( sR, s G, sB). Thus, representing the image by its
chromaticity values using (Eq. D.15) will eliminate any dependency on light geometry, producing a color normalized image:
r=
sG sB sR , g= , and b = sR + sG + sB sR + sG + sB sR + sG + sB (Eq. D.17)
Figure D.1
Chromaticity Diagram, by the CIE (Commission Internationale de l'Eclairage – the international Commission on Illumination). [112]
150
Appendix D
Color Models
Appendix
E
A Threshold Selection Method from Gray‐Level Histograms It's important in image processing to select an adequate threshold of gray level for extracting objects from their background. In ideal case, the histogram has a deep and sharp valley between the peaks representing objects and background, respectively, so the threshold can be chosen at the bottom of this valley [63]. However, for more real pictures, it is often difficult to detect the valley bottom precisely, especially in cases when the valley is flat and broad, filled with noise, or when the two peaks are extremely unequal in height, often producing no traceable valleys. Nobuyuki Otsu [63] proposed the following algorithm that computes a global image threshold using only the gray-level histogram without other priori knowledge. Let L be the number of gray levels in a typical image [1,2,L, L − 1] . The number of pixels in level i is denoted by ni , where 0 ≤ i < L , and N is the total number of pixels in the whole image; N = n1 + n 2 + L + n L −1 . The graylevel histogram is regarded as a normalized probability distribution pi, where:
p i = ni / N , Appendix E
pi ≥ 0,
L −1
∑p i =0
i
=1
(Eq. E.1)
A Threshold Selection Method from Gray-Level Histograms 151
Then we Dichotomize the pixels of the given image into two classes C0 and C1 (representing the background and objects, or vice-versa) by threshold k; C0 denotes the pixels with levels [0, L, k ] , while C1 denotes the pixels with levels [k + 1,L, L − 1] . The probability of class occurrence 'ω' – OMEGA – is given by: k
ω 0 = Pr(C 0 ) = ∑ pi
(Eq. E.2)
i =0
ω1 = Pr(C1 ) =
L −1
∑p
i = k +1
(Eq. E.3)
i
while the class mean level 'µ' – MU – is given by: k
k
i =0
i =0
µ 0 = ∑ i Pr(i | C 0 ) = ∑ ipi / ω 0 µ1 =
L −1
∑ i Pr(i | C1 ) =
i = k +1
L −1
∑ ip
i = k +1
i
/ ω1
(Eq. E.4)
(Eq. E.5)
and the total mean level of the given picture is: L −1
µ T = ∑ ipi
(Eq. E.6)
i =0
In order to evaluate the "goodness" of the threshold at level k, the author used the total variance of levels 'η' – ETA – as a discriminant criterion measure (i.e. measure of class separability).
η = σ B2 / σ T2
(Eq. E.7)
where
σ B2 = ω 0 ( µ 0 − µ T ) 2 + ω1 ( µ1 − µ T ) 2 = ω 0ω1 ( µ1 − µ 0 ) 2
(Eq. E.8)
and 152 Appendix E
A Threshold Selection Method from Gray-Level Histograms
L −1
σ T2 = ∑ (i − µ T ) 2 pi
(Eq. E.9)
i =0
It's noticed that σ T2 is independent of k, while σ B2 is a function of the threshold k based on first-order statistics (class means). The optimal threshold k* that maximizes the η, or equivalently maximizes σ B2 , is selected by the following sequential search:
σ B2 (k *) = max σ B2 (k ) 0≤ k < L −1
(Eq. E.10)
The lower bound {η = 0} is attainable only by images having a single gray level, and the upper bound {η = 1} is attainable only by two-valued images. The effective range of the gray-level histogram S* (the range of k over which the maximum is sought) can be restricted to: S * = {k ; ω 0ω1 > 0} .
Appendix E
A Threshold Selection Method from Gray-Level Histograms 153
References [1] Alireza Osareh, “Automated Identification of Diabetic Retinal Exudates and the Optic Disc,” Ph.D. Thesis, Department of Computer Science, Faculty of Engineering, University of Bristol, January, 2004. [2] G. F. Luger, Artificial Intelligence Structures and Strategies or Complex Problem Solving, 4th Edition, Pearson Education, 2002. [3] Rezaul Begg and Marimuthu Palaniswami, Computational Intelligence for Movement Sciences: Neural Networks and Other Emerging Techniques, IDEA Group Publishing, February 2006. [4] R. C. Gonzalez and R. E. Woods, Digital Image Processing, 2nd Edition, Prentice-Hall, 2002. [5] C. Sinthanayothin, “Image analysis for automatic diagnosis of diabetic retinopathy,” Ph.D. Thesis, University of London (King's College London), September 1999. [6] A-H. El-Gilany, S. El-Fedawy, and M. Tharwat, “Causes of blindness and needs of the blind in Mansoura, Egypt,” Eastern Mediterranean Health Journal, vol. 8, no. 1, Jan. 2002. [7] J. Staal, M. D. Abràmoff, M. Niemeijer, M. A. Viergever, and B. van-Ginneken, “Ridgebased vessel segmentation in color images of the retina,” IEEE Transactions on Medical Imaging, vol. 23, no. 4, pp. 501–509, April 2004. [8] W. H. Herman, R. E. Aubert, M. A. Ali, E. S. Sous, and A. Badran, “Diabetes mellitus in Egypt: risk factors, prevalence and future burden,” Eastern Mediterranean Health Journal, vol. 3, no. 1, pp. 144–148, 1997. [9] M. El-Shazly, M. Zeid, and A. Osman, “Risk factors for eye complications in patients with diabetes mellitus: development and progression,” Eastern Mediterranean Health Journal, vol. 6, no. 2/3, pp. 313–325, 2000, [10] Alireza Osareh, Majid Mirmehdi, Barry Thomas, and Richard Markham, “Classification and Localisation of Diabetic-Related Eye Disease,” Proceedings of the 7th
154
References
European Conference on Computer Vision (ECCV), pp. 502–516, Springer-Verlag, LNCS 2353, May 2002. [11] K. Facey, E. Cummins, K. Macpherson, A. Morris, L. Reay, and J. Slattery, “Organisation of services for diabetic retinopathy screening,” Health Technology Assessment Report 1, Glasgow: Health Technology Board for Scotland, April 2002. [12] H. Li and O. Chutatape, “Fundus image features extraction,” Proceedings of the 22nd Annual IEEE Engineering in Medicine and Biology Society (EMBS) International Conference, July 23-28, 2000, Chicago IL., pp. 3071–3073, 2000. [13] S. C. Siu, T. C. Ko, K. W. Wong, and W. N. Chan, “Effectiveness of non-mydriatic retinal photography and direct ophthalmoscopy in detecting diabetic retinopathy,” Hong Kong Medical Journal, vol. 4, no. 4, pp. 367–370, December 1998. [14] C. Sinthanayothin, J. F. Boyce, T. H. Williamson, H. L. Cook, E. Mensah, S. Lal, and D. Usher, “Automated detection of diabetic retinopathy on digital fundus images,” Diabetes UK – Diabetic Medicine, vol. 19, pp. 105–112, 2002. [15] R. A. Abdel-Ghafar, T. Morris, T. Ritchings, and I. Wood, “Detection and characterisation of the optic disk in glaucoma and diabetic retinopathy,” Medical Image Understanding and Analysis Conference (MIUA), Imperial College London, September 2324, 2004. [16] T. Morris and Z. Newell, “Segmentation of retinal images guided by the wavelet transform,” Proceedings of the International Workshop on Image and Signal Processing '96, Elsevier, Amsterdam, 1996. [17] T. Teng, M. Lefley, and D. Claremont, “Progress towards automated diabetic ocular screening: a review of image analysis and intelligent systems for diabetic retinopathy,” Medical & Biological Engineering & Computing, vol. 40, pp. 2–13, 2002. [18] A. Hoover and M. Goldbaum, “Locating the optic nerve in a retinal image using the fuzzy convergence of the blood vessels,” IEEE Transactions on Medical Imaging, vol. 22, no. 8, pp. 951–958, 2003. [19] Frank ter Haar, “Automatic localization of the optic disc in digital colour images of the human retina,” M.S. Thesis, Utrecht University, the Netherlands, December 16, 2005. [20] Bernd Liesenfeld, Eva Kohner, Wolfgang Piehlmeter, Stefan Kluthe, Steve Aldington, Massimo Porta, Toke Bek, Marco Obermaier, Hans Mayer, Gerd Mann, Rolf Holle, and
References
155
Karl-Dieter Hepp, “A telemedical approach to the screening of diabetic retinopathy: Digital fundus photography”, Diabetes Care, vol. 23, no. 3, pp.345–348, March 2000. [21] Peter Koch Jensen and Erik Scherfig, “Resolution of retinal digital colour images,” Acta Ophthalmologica Scandinavica, vol. 77, no. 5, pp. 526–529, October 1999. [22] Keith A. Goatman, A. David Whitwam, A. Manivannan, John A. Olson, and Peter F. Sharp, “Colour normalisation of retinal images,” Proceedings of Medical Image Understanding and Analysis, 2003. [23] L. Gagnon, M. Lalonde, M. Beaulieu, and M.-C. Boucher, “Procedure to detect anatomical structures in optical fundus images,” Proceedings of Conference Medical Imaging: Image Processing (SPIE #4322), San Diego, pp.1218–1225, February 19-22, 2001. [24] University Medical Center Utrecht, Image Sciences Institute, Research section, Digital Retinal Image for Vessel Extraction (DRIVE) database, [Online]. Available: http://www.isi.uu.nl/Research/Databases/DRIVE [25] G. Yang, L. Gagnon, S. Wang, and M.-C. Boucher, “Algorithm for detecting microaneurysms in low-resolution color retinal images,” Proceedings of Vision Interface 2001, Ottawa, pp. 265–271, June 7-9, 2001. [26] D. J. Cornforth, H. J. Jelinek, J. J. G. Leandro, J. V. B. Soares, R. M. Cesar, Jr., M. J. Cree, P. Mitchell, and T. Bossomaier, “Development of retinal blood vessel segmentation methodology using wavelet transforms for assessment of diabetic retinopathy,” Proceedings of the 8th Asia Pacific Symposium on Intelligent and Evolutionary Systems, Cairns Australia, December 6-7, 2004. [27] STructured Analysis of the REtina (STARE) project website. Clemson Univ., Clemson, SC. [Online]. Available: http://www.ces.clemson.edu/~ahoover/stare [28] F. Zana and J. C. Klein, “A Multimodal Registration Algorithm of Eye Fundus Images Using Vessels Detection and Hough Transform”, IEEE Transactions on Medical Imaging, vol. 18, no. 5, pp. 419–428, May 1999. [29] Luo Gang, Opas Chutatape, and Shankar M. Krishnan, “Detection and Measurement of Retinal Vessels in Fundus Images Using Amplitude Modified Second-Order Gaussian Filter”, IEEE Transactions on Biomedical Engineering, vol. 49, no. 2, pp. 168–172, February 2002.
