CHAPTER - 1

INTRODUCTION

Medical imaging has revolutionized the medicine by providing cost-

efficient healthcare and effective diagnosis in all major disease areas.

Medical imaging allows scientists and physicians to understand

potential life-saving information using less invasive techniques.

In medical imaging the quality of the image acquisition and the

image interpretation determines the accuracy of diagnosis. Computers

have a huge impact on the acquisition of medical images. They

perform multi-pronged functions like controlling imaging hardware,

performing reconstruction, post-processing of the image data and

storing the scans. In contrast, the role of computers in the

interpretation of medical images has so far been limited.

Interpretation remains an almost exclusively human domain.

Recent years have witnessed pioneering work in the area of medical

imaging. Applications that can interpret an image are being developed,

which in turn can aid a physician in detecting possible subtle

abnormalities. The computer indicates places in the image that

require extra attention from the physician because they could be

abnormal. These technologies are called Computer Aided Diagnosis

(CAD). Studies on CAD systems show that CAD can be helpful to

improve diagnostic accuracy of physicians and lighten the burden of

increasing workload. The most established CAD applications in

medical fields are the use of automated systems in mammography,

chest computed tomography and radiography.

This thesis describes components of an automatic system that can

aid in the detection of diabetic retinopathy. Diabetic retinopathy is an

eye disease and a general complication of diabetes that causes vision

loss, if left undiagnosed at the initial stage. As the number of diabetes

affected people is increasing worldwide, the need for automated

detection methods of diabetic retinopathy will increase as well. To

automatically detect diabetic retinopathy, a computer has to interpret

and analyze digital images of the retina.

The Fundus Image Analysis system described in this thesis is

developed to assist ophthalmologists diagnosis by providing second

opinion and also functions as an automatic tool for the mass

screening of diabetic retinopathy. Colour fundus images are used by

ophthalmologists to study eye diseases like diabetic retinopathy.

Extraction of the normal features like optic disk, fovea and blood

vessels; and abnormal features like exudates, cottonwool spots,

microaneurysms and hemorrhages from colour fundus images are

used in fundus image analysis system for comprehensive analysis and

grading of diabetic retinopathy. This CAD system also provides the

spatial distribution of abnormalities based on fovea such that an

ophthalmologist can make a detailed diagnosis.

This introductory Chapter presents some background information

on the anatomy of the eye, diabetic retinopathy, and diabetic

retinopathy screening.

1.1. ANATOMY OF THE EYE

The human eye functions like a camera. Human eye receives light,

converting it into an electric signal and transmitting the same to the

brain with the help of optic nerve where the electrical signal is

converted into vision. Figure 1.1 illustrates the cross-section of

human eye and points out its major components. Amongst the various

ocular structures, only the anatomical parts of the retina are

explained which are more appropriate to this research work.

Light enters the human eye through the pupil and is directed on

the retina that is similar to a camera film. Like film, the retina is

comprised of several layers with different functions. The first layer

that receives light is called the nerve fiber layer. The majority of the

retinal blood vessels are under this layer which nourishes the internal

parts of the retina. The outermost layer contains millions of

photoreceptors responsible for receiving light rays, converting them

into electrical impulses. These electrical impulses are transformed into

images when they are transmitted to the brain.

The outlying parts of the retina are responsible for peripheral

vision. Macula is the central area of the retina, temporal to the optic

disk. It is responsible to have fine central vision and colour vision.

The center of macula is called fovea. This region of the retina is the

most sensitive region. The diameter of the macula is about 4 to 5 mm,

which enable us in appreciating details and performing tasks, like

reading.

Optic disk (or optic nerve head) is the bright yellowish disk, from

which, blood vessels and optic nerve fibres emerge. Optic disk

transmits electrical impulses from the retina to the brain and it is

about 3mm nasal to fovea. It measures 1.5 to 2 mm in diameter.

The retinal blood vessels are originated from the central retinal

artery and vein that lie in the optic nerve head. These blood vessels

nourish the internal parts of retina and radiate out from the optic

disk. Fig.1.2 shows a typical normal retina image with highlighted

areas of macula, fovea, optic disk and blood vessels.