156
References
[30] Subhasis Chaudhuri, Shankar Chatterjee, Norman Katz, Mark Nelson, and Michael Goldbaum, “Detection of blood vessels in retinal images using two-dimensional matched filters”, IEEE Transactions on Medical Imaging, vol. 8, no. 3, pp. 263–269, Sept. 1989. [31] R. C. Gonzalez, R. E. Woods, and S. L. Eddins, Digital Image Processing Using MATLAB, Pearson Education, 2004. [32] C. Sinthanayothin, J. F. Boyce, H. L. Cook and T. H. Williamson, “Automated localisation of the optic disk, fovea, and retinal blood vessels from digital colour fundus images,” British Journal of Ophthalmology, vol. 83, no. 8, pp. 902–910, 1999. [33] Di Wu, Ming Zhang, Jyh-Charn Liu, and Wendall Bauman, “On the Adaptive Detection of Blood Vessels in Retinal Images,” IEEE Transactions on Biomedical Engineering, vol. 53, no. 2, pp. 341–343, February 2006. [34] Tusheng Lin and Yibin Zheng, “Adaptive image enhancement for retinal blood vessel segmentation,” Electronics Letters, vol. 38, no. 19, pp. 1090–1091, September 12, 2002. [35] M. Foracchia, E. Grisan, and A. Ruggeri, “Luminosity and contrast normalization in retinal images,” Medical Image Analysis, vol. 3, no. 9, pp. 179–190, 2005. [36] Graham D. Finlayson, Bernt Schiele, and James L. Crowley, “Comprehensive colour image normalization,” Proceedings of the European Conference on Computer Vision, Springer-Verlag, pp.475–490, 1998. [37] Marc Lalonde, Langis Gagnon, and Marie-Carole Boucher, “Automatic visual quality assessment in optical fundus images,” Proceedings of Vision Interface 2001, Ottawa, pp. 259–264, June 7-9, 2001. [38] Samuel C. Lee and Yiming Wang, “Automatic retinal image quality assessment and enhancement,” Proceedings of SPIE Conference on Image Processing, vol. 3661, pp. 1581– 1590, February 1999. [39] A. Toniappa, S. A. Barman, E. Corvee, M. J. Moseley, K. Cocker, and A. R. Fielder, “Image quality assessment in retinal images of premature infants taken with RetCam 120 digital fundus camera,” The Imaging Science Journal, vol. 53, no. 1, pp. 51–59, March 2005. [40] Rayleigh distribution: From Wikipedia http://en.wikipedia.org/wiki/Rayleigh_distribution
References
–
the
free
encyclopedia,
157
[41] Alan D. Fleming, Sam Philip, Keith A. Goatman, John A. Olson, and Peter F. Sharp, “Automated Assessment of Diabetic Retinal Image Quality Based on Clarity and Field Definition,” Investigative Ophthalmology and Visual Science, vol. 47, no. 3, pp. 1120–1125, March 2006. [42] Huiqi Li and Opas Chutatape, “Automatic Location of Optic Disc in Retinal Images,” IEEE International Conference on Image Processing, vol. 2, pp. 837–840, Oct. 7-10, 2001. [43] James Lowell, Andrew Hunter, David Steel, Ansu Basu, Robert Ryder, Eric Fletcher, and L. Kennedy, “Optic Nerve Head Segmentation,” IEEE Transactions on Medical Imaging, vol. 23, no. 2, pp. 256–264, February 2004. [44] Huiqi Li and Opas Chutatape, “A Model-Based Approach for Automated Feature Extraction in Fundus Images,” Ninth IEEE International Conference on Computer Vision (ICCV'03), vol. 1, pp. 394–399, 2003. [45] M. Lalonde, M. Beaulieu and L. Gagnon, “Fast and robust optic disk detection using pyramidal decomposition and Hausdorff-based template matching,” IEEE Transactions on Medical Imaging, vol. 20, no. 11, pp. 1193–1200, November 2001. [46] Thomas Walter and Jean-Claude Klein, “Segmentation of color fundus images of the human retina: Detection of the optic disc and the vascular tree using morphological techniques,” Proceedings of the 2nd International Symposium on Medical Data Analysis, pp. 282–287, 2001. [47] Radim Chrástek, Matthias Wolf, Klaus Donath, Georg Michelson, Heinrich Niemann, “Optic Disc Segmentation in Retinal Images,” Bildverarbeitung für die Medizin 2002, pp. 263–266, 2002. [48] Alireza Osareh, Majid Mirmehdi, Barry Thomas, and Richard Markham, “Comparison of colour spaces for optic disc localisation in retinal images,” Proceedings of the 16th International Conference on Pattern Recognition, pp. 743–746, 2002. [49] S. F. Barrett, E. Naess, and T. Molvik, “Employing the Hough transform to locate the optic disk,” Biomedical Sciences Instrumentation, vol. 37, pp. 81–86, February 2001. [50] M. Niemeijer, J. Staal, B. van Ginneken, M. Loog, and M. D. Abràmoff, “Comparative study of retinal vessel segmentation methods on a new publicly available database,” SPIE Medical Imaging, vol. 5370, pp. 648–656, 2004.
158
References
[51] Adam Hoover and Michael Goldbaum, “Fuzzy convergence,” Proceedings of IEEE Computer Society Conference on Computer Vision and Pattern Recognition, Santa Barbara, California, USA, pp. 716–721, 1998. [52] A. Hoover, V. Kouznetsova, and M. Goldbaum, “Locating blood vessels in retinal images by piecewise threshold probing of a matched filter response,” IEEE Transactions on Medical Imaging, vol. 19, no. 3, pp. 203–210, March 2000. [53] M. Foracchia, E. Grisan, and A. Ruggeri, “Detection of Optic Disc in Retinal Images by Means of a Geometrical Model of Vessel Structure,” IEEE Transactions on Medical Imaging, vol. 23, no. 10, pp. 1189–1195, October 2004. [54] Alan D. Fleming, Keith A. Goatman, Sam Philip, John A. Olson, Peter F. Sharp, “Automatic detection of retinal anatomy to assist diabetic retinopathy screening,” Physics in Medicine and Biology, vol. 52, no. 2, pp. 331–345, January 21, 2007. [55] M. Goldbaum, S. Moezzi, A. Taylor, S. Chatterjee, J. Boyd, E. Hunter, R. Jain, “Automated diagnosis and image understanding with object extraction, object classification, and inferencing in retinal images,” Proceedings of IEEE International Congress of Image Processing, Los Alamitos, California, USA, vol. 3, pp. 695–698, 1996. [56] Kenneth W. Tobin, Edward Chaum, V. Priya Govindasamy, Thomas P. Karnowski, and Omer Sezer, “Characterization of the optic disc in retinal imagery using a probabilistic approach,” Medical Imaging 2006: Image Processing. Proceedings of the SPIE, vol. 6144, no. 4, pp. 1088–1097, March 2006. [57] M. D. Abràmoff, and M. Niemeijer, “The automatic detection of the optic disc location in retinal images using optic disc location regression,” Proceedings of the IEEE EMBC 2006, (in press). [58] Nahed H. Solouma, Abou-Bakr M. Youssef, Yehia A. Badr, and Yasser M. Kadah, “A New Real-Time Retinal Tracking System for Image-Guided Laser Treatment,” IEEE Transactions on Biomedical Engineering, vol. 49, no. 9, pp. 1059–1067, September 2002. [59] Jin Yu, Syed Sibte Raza Abidi, Paul Habib Artes, Andrew McIntyre, and Malcolm Heywood, “Automated Optic Disk Analysis for Diagnostic Support to Classify Glaucoma Patients,” 18th IEEE International Symposium on Computer-Based Medical Systems, Dublin, June 23-24 2005.
References
159
[60] Axel Pinz, Stefan Bernögger, Peter Datlinger, and Andreas Kruger, “Mapping the Human Retina,” IEEE Transactions on Medical Imaging, vol. 17, no. 4, pp. 606–619, August 1998. [61] N. Patton, T. Aslam, T. MacGillivray, A. Pattie, I. Deary, and B. Dhillon, “Retinal vascular image analysis as a potential screening tool for cerebrovascular disease: a rationale based on homology between cerebral and retinal microvasculatures,” J. Anat., vol. 206, no. 4, pp. 319–348, 2005. [62] Xiaohong Gao, Anil Bharath, Alice Stanton, Alun Hughes, Neil Chapman, and Simon Thom, “Measurement of Vessel Diameters on Retinal Images for Cardiovascular Studies,” Proceedings of the Medical Image Understanding and Analysis Conference (MIUA'01), the University of Birmingham, July 16-17, 2001. [63] Nobuyuki Otsu, “A Threshold Selection Method from Gray Level Histograms,” IEEE Transactions on Systems, Man, and Cybernetics, vol. SMC-9, no. 1, pp. 62–66, Jan. 1979. [64] O. Chutatape, Liu Zheng, and S. M. Krishnan, “Retinal Blood Vessel Detection and Tracking by Matched Gaussian and Kalman Filters,” Proceedings of the 20th Annual International Conference of the IEEE Engineering in Medicine and Biology Society, vol. 20, no. 6, pp. 3144–3149, 1998. [65] Frédéric Zana and Jean-Claude Klein, “Segmentation of Vessel-Like Patterns Using Mathematical Morphology and Curvature Evaluation,” IEEE Transactions on Image Processing, vol. 10, no. 7, pp. 1010–1019, July 2001. [66] João V. B. Soares, Jorge J. G. Leandro, Roberto M. Cesar Jr., Herbert F. Jelinek, and Michael J. Cree, “Retinal Vessel Segmentation Using the 2-D Gabor Wavelet and Supervised Classification,” IEEE Transactions on Medical Imaging, vol. 25, no. 9, pp. 1214–1222, September 2006. [67] Human-machine integration for vessel segmentation project [Online]. Available: http://retina.iv.fapesp.br [68] A. P. Condurache, and T. Aach, “Vessel Segmentation in 2D-Projection Images Using a Supervised Linear Hysteresis Classifier,” Proceedings of the 18th International Conference on Pattern Recognition (ICPR'06), Hong Kong - China, pp. 343–346, August 20-24, 2006.
160
References
[69] Vessel segmentation and analysis (VegSeg) software tool [Online]. Available: http://www.isip.uniluebeck.de/index.php?id=57 [70] Ramesh Jain, Rangachar Kasturi, and Brian G. Schunck, Machine Vision, Computer Science Series, McGraw-Hill, 1995. [71] Danny Y. Lin, Mark S. Blumenkranz, Rosemary J. Brothers, and David M. Grosvenor, “The sensitivity and specificity of single-field nonmydriatic monochromatic digital fundus photography with remote image interpretation for diabetic retinopathy screening: A comparison with ophthalmoscopy and standardized mydriatic color photography,” American Journal of Ophthalmology, vol. 134, Issue 2, pp. 204–213, August 2002. [72] Marilyn James, David A. Turner, Deborah M. Broadbent, Jiten Vora, and Simon P. Harding, “Cost effectiveness analysis of screening for sight threatening diabetic eye disease,” BMJ, vol. 320, pp. 1627–1631, June 17, 2000. [73] P. Kahai, K. R. Namuduri, and H. Thompson, “A Decision Support Framework for Automated Screening of Diabetic Retinopathy,” International Journal of Biomedical Imaging, vol. 2006, Article ID 45806, pp. 1–8, October 8, 2005. [74] B. Raman, E.S. Bursell, M. Wilson, G. Zamora, I. Benche, S.C.Nemeth, and P. Soliz, “The Effects of Spatial Resolution on an Automated Diabetic Retinopathy Screening System’s Performance in Detecting Microaneurysms for Diabetic Retinopathy,” Proceedings of the 17th IEEE Symposium on Computer-Based Medical Systems (CBMS’04), 2004. [75] Keith A. Goatman, Michael J. Cree, John A. Olson, John V. Forrester, and Peter F. Sharp, “Automated Measurement of Microaneurysm Turnover,” Investigative Ophthalmology & Visual Science, vol. 44, no. 12, pp. 5335–5341, December 2003. [76] M. J. Cree, E. Gamble, and D. Cornforth, “Colour Normalisation to Reduce InterPatient and Intra-Patient Variability in Microaneurysm Detection in Colour Retinal Images,” Proceedings of the Australian Pattern Recognition Society (APRS) Workshop on Digital Image Computing (WDIC'05), Griffith University, Southbank, Brisbane, Australia, February 21, 2005. [77] Huiqi Li and Opas Chutatape, “Automated Feature Extraction in Color Retinal Images by a Model Based Approach,” IEEE Transactions on Biomedical Engineering, pp. 1–9, 2004.