Fig1.1. A Cross Section of the Human Eye

1.2. DIABETIC RETINOPATHY

Diabetic retinopathy is the prime cause of vision loss amongst the

working age population of the developing and the developed countries.

Diabetic patients are 25 times more probable to become blind than

non-diabetic patients [1]. Diabetic retinopathy is a complication of

diabetes to the retina. Both the forms of diabetes i.e. diabetes mellitus

and diabetes insepidous, leads to diabetic retinopathy eventually after

some time. It is a very asymptomatic disease in the early stages and it

could lead to permanent vision loss if untreated for long time. The

problem here is the patients may not know about it until it reaches

advanced stages. Once it reaches advanced stages vision loss becomes

Fig.1.2. A Typical Retinal Image from the Left Eye Showing Retinal

Vasculature, Optic Disk, Macula and Fovea.

inevitable. As diabetic retinopathy is the third major cause of

blindness particularly in India, there is an immediate requirement to

develop efficient diagnosis methods for this problem. The age of onset

and the duration of the diabetes are the two most important issues

that determine the incidence of diabetic retinopathy. Among the

patients below the age of 30 years, when first diagnosed with diabetes,

the prevalence is 17% during the first 5 years. This increases to 97%

after 15 years of diabetes [2]. Amongst the patients above the age of

30 years at the onset of diabetes, 20% have showed signs of

retinopathy immediately after presentation and this increased to 78%

after 15 years of diabetes [3].

Diabetic retinopathy occurs because of microangiopathy which in

turn affects the retinal precapillary arterioles, capillaries and venules.

It is caused by microvascular leakages from the breakdown of the

internal blood-retinal barrier and microvascular occlusion. Due to the

progressive damage of the microvascular system, loss of vision and

blindness can occur as shown in Fig. 1.3.

Microaneurysms are the first clinically noticeable signs of diabetic

retinopathy. They appear as small red dots of 10 to 100 microns

diameter. Microaneurysms exist normally temporal to the macula

(Fig. 1.4(a)). Microaneurysms arise due to high sugar levels in the

blood which causes the walls of tiny blood vessels to distend.

As the disease progresses, microaneurysms will be ruptured. This

results in retinal hemorrhages either superficially or in deeper layers

of the retina (Fig. 1.4(a)). As the retinal blood vessels become more

damaged and permeable, their number will increase. Retinal

hemorrhages look either as small red dots or blots identical to

microaneurysms or as larger flame-shaped hemorrhages.

The vessels besides leaking blood also cause the leakage of lipids

and proteins paving way for the appearance of small bright dots called

exudates (Fig. 1.4(b)). They are seen on the retina as typical bright,

reflective white or cream coloured lesions that indicate increased

blood vessel permeability and an allied risk of retinal edema. If this

takes place on the macula region vision may be lost.

As the disease advances further, multiple small patches of the

retina become ischemic deprived of blood. These ischemic regions are

observable on the retina as fluffy whitish blobs called cottonwool spots

(Fig. 1.4(c)). As a response to the development of ischemic areas, the

eye starts growing new blood vessels to provide more oxygen to the

retina. These newly grown blood vessels, called Neovascularization,

Fig.1.3. Effect of Diabetic Retinopathy on Vision (a) Without

Retinopathy (b) With Retinopathy

(a)

(b) (c)

Fig.1.4. Abnormal Diabetic Retinopathy Images (a) An Image Containing Microaneurysms and Hemorrhages (b) Diabetic

Retinopathy with Retinal Exudates in Macula Region (c) Diabetic

Retinopathy with Cottonwool Spots.

are delicate and weak having a greater risk of rupturing. These newly

developed blood vessels cause large hemorrhages than normal vessels.