References
161
[78] Xiaohui Zhang and Opas Chutatape, “Detection and Classification of Bright Lesions in Color Fundus Images,” Proceedings of the International Conference on Image Processing (ICIP), pp. 139–142, 2004. [79] Xiaohui Zhang and Opas Chutatape, “Top-down and Bottom-up Strategies in Lesion Detection of Background Diabetic Retinopathy,” Proceedings of the 2005 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR’05), 2005. [80] Gang Luo, Opas Chutatape, Huiqi Li, and Shankar. M. Krishnan, “Abnormality Detection in Automated Mass Screening System of Diabetic Retinopathy,” Proceedings of the Fourteenth IEEE Symposium on Computer-Based Medical Systems (CBMS’01), 2001. [81] Thomas Walter, Jean-Claude Klein, Pascale Massin, and Ali Erginay, “A Contribution of Image Processing to the Diagnosis of Diabetic Retinopathy - Detection of Exudates in Color Fundus Images of the Human Retina,” IEEE Transactions on Medical Imaging, vol. 21, no. 10, pp. 1236–1243, October 2002. [82] Thomas Walter and Jean-Claude Klein, “A Computational Approach to Diagnosis of Diabetic Retinopathy,” Proceedings of the 6th Conference on Systemics, Cybernetics and Informatics (SCI), Orlando – Florida, USA, pp. 521–526, July 15-18, 2002. [83] Huan Wang, Wynne Hsu, Kheng Guan Goh, and Mong Li Lee, “An Effective Approach to Detect Lesions in Color Retinal Images,” Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Hilton Head Island, SC - USA, vol. 2, pp. 181–186, June 13-15, 2000. [84] Meindert Niemeijer, “Automatic Detection of Diabetic Retinopathy in Digital Fundus Photographs,” Ph.D. Thesis, Image Sciences Institute, University Medical Center Utrecht, Utrecht, the Netherlands, December 20, 2006. [85] Stephen J. Preece and Ela Claridge, “Physics Based Medical Image Understanding of the Colouration of the Ocular Fundus with Application to Detection of Diabetic Retinopathy,” Proceedings of Medical Image Understanding and Analysis (MIUA), University College London, United Kingdom, 2000. [86] Seong Ho Park, Jin Mo Goo, and Chan-Hee Jo, “Receiver Operating Characteristic (ROC) Curve: Practical Review for Radiologists,” Korean J. Radiol., vol. 5, pp. 11–18, March 2004. [87] MedCalc Software, Mariakerke, Belgium, [Online]. Available: http://www.medcalc.be
162
References
[88] Alejandro F. Frangi, Wiro J. Niessen, Koen L. Vincken, and Max A. Viergever, “Multiscale vessel enhancement filtering,” Lecture Notes in Computer Science, SpringerVerlag, vol. 1496, pp.130–137, 1998. [89] C. Lorenz, I.-C. Carlsen, T. M. Buzug, C. Fassnacht, and J. Weese, “Multi-scale line segmentation with automatic estimation of width, contrast, and tangential direction in 2D and 3D medical images,” Lecture Notes in Computer Science, vol. 1205, pp. 233–242, 1997. [90] R. O. Duda, P. E. Hart, and D. G. Stork, Pattern Classification, 2nd Edition, Wiley, 2000. [91] Issam El-Naqa, Yongyi Yang, Miles N. Wernick, Nikolas P. Galatsanos, and Robert M. Nishikawa, “A Support Vector Machine Approach for Detection of Microcalcifications,” IEEE Transactions on Medical Imaging, vol. 21, no. 12, pp. 1552–1563, December 2002. [92] X. Zhang, X. L. Xiao, J. W. Tian, J. Liu, and G. Y. Xu, “Application of support vector machines in classification of magnetic resonance images,” International Journal of Computers and Applications, vol. 28, no. 2, pp. 122–128, April 2006. [93] N. Piliouras, I. Kalatzis, N. Dimitropoulos, and D. Cavouras, “Development of the cubic least squares mapping linear-kernel support vector machine classifier for improving the characterization of breast lesions on ultrasound,” Computerized Medical Imaging and Graphics, vol. 28, no. 5, pp. 247–255, July 2004. [94] T. Joachims, “Making Large-Scale Support Vector Machine Learning Practical,” Advances in Kernel Methods: Support Vector Learning, MIT Press, pp. 169–184, 1999. [95] S. Canu, Y. Grandvalet, V. Guigue, and A. Rakotomamonjy, “SVM and Kernel Methods Matlab Toolbox,” Perception Systèmes et Information, INSA de Rouen, Rouen France, 2005 [Online]. Available: http://asi.insa-rouen.fr/~arakotom/toolbox/index.html [96] Advanced Retinal Imaging Center: A Historical Perspective - New York Eye & Ear Infirmary, Sept. 9, 2005 [Online]. Available: http://www.nyee.edu/aric_03.html?print=1 [97] Eye Institute at Hartford Hospital website, September 8, 2005 [Online]. Available: http://www.harthosp.org/eyes/images/diabetic_03.jpg [98] Digital reference of ophthalmology - Edward S. Harkness Eye Institute, Columbia University, September 9, 2005 [Online]. Available: http://dro.hs.columbia.edu/pucker1.htm [99] Wendy Strouse Watt, September http://www.mdsupport.org/library/angio.html
References
8,
2005
[Online].
Available:
163
[100] Nisha Sachdev, Gerard Cairns, and Charles McGhee, “An introduction to wavefront aberrometry of the eye,” New Zealand Optics Magazine, Eye on Ophthalmology, May 2002. [101] T. Nirmaier, G. Pudasaini, and J. Bille, “CMOS-based Hartmann-Shack sensor for real time adaptive optics,” 13th IEEE Real Time Conference, Montréal, Canada, May 18-23, 2003. [102] B. C. Platt and R. Shack, “History and principles of Shack-Hartmann wavefront sensing,” Journal of Refractive Surgery, vol. 17, pp. S573–577, September/October 2001. [103] Antonio Valerio Netto and Maria Cristina F. De Oliveira, “Image processing and analysis to diagnose ocular refraction errors,” Proceedings of the XV Brazilian Symposium on Computer Graphics and Image Processing (SIBGRAPI’02), IEEE 2002. [104] Antonio Valerio Netto and Maria Cristina F. De Oliveira, “A Proposal for Applying Artificial Intelligence to Measure Ocular Refractive Errors,” XXVIII Conferência LatinoAmericana de Informática CLEI´2002, Montevidéo – Uruguay, vol. 1, pp. 1–9, 25 al 29 de Noviembre de 2002. [105] CASIA Iris Image Database, National Laboratory of Pattern Recognition (NLPR), Institute of Automation (IA), Chinese Academy of Sciences (CAS) website, [Online]. Available: http://www.sinobiometrics.com/casiairis.htm [106] J. Daugman, “How iris recognition works,” IEEE Transactions on Circuits and Systems for Video Technology, vol. 14, no. 1, January 2004. [107] S. N. Yanushkevich, A. Stoica, S. N. Srihari, V. P. Shmerko, and M. L. Gavrilova, “Simulation of biometric information: The new generation of biometric systems,” Proceedings of the International Workshop on Modeling and Simulation in Biometric Technology (BT2004), Calgary, AB, Canada, pp.87–98, 2004. [108] Webpage for J. Daugman, “How iris recognition is used at airports,” University of Cambridge – Computer Laboratory, September 10, 2005, [Online]. Available: http://www.cl.cam.ac.uk/users/jgd1000/ [109] M. Arab, “Diabetes mellitus in Egypt,” World Health Statistics Quarterly (Rapport Trimestriel de Statistiques Sanitaires Mondiales), 45(4), pp. 334–337, 1992. [110] N. M. Kamel, Y. A. Badawy, N. A. El-Zeiny, and I. A. Merdan, “Sociodemographic determinants of management behaviour of diabetic patients Part I. Behaviour of patients in
164
References
relation to management of their disease,” Eastern Mediterranean Health Journal, vol. 5, Issue 5, pp. 967–973, 1999. [111] N. M. Kamel, Y. A. Badawy, N. A. El-Zeiny, and I. A. Merdan, “Sociodemographic determinants of management behaviour of diabetic patients Part II. Diabetics' knowledge of the disease and their management,” Eastern Mediterranean Health Journal, vol. 5, Issue 5, pp. 974–983, 1999. [112] Australian Government – Department of Defense website, February 27, 2006 [Online].Available:http://www.defence.gov.au/dpe/dhs/infocentre/publications/journals/NoI Ds/adfhealth_oct05/ADFHealth_6_2_62-66-5.gif
References
165
ﳎﻤﻮﻋﺎﺕ ﺻﻐﲑﺓ ﻭﻛﺬﻟﻚ ﺯﻣﻦ ﺍﻟﺘﺪﺭﻳﺐ ﺍﻟﻔﻌﺎﻝ ﳘﺎ ﻋﻨﺼﺮﻳﻦ ﺣﺎﲰﲔ ﺑﺎﻟﻨﺴﺒﺔ ﻟﻸﻧﻈﻤﺔ ﺍﻟﱵ ﲢﺘﺎﺝ ﺩﺍﺋﻤﺎ ﺍﱃ
ﺍﻟﺘﻌﺪﻳﻞ ﻭﺍﻟﻀﺒﻂ ﻣﻊ ﺍﻟﺒﻴﺎﻧﺎﺕ ﺍﳌﺨﺘﻠﻔﺔ .ﻗﺎﻋﺪﺓ ﺑﻴﺎﻧﺎﺕ ﺍﻟـ DRIVEﺍﳌﻌﻴﺎﺭﻳﺔ ﻭﺍﳌﺘﺎﺣﺔ ﻋﻠﻨﺎ ﺍﺳﺘﺨﺪﻣﺖ
ﻟﺘﻘﻴﻴﻢ ﺍﺩﺍﺀ ﺍﻟﻄﺮﻳﻘﺔ ﺍﳌﺴﺘﺤﺪﺛﺔ .ﺍﻟﺘﺠﺎﺭﺏ ﺗﻜﺸﻒ ﺍﻥ ﺍﳌﻨﻄﻘﺔ ﲢﺖ ﻣﻨﺤﲎ ﺍﳋﻮﺍﺹ ﺍﻟﻌﺎﻣﻠﺔ ﻟﻠﻤﺴﺘﻘﺒﻞ
) (AUCﺑﻠﻐﺖ ﻣﺎ ﻗﻴﻤﺘﻪ 0.9537ﻭﳝﻜﻦ ﻣﻘﺎﺭﻧﺘﻬﺎ ﺍﱃ ﺣﺪ ﻛﺒﲑ ﻣﻊ ﺍﻟـ) (AUCsﺍﳌﺴﺠﻠﺔ ﺳﺎﺑﻘﺎ ﻭﺍﻟﱴ ﺗﺘﺮﺍﻭﺡ ﻣﻦ 0.7878ﺍﱃ .0.9614
ﻭﺃﺧﲑﺍ ،ﺍﳉﺰﺀ ﺍﻟﺮﺍﺑﻊ ﻣﻦ ﺍﻻﻋﻤﺎﻝ ﺍﳌﻌﺮﻭﺿﺔ ﻳﺘﻨﺎﻭﻝ ﺍﻟﻜﺸﻒ ﺍﻵﱄ ﻟﻠﺘﺤﻠﺒﺎﺕ ﺍﻟﺼﻌﺒﺔ ،ﺑﺎﻋﺘﺒﺎﺭﻫﺎ
ﺍﻫﻢ ﻣﻈﺎﻫﺮ ﺍﻟـ) .(DRﺃﺳﺎﻟﻴﺐ ﺍﻟﺘﻌﺎﻣﻞ ﻣﻊ ﺍﻛﺘﺸﺎﻑ ﺗﻘﺮﺣﺎﺕ ﺍﻟﺸﺒﻜﻴﺔ ﺍﻟﺴﺎﻃﻌﺔ ﺑﺸﻜﻞ ﻋﺎﻡ ﻗﺪ ﰎ ﺍﺳﺘﻌﺮﺍﺿﻬﺎ .ﰒ ﰎ ﺍﻗﺘﺮﺍﺡ ﺍﺳﻠﻮﺏ ﻣﺒﲎ ﻋﻠﻰ ﺃﺳﺎﺱ ﺍﺳﺘﺠﺎﺑﺔ ﻣﺎﻛﻴﻨﺔ ﺍﳌﺘﺠﻪ ﺍﻟﺪﺍﻋﻢ ﻭﺍﺳﻌﺔ ﺍﻟﻨﻄﺎﻕ ﺍﳌﺪﺭﺑﺔ
ﻻﺳﺘﻘﻄﺎﻉ ﺍﻟـ) (RVﺁﻟﻴﺎ .ﺍﻟﻄﺮﻳﻘﺔ ﺍﳌﻘﺘﺮﺣﺔ ﺗﺴﺘﺨﺪﻡ ﻧﺴﺨﺔ ﻣﻌﻜﻮﺳﺔ ﻭﻣﻐﻠﻘﺔ ﻣﻦ ﺍﺳﺘﺠﺎﺑﺔ ﻣﺎﻛﻴﻨﺔ ﺍﳌﺘﺠﻪ
ﺍﻟﺪﺍﻋﻢ ﻭﺍﺳﻌﺔ ﺍﻟﻨﻄﺎﻕ ﻟﺘﻘﺮﻳﺮ ﺍﳌﻨﺎﻃﻖ ﺍﶈﺘﻤﻠﺔ ﻟﻮﺟﻮﺩ ﺍﻟﺘﺤﻠﺒﺎﺕ ﻭﻓﻘﺎ ﳋﺼﺎﺋﺺ ﻛﻞ ﻣﻨﻄﻘﺔ .ﻭﻗﺪ ﰎ ﲡﺰﺋﺔ
ﺍﻟﺼﻮﺭﺓ ﺍﱃ ﻣﻨﺎﻃﻖ ﺑﺎﺳﺘﺨﺪﺍﻡ ﺧﻮﺍﺭﺯﻣﻴﺔ ﺍﻻﺭﺽ ﺍﳌﺮﺗﻔﻌﺔ – ﺍﻯ ﺧﻂ ﺍﻟﺘﻘﺴﻴﻢ – ﺑﲔ ﲡﻤﻌﲔ ﻣﺎﺋﻴﲔ .ﺣﻘﻖ
ﺍﻻﺳﻠﻮﺏ ﺍﳌﻘﺘﺮﺡ ﺩﻗﻪ ﻗﺪﺭﻫﺎ ٪100ﰱ ﺍﻟﻜﺸﻒ ﻋﻦ ﺍﻟﺘﺤﻠﺒﺎﺕ ﺍﻟﺼﻌﺒﺔ ،ﻭﺩﻗﺔ ﻗﺪﺭﻫﺎ ٪90ﰱ ﺍﻻﺷﺎﺭﺓ ﺍﱃ
ﺍﻟﺼﻮﺭ ﺍﻟﱴ ﻻ ﺗﺸﻤﻞ ﺍﻱ ﺷﻜﻞ ﻣﻦ ﺍﺷﻜﺎﻝ ﺍﻟﺘﻘﺮﺣﺎﺕ ﺍﻟﺴﺎﻃﻌﺔ.
ﺍﻟﻜﻠﻤﺎﺕ ﺍﻟﺮﺋﻴﺴﻴﺔ -ﻣﻌﺎﳉﺔ ﺍﻟﺼﻮﺭ ﺍﻟﻄﺒﻴﺔ ،ﺍﻟﺬﻛﺎﺀ ﺍﻻﺻﻄﻨﺎﻋﻲ ،ﺍﻟﺘﺠﻬﻴﺰ ،ﲢﻠﻴﻞ ﺻﻮﺭ
ﺍﻟﺸﺒﻜﻴﺔ/ﻗﻌﺮ ﺍﻟﻌﲔ ،ﲢﺪﻳﺪ ﻣﺮﻛﺰ ﺍﻟﻘﺮﺹ ﺍﻟﺒﺼﺮﻱ ) ،(ODﺍﺳﺘﻘﻄﺎﻉ ﺍﻻﻭﻋﻴﺔ ﺍﻟﺪﻣﻮﻳﺔ ،ﻛﺸﻒ ﺍﻟﺘﺤﻠﺒﺎﺕ،
ﻛﺸﻒ ﺃﻣﺮﺍﺽ ﺍﻟﺸﺒﻜﻴﺔ ﺍﳌﺘﻌﻠﻘﺔ ﺑﺎﻟﺴﻜﺮﻱ ).(DR
) (Abstract & Keywords in Arabicاﻟﻤﻘﺪﻣﺔ واﻟﻜﻠﻤﺎت اﻟﺮﺋﻴﺴﻴﺔ
ﺍﻥ ﺍﻛﺘﺸﺎﻑ ﻣﻌﺎﱂ ﺍﻟﺸﺒﻜﻴﺔ )ﺍﻻﻭﻋﻴﺔ ﺍﻟﺪﻣﻮﻳﺔ ،ﺍﻟﻘﺮﺹ ﺍﻟﺒﺼﺮﻯ ،ﻭﺍﻟﺒﻘﻌﺔ ﺍﻟﺪﺍﻛﻨﺔ( ﻳﻌﻄﻰ ﺍﻻﻃﺎﺭ
ﺍﻟﺬﻯ ﳝﻀﻰ ﻣﻨﻪ ﺍﻟﺘﺤﻠﻴﻞ ﺍﻻﱄ ﻭﺍﻟﺘﻔﺴﲑ ﺍﻟﺒﺸﺮﻯ ﻟﻠﺸﺒﻜﻴﻪ .ﻭﻋﻠﻴﻪ ﻓﺎﺎ ﺳﺘﻜﻮﻥ ﺫﺍﺕ ﻣﻌﻮﻧﺔ ﻋﺎﻟﻴﺔ ﰲ ﺍﳌﺴﺘﻘﺒﻞ
ﻻﻛﺘﺸﺎﻑ ﻭﻗﻴﺎﺱ ﺍﻻﻣﺮﺍﺽ ﰲ ﺍﳌﻨﺎﻃﻖ ﺍﳌﺬﻛﻮﺭﺓ ﺍﻋﻼﻩ .ﻭﻋﻼﻭﺓ ﻋﻠﻰ ﺫﻟﻚ ﻓﺎﻥ ﺍﳌﻌﺎﱂ ﺍﻷﺳﺎﺳﻴﺔ ﳝﻜﻦ ﺍﻥ ﺗﺴﺘﺨﺪﻡ ﻛﻤﻌﺎﻳﲑ ﺗﺴﻤﺢ ﺑﺎﻟﺘﺨﻠﺺ ﻣﻦ ﺍﻟﺼﻮﺭ ﺫﺍﺕ ﺍﻟﻨﻮﻋﻴﺔ ﺍﻟﺴﻴﺌﺔ ﺟﺪﺍ ﻟﺘﻘﻴﻴﻢ ﺍﻟﺸﺒﻜﻴﺔ.
ﺍﳉﺰﺀ ﺍﻟﺜﺎﱐ ﻣﻦ ﻫﺬﻩ ﺍﻻﻃﺮﻭﺣﻪ ﻳﺘﻨﺎﻭﻝ ﺍﻟﻜﺸﻒ ﻋﻦ ﻣﻮﻗﻊ ﺍﻟﻘﺮﺹ ﺍﻟﺒﺼﺮﻯ ) Optic
” (Disc “ODﺑﺎﻋﺘﺒﺎﺭﻩ ﺍﺣﺪ ﺍﳋﻄﻮﺍﺕ ﺍﻻﺳﺎﺳﻴﺔ ﰱ ﺗﻄﻮﻳﺮ ﻧﻈﻢ ﺍﻟﻔﺮﺯ ﺍﻵﱄ ﻟﻠـ) .(DRﻭﻧﻘﺪﻡ ﰲ ﻫﺬﻩ ﺍﻟﺪﺭﺍﺳﻪ ﺍﺳﻠﻮﺑﺎ ﺍﻟﻴﺎ ﻟﺘﺤﺪﻳﺪ ﻣﻮﻗﻊ ﺍﻟـ) (ODﻣﻦ ﺍﻟﺼﻮﺭ ﺍﻟﺮﻗﻤﻴﻪ ﻟﻠﺸﺒﻜﻴﺔ ﻋﻠﻰ ﺍﺳﺎﺱ ﻣﻄﺎﺑﻘﺔ ﳕﻂ ﺍﻻﲡﺎﻫﺎﺕ
ﺍﳌﺘﻮﻗﻌﺔ ﻟﻼﻭﻋﻴﻪ ﺍﻟﺪﻣﻮﻳﻪ ﰱ ﺍﻟﺸﺒﻜﻴﺔ .ﻭﻣﻦ ﻫﻨﺎ ،ﻗﻘﺪ ﰎ ﺍﻗﺘﺮﺍﺡ ﻣﺮﺷﺢ ﻣﻄﺎﺑﻖ ﺑﺴﻴﻂ ﻟﻴﻄﺎﺑﻖ ﺗﻘﺮﻳﺒﻴﺎ ﺍﲡﺎﻩ
ﺍﻻﻭﻋﻴﺔ ﺍﻟﺪﻣﻮﻳﺔ ﰱ ﳏﻴﻂ ﺍﻟـ) .(ODﺍﻟﻄﺮﻳﻘﺔ ﺍﳌﻘﺘﺮﺣﺔ ﻗﻴﻤﺖ ﺑﺎﺳﺘﺨﺪﺍﻡ ﳎﻤﻮﻋﺔ ﻓﺮﻋﻴﺔ ﻣﻦ ﻗﺎﻋﺪﺓ ﺑﻴﺎﻧﺎﺕ
ﻣﺸﺮﻭﻉ ﺍﻟـ STAREﺗﺘﻀﻤﻦ 81ﺻﻮﺭﺓ ﻟﻠﺸﺒﻜﻴﺔ ﺍﻟﻄﺒﻴﻌﻴﺔ ﻭﺍﳌﺮﻳﻀﺔ ،ﻭﻗﺪ ﺍﺳﺘﺨﺪﻣﺖ ﻣﻦ ﻗﺒﻞ ﺑﻮﺍﺳﻄﺔ
ﲨﻠﺔ ﺍﻻﻋﻤﺎﻝ ﺍﻟﺴﺎﺑﻘﺔ ﻟﺘﺤﺪﻳﺪ ﻣﻮﻗﻊ ﺍﻟـ) .(ODﻭﻟﻘﺪ ﰎ ﲢﺪﻳﺪ ﻣﺮﻛﺰ ﺍﻟـ) (ODﺑﻄﺮﻳﻘﺔ ﺻﺤﻴﺤﺔ ﰲ 80
ﻣﻦ ﺍﺻﻞ 81ﺻﻮﺭﺓ ) .(٪98.77ﻭﺑﺎﻻﺿﺎﻓﺔ ﺍﱃ ﺫﻟﻚ ،ﰎ ﲢﺪﻳﺪ ﻣﺮﻛﺰ ﺍﻟـ) (ODﺑﻄﺮﻳﻘﺔ ﺻﺤﻴﺤﺔ ﰲ ﲨﻴﻊ ﺍﻟﺼﻮﺭ ﺍﻟـ (٪100) 40ﺑﺎﺳﺘﺨﺪﺍﻡ ﻗﺎﻋﺪﺓ ﺑﻴﺎﻧﺎﺕ ﺍﻟـ DRIVEﺍﳌﺘﺎﺣﺔ ﻋﻼﻧﻴﻪ.
ﺍﳉﺰﺀ ﺍﻟﺜﺎﻟﺚ ﻣﻦ ﺍﻻﻃﺮﻭﺣﻪ ﻳﺘﻨﺎﻭﻝ ﺍﺳﺘﻘﻄﺎﻉ ﺍﻻﻭﻋﻴﺔ ﺍﻟﺪﻣﻮﻳﺔ ﰱ ﺍﻟﺸﺒﻜﻴﺔ ) (RVﻛﺮﻛﻴﺰﺓ ﺍﺧﺮﻯ
ﻣﻦ ﺍﻟﺮﻛﺎﺋﺰ ﺍﻻﺳﺎﺳﻴﺔ ﻟﺘﻄﻮﻳﺮ ﻧﻈﻢ ﺍﻟﻔﺮﺯ ﺍﻻﱃ ﻟﻠﺸﺒﻜﻴﺔ ،ﻭﺫﻟﻚ ﻻﻥ ﺍﻟـ) (RVﻫﻰ ﲟﺜﺎﺑﺔ ﺍﳌﻌﻠﻢ ﺍﻟﺮﺋﻴﺴﻰ ﻭﺍﻟﺬﻯ ﻳﺒﺪﺃ ﻣﻦ ﻋﻨﺪﻩ ﺍﳌﺰﻳﺪ ﻣﻦ ﺍﻟﺘﺤﻠﻴﻞ .ﻣﺆﺧﺮﺍ ،ﺍﺛﺒﺖ ﺍﻟﺘﺼﻨﻴﻒ ﺍﳌﺸﺮﻑ ﻋﻠﻴﻪ ﺍﻧﻪ ﺍﻛﺜﺮ ﻛﻔﺎﺀﻩ ﻭﺩﻗﺔ ﻟﻌﻤﻠﻴﺔ ﺍﻻﺳﺘﻘﻄﺎﻉ .ﻋﻼﻭﺓ ﻋﻠﻰ ﺫﻟﻚ ،ﻓﺎﻥ ﻣﻴﺰﺍﺕ ﻭﻣﻌﺎﱂ ﺟﺪﻳﺪﺓ ﻗﺪ ﺍﺳﺘﺨﺪﻣﺖ ﰲ ﲨﻠﺔ ﺍﻟﺪﺭﺍﺳﺎﺕ ﺍﳊﺪﻳﺜﺔ ﺗﻈﻬﺮ
ﻣﻘﺪﺭﺓ ﻋﺎﻟﻴﺔ ﰱ ﺍﻟﻔﺼﻞ ﺑﲔ ﻓﺌﱴ ﺍﻻﻭﻋﻴﺔ ﻭﻏﲑ ﺍﻻﻭﻋﻴﺔ .ﺍﺳﺘﻌﻤﻞ ﻫﺬﺍ ﺍﻟﻌﻤﻞ ﻣﺎﻛﻴﻨﺔ ﺍﳌﺘﺠﻪ ﺍﻟﺪﺍﻋﻢ ﻭﺍﺳﻌﺔ
ﺍﻟﻨﻄﺎﻕ ) (Large-Scale SVMﻻﺳﺘﻘﻄﺎﻉ ﺍﻟـ) (RVﺁﻟﻴﺎ ،ﻣﺴﺘﺨﺪﻣﺎ ﺧﻠﻴﻄﺎ ﻣﻦ ﲢﺴﻴﻨﺎﺕ
"ﻣﻮﳚﺎﺕ ﻏﺎﺑﻮﺭ ﺛﻨﺎﺋﻴﺔ ﺍﻻﺑﻌﺎﺩ ،ﺍﻋﻠﻰ ﺍﻟﻘﺒﻌﻪ ،ﻭ ﺍﻻﺳﺲ ﺍﳍﺴﻴﻪ" ﻛﻤﻼﻣﺢ ﻟﻜﻞ ﻧﻘﻄﺔ ﻣﻨﻔﺮﺩﺓ ﻋﻠﻰ ﺷﺎﺷﺔ
ﺍﻟﻜﻤﺒﻴﻮﺗﺮ ) .(pixelﺍﻟﻄﺮﻳﻘﺔ ﺍﳌﻘﺘﺮﺣﺔ ﻗﻠﻠﺖ ﻋﺪﺩ pixelsﺍﻟﺘﺪﺭﻳﺐ ﺑﺸﻜﻞ ﻣﻠﺤﻮﻅ ﺣﻴﺚ ﺍﻧﻨﺎ
ﳓﺘﺎﺝ ﺍﱃ 2000ﻓﻘﻂ ﺑﺪﻻ ﻣﻦ ﺍﻟـ 1ﻣﻠﻴﻮﻥ ﺍﳌﻌﺮﻭﺿﺔ ﻣﺆﺧﺮﺍ ﰲ ﺍﻟﺪﺭﺍﺳﺎﺕ .ﻭﻧﺘﻴﺠﺔ ﻟﺬﻟﻚ ،ﺍﳔﻔﺾ ﻣﺘﻮﺳﻂ ﻓﺘﺮﺓ ﺍﻟﺘﺪﺭﻳﺐ ﺍﱃ 3.75ﺛﻮﺍﻥ ﺑﺪﻻ ﻣﻦ ﺍﻟـ 9ﺳﺎﻋﺎﺕ ﺍﻟﱵ ﺳﺒﻖ ﺗﺴﺠﻴﻠﻬﺎ ﰲ ﺍﻟﺪﺭﺍﺳﺎﺕ .ﺍﻣﺎ
ﺍﺳﺘﻘﻄﺎﻉ ﺍﻟـ) (RVﻣﻦ ﺻﻮﺭﺓ ﻓﻴﺤﺘﺎﺝ ﺍﱃ 30ﺛﺎﻧﻴﺔ ﻓﻘﻂ .ﻭﺟﺪﻳﺮ ﺑﺎﻟﺬﻛﺮ ﺃﻥ ﺍﻟﺘﺪﺭﻳﺐ ﺑﺎﺳﺘﺨﺪﺍﻡ ) (Abstract & Keywords in Arabicاﻟﻤﻘﺪﻣﺔ واﻟﻜﻠﻤﺎت اﻟﺮﺋﻴﺴﻴﺔ
ﺍﳌﻘﺪﻣﺔ ﺍﻟﺴﻜﺮﻱ ﻫﻮ ﻣﺮﺽ ﻳﺆﺛﺮ ﻋﻠﻰ ﺣﻮﺍﱃ 5.5ﰱ ﺍﳌﺎﺋﺔ ﻣﻦ ﺳﻜﺎﻥ ﺍﻟﻌﺎﱂ .ﰲ ﻣﺼﺮ ﺣﻮﺍﱄ 9ﻣﻼﻳﲔ )ﺃﻯ ﺃﻛﺜﺮ
ﻣﻦ ٪13ﻣﻦ ﺍﻟﺴﻜﺎﻥ ≤ 20ﻋﺎﻣﺎ( ﺳﻴﺼﺎﺑﻮﻥ ﺑﺎﻟﺴﻜﺮﻱ ﲝﻠﻮﻝ ﻋﺎﻡ ،2025ﰲ ﺣﲔ ﺍﻥ ﺍﻟﺪﺭﺍﺳﺎﺕ ﺍﻻﺧﲑﺓ ﻣﻦ ﻋﻤﺎﻥ ﻭﺑﺎﻛﺴﺘﺎﻥ ﺗﺸﲑ ﺍﱃ ﺍﻥ ﺫﻟﻚ ﻗﺪ ﻳﻜﻮﻥ ﻇﺎﻫﺮﺓ ﺍﻗﻠﻴﻤﻴﻪ .ﻭﺑﺎﻟﺘﺎﱄ ،ﻓﺎﻥ ﺣﻮﺍﱄ ٪10ﻣﻦ ﲨﻴﻊ ﻣﺮﺿﻰ ﺍﻟﺴﻜﺮﻱ ﻣﺼﺎﺑﻮﻥ ﺑﺄﻣﺮﺍﺽ ﺍﻟﺸﺒﻜﻴﺔ ﺍﳌﺘﻌﻠﻘﺔ ﺑﺎﻟﺴﻜﺮﻱ )Diabetic-
”(Retinopathy “DR؛ ﺍﺣﺪ ﺍﻛﺜﺮ ﻣﻀﺎﻋﻔﺎﺕ ﻣﺮﺽ ﺍﻟﺴﻜﺮﻱ ﻭﺍﻟﱵ ﻫﻲ ﺍﻟﺴﺒﺐ ﺍﻟﺮﺋﻴﺴﻲ
ﻟﻠﻌﻤﻰ ﰲ ﺍﻟﻌﺎﱂ ﺍﻟﻐﺮﰊ ،ﻭﻫﺬﺍ ﳛﺘﻤﻞ ﺃﻥ ﻳﻜﻮﻥ ﺻﺤﻴﺤﺎ ﰲ ﻫﻮﻧﻎ ﻛﻮﻧﻎ ،ﻭﻣﺼﺮ ﺃﻳﻀﺎ .ﻋﻼﻭﺓ ﻋﻠﻰ ﺫﻟﻚ ،ﺍﻥ
ﻣﺮﺿﻰ ﺍﻟﺴﻜﺮﻱ ﻣﻦ ﺍﳌﺘﻮﻗﻊ ﺍﻥ ﻳﺼﺎﺑﻮﺍ ﺑﺎﻟﻌﻤﻰ ﲞﻤﺴﺔ ﻭﻋﺸﺮﻳﻦ ﺿﻌﻔﺎ ﻣﻘﺎﺭﻧﺔ ﺑﻐﲑ ﺍﳌﺼﺎﺑﲔ ﺑﺎﻟﺴﻜﺮﻯ .ﻭﻧﻈﺮﺍ
ﻟﺘﺰﺍﻳﺪ ﻋﺪﺩ ﺍﳌﺮﺿﻰ ،ﻭﺃﻃﺒﺎﺀ ﺍﻟﻌﻴﻮﻥ ﻏﲑ ﻛﺎﻓﲔ ﻟﻔﺤﺼﻬﻢ ،ﻓﺎﻥ ﺍﻟﻔﺤﺺ ﺍﻻﺗﻮﻣﺎﺗﻴﻜﻰ ﳝﻜﻦ ﺍﻥ ﻳﻘﻠﻞ ﻣﻦ ﺧﻄﺮ ﺍﻟﻌﻤﻰ ﺑﻨﺴﺒﺔ ٪50ﻭﻳﻮﻓﺮ ﻭﻓﻮﺭﺍﺕ ﻛﺒﲑﺓ ﰲ ﺍﻟﺘﻜﺎﻟﻴﻒ ،ﻭﳜﻔﺾ ﺍﻟﻀﻐﻂ ﻋﻠﻰ ﺍﻟﺒﲎ ﺍﻟﺘﺤﺘﻴﺔ ﻭﺍﳌﻮﺍﺭﺩ ﺍﳌﺘﺎﺣﺔ.
ﺍﻥ ﺻﻮﺭ ﺍﻟﺸﺒﻜﻴﺔ ﺍﳒﺢ ﺑﻜﺜﲑ ﰱ ﺍﻛﺘﺸﺎﻑ ﺍﻟﺴﻜﺮﻯ ﻣﻦ ﺍﻟﻜﺸﻒ ﺍﳌﺒﺎﺷﺮ
) .(ophthalmoscopyﻛﻤﺎ ﺍﻥ ﺍﻟﺼﻮﺭ ﺍﻟﺮﻗﻤﻴﻪ ﻟﻘﻌﺮ ﺍﻟﻌﲔ ﻻ ﲢﺘﺎﺝ ﺍﱃ ﺣﻘﻦ ﺍﳉﺴﻢ ﲟﺎﺩﺗﻰ
ﺍﻟـ fluorescein dyeﺍﻭ ﺍﻟـ ،indocyanine greenﻭﺑﺎﻟﺘﺎﱄ ﻻ ﺗﺘﻄﻠﺐ ﺍﻱ ﻣﺘﺨﺼﺼﲔ .ﺍﻥ ﺻﻮﺭ ﻗﻌﺮ ﺍﻟﻌﲔ ﺍﻟﺮﻗﻤﻴﻪ ﲢﻠﻞ ﺭﻭﺗﻴﻨﻴﺎ ﺑﻮﺍﺳﻄﺔ ﻧﻈﻢ ﺍﻟﻔﺮﺯ ،ﻭﻧﻈﺮﺍ ﻟﻌﻤﻠﻴﺔ ﺍﻗﺘﻨﺎﺀ ﻫﺬﻩ ﺍﻟﺼﻮﺭ ،ﻓﺎﻥ ﻛﺜﲑﺍ ﻣﻨﻬﺎ ﺫﻭ ﻧﻮﻋﻴﺔ ﺭﺩﻳﺌﺔ ﺗﻌﻮﻕ ﺍﳌﺰﻳﺪ ﻣﻦ ﺍﻟﺘﺤﻠﻴﻞ .ﺍﻟﺪﺭﺍﺳﺎﺕ ﺍﳊﺪﻳﺜﺔ ﻻ ﺗﺰﺍﻝ ﺗﺼﺎﺭﻉ ﻣﺴﺄﻟﺔ
ﲡﻬﻴﺰ ﺻﻮﺭ ﺍﻟﺸﺒﻜﻴﺔ ،ﻭﻳﺮﺟﻊ ﺫﻟﻚ ﺃﺳﺎﺳﺎ ﺍﱃ ﻗﻠﺔ ﺍﳌﻮﺟﻮﺩ ﻣﻦ ﺍﻟﺪﺭﺍﺳﺎﺕ ﺍﳌﻘﺎﺭﻧﺔ .ﻋﻼﻭﺓ ﻋﻠﻰ ﺍﻥ ﺍﻻﺳﺎﻟﻴﺐ ﺍﳌﺘﺎﺣﺔ ﻟﺘﺠﻬﻴﺰﻫﺎ ﻻ ﳚﺮﻱ ﺗﻘﻴﻴﻤﻬﺎ ﻋﻠﻰ ﻧﻄﺎﻕ ﻗﻮﺍﻋﺪ ﺍﻟﺒﻴﺎﻧﺎﺕ ﺍﳌﻌﻴﺎﺭﻳﺔ ﻭﺍﳌﺘﺎﺣﺔ ﻟﻠﺠﻤﻬﻮﺭ. ﺍﳉﺰﺀ ﺍﻷﻭﻝ ﻣﻦ ﻫﺬﻩ ﺍﻻﻃﺮﻭﺣﻪ ﻳﻨﺎﻗﺶ ﺍﳌﻨﻬﺠﻴﺎﺕ ﺍﻷﺭﺑﻊ ﺍﻟﺮﺋﻴﺴﻴﺔ ﻟﺘﺠﻬﻴﺰ ﺻﻮﺭ ﺍﻟﺸﺒﻜﻴﺔ )ﺍﻧﺘﺎﺝ
ﺍﻟﻘﻨﺎﻉ ،ﻣﺴﺎﻭﺍﺓ ﺍﻻﺿﺎﺀﻩ ،ﺗﻌﺰﻳﺰ ﺍﻟﺘﺒﺎﻳﻦ ،ﻭﺗﻄﺒﻴﻊ ﺍﻟﻠﻮﻥ( ،ﻭﺗﺄﺛﲑﻫﺎ ﻋﻠﻰ ﺗﺸﺮﻳﺢ ﺍﻟﺸﺒﻜﻴﺔ ﺁﻟﻴﺎ .ﰲ ﻛﻞ
ﻣﻨﻬﺠﻴﻪ ،ﲤﺖ ﻣﻘﺎﺭﻧﺔ ﻟﻘﻴﺎﺱ ﺍﻷﺩﺍﺀ ﺑﲔ ﺍﻻﺳﺎﻟﻴﺐ ﺍﳌﺘﺎﺣﺔ ﻋﻠﻰ ﺃﺳﺎﺱ ﺍﻟﻘﻴﺎﺳﺎﺕ ﺍﳌﻨﺎﺳﺒﻪ ﻭﺑﺎﺳﺘﺨﺪﺍﻡ ﺍﺛﻨﺘﲔ ﻣﻦ
ﻗﻮﺍﻋﺪ ﺍﻟﺒﻴﺎﻧﺎﺕ ﺍﻟﺸﺒﻜﻴﺔ ﺍﳌﺘﺎﺣﺔ ﻋﻠﻨﺎ .ﻭﺑﺎﻻﺿﺎﻓﺔ ﺍﱃ ﺫﻟﻚ ،ﺍﻗﺘﺮﺣﻨﺎ "ﺍﻟﺘﻄﺒﻴﻊ ﺍﻟﺸﺎﻣﻞ" ﻭ"ﺗﻌﺰﻳﺰ ﺍﻟﺘﺒﺎﻳﻦ ﺍﶈﻠﻲ
ﺍﳌﺴﺒﻮﻕ ﲟﺴﺎﻭﺍﺓ ﺍﻻﺿﺎﺀﻩ" ﻭﺍﻟﻠﺘﲔ ﺳﺠﻠﺘﺎ ﻧﺘﺎﺋﺞ ﻣﻘﺒﻮﻟﺔ ﻋﻨﺪ ﺗﻄﺒﻴﻘﻬﻤﺎ ﻟﺘﻄﺒﻴﻊ ﺍﻟﻠﻮﻥ ﻭﺗﻘﻮﻳﺔ ﺍﻟﺘﺒﺎﻳﻦ ﻋﻠﻰ
ﺍﻟﺘﻮﺍﱄ. ) (Abstract & Keywords in Arabicاﻟﻤﻘﺪﻣﺔ واﻟﻜﻠﻤﺎت اﻟﺮﺋﻴﺴﻴﺔ
This page is left blank intentionally
2
References
ﻛﻠﻴﺔ ﺍﳊﺎﺳﺒﺎﺕ ﻭﺍﳌﻌﻠﻮﻣﺎﺕ ﻗﺴﻢ ﻋﻠﻮﻡ ﺍﳊﺎﺳﺐ
ﺗﻘﻨﻴﺎﺕ ﺍﻟﺬﹼﻛﺎﺀ ﺍﻹﺻﻄﻨﺎﻋ ﻰ ﻟﺘﺼﻨﻴﻒ ﺍﻟﻨﻤﺎﺫﺝ ﺍﻟﻌﻴﻨﻴﺔ
ﻋﻤﺮﻭ ﺃﲪﺪ ﺻﱪﻯ ﻋﺒﺪ ﺍﻟﺮﲪﻦ ﻏﻨﻴﻢ
ﺍﳌﺸﺮﻓﻮﻥ: ﺍ .ﺩ /ﻋﺎﻃﻒ ﺯﻛﻰ ﻏﻠﻮﺵ ﺍﺳﺘﺎﺫ ﰱ ﻋﻠﻮﻡ ﺍﳊﺎﺳﺐ ،ﻛﻠﻴﺔ ﺍﳊﺎﺳﺒﺎﺕ ﻭﺍﳌﻌﻠﻮﻣﺎﺕ ،ﺟﺎﻣﻌﺔ ﺣﻠﻮﺍﻥ ،ﺍﻟﻘﺎﻫﺮﺓ، ﻣﺼﺮ .
ﺍ .ﻡ .ﺩ /ﻋﻠﻴﺎﺀ ﻋﺒﺪﺍﳊﻠﻴﻢ ﻋﺒﺪﺍﻟﺮﺍﺯﻕ ﻳﻮﺳﻒ ﺍﺳﺘﺎﺫ ﻣﺴﺎﻋﺪ ﰱ ﻋﻠﻮﻡ ﺍﳊﺎﺳﺐ ،ﻛﻠﻴﺔ ﺍﳊﺎﺳﺒﺎﺕ ﻭﺍﳌﻌﻠﻮﻣﺎﺕ ،ﺟﺎﻣﻌﺔ ﺣﻠﻮﺍﻥ، ﺍﻟﻘﺎﻫﺮﺓ ،ﻣﺼﺮ. ﺩ /ﺣﺴﺎﻡ ﺍﻟﺪﻳﻦ ﳏﻤﻮﺩ ﺑﻜﲑ ﻣﺪﺭﺱ ﰱ ﻃﺐ ﺍﻟﻌﻴﻮﻥ ،ﻛﻠﻴﺔ ﺍﻟﻄﺐ ،ﺟﺎﻣﻌﺔ ﺍﻟﻘﺎﻫﺮﺓ ،ﺍﻟﻘﺎﻫﺮﺓ ،ﻣﺼﺮ.
ﺭﺳﺎﻟﺔ ﻣﻘﺪﻣﺔ ﺍﱃ ﺟﺎﻣﻌﺔ ﺣﻠﻮﺍﻥ ﲟﻮﺟﺐ ﻣﺘﻄﻠﺒﺎﺕ ﺩﺭﺟﺔ ﻣﺎﺟﺴﺘﲑ ﺍﻟﻌﻠﻮﻡ ﰱ ﻋﻠﻮﻡ ﺍﳊﺎﺳﺐ ﻣﻦ ﻛﻠﻴﺔ ﺍﳊﺎﺳﺒﺎﺕ ﻭﺍﳌﻌﻠﻮﻣﺎﺕ ،ﻗﺴﻢ ﻋﻠﻮﻡ ﺍﳊﺎﺳﺐ.
ﻣﺎﻳـﻮ )ﺃﻳـﺎﺭ( 2007