1.3. SCREENING FOR DIABETIC RETINOPATHY

According to recent reports incidence of diabetes is about 12 to

14% in the urban population of India of which over 20% of patients

are likely to be suffering from diabetic retinopathy [4]. In the rural

population, the prevalence of diabetes is about 5%. Early detection of

diabetic retinopathy and treatment can prevent visual impairment and

most of the patients can be saved from vision loss. Screening is an

effective way for early detection of diabetic retinopathy. Studies have

revealed that people who suffer from diabetes benefit from regularly

attending a screening session [5-7]. In this screening session the

retinas of both eyes are examined by an ophthalmologist and if

diabetic retinopathy is detected the patient can be followed up.

Examples of large scale screening programs conducted in India are

Telescreening for Diabetic Retinopathy: Taking eye care to rural

south India and Comprehensive diabetic retinopathy Project [8].

Traditionally the retina is observed either directly using an

ophthalmoscope or indirectly via digital photographs that are taken by

a fundus camera. The ophthalmoscope is a small portable instrument

containing a light source and a set of lenses. A digital fundus camera

is a low power microscope with a camera attached and designed to

photograph the interior layers of the eye. For large scale screening, it

has been shown that fundus images are more reliable than

ophthalmoscope in the detection of diabetic retinal lesions [912].

Digital fundus photography allows instantaneous examination of the

retina as and when necessary with a quick storage, access to the

images and decoupling of the acquisition and interpretation stages of

the screening.

In order to identify diabetic retinopathy at the beginning stages,

screening should be done periodically. Moreover, majority of the

diabetic population is living in the vast rural India, so large numbers

of screening camps are needed. This large scale repeated mass

screenings lead to generation of numerous fundus images that are to

be evaluated. It requires lot of time and many technicians, which are

currently not available. Typically, 90% of the images thus generated in

such scenario would be normal. Hence a fully automated, computer

based diabetic retinopathy recognition systems which can filter out

many of these normal images will save the time and reduce the

number of retinal images that are to be examined by the physicians.

With rapid development of technology, latest techniques for

screening are available which include digital photographic and

computerized techniques for detection and assessment of retinopathy.

A computer might be employed to make a selection of the images

stored on the central server. Possible suspect images (patients) could

be shown to the ophthalmologist while certainly normal images could

be stored immediately. This could potentially lower the total workload

of the ophthalmologists. This thesis mainly focuses on the

development of an automated Fundus Image Analysis system for such

a pre-selection of retinal images in diabetic retinopathy screening.

1.4. AIMS AND OBJECTIVES

The main aim of this research is to develop reliable and accurate

image processing and pattern recognition methods for automatic

fundus image analysis to aid ophthalmologists diagnosis and to be

used as an automatic tool for the mass screening of diabetic

retinopathy. Given a low quality colour fundus image, the proposed

fundus image analysis system should extract the fundal landmark

features such as the optic disk, fovea and the retinal blood vessels as

reference coordinates before the system can achieve more complex

tasks of identifying pathological entities such as hard exudates,

cottonwool spots, hemorrhages and microaneurysms. The system

must be able to do it all the time irrespective of variability in colour,

illumination levels and amount of noise. It would be good if the

algorithms of the system are fast and computationally light.

The objectives of this research work are

To Segment Retinal Blood Vessel Tree - This is a must for the

identification of optic disk and for the elimination of vascular

structures from the search of possible non-vascular lesions.

To Identify the Position of the Optic Disk - It provides the main

landmark of the retinal coordinates and to exclude it from the set

of possible lesions.

To Detect the Contour of the Optic Disk - To assess the progress

of eye disease and treatment results.

To Identify Fovea, Vascular Arcade and to Establish Polar

Fundal Coordinate System Centered on Fovea - This co-ordinate

system will be helpful to the ophthalmologists to analyze the

severity of diabetic retinopathy.

To Detect Bright Lesions such as Hard Exudates and

Cottonwool Spots The bright lesions are the visible sign of

diabetic retinopathy and indicators of co-existent retinal edema.

Automated early detection of bright lesions can assist

ophthalmologists prevent the spread of the disease more effectively.

To Detect Red Lesions such as Microaneurysms and

Hemorrhages As red lesions are the first clinically observable

lesions indicating diabetic retinopathy, their detection is critical for

a diabetic retinopathy screening system.

1.5. FUNDUS IMAGE ANALYSIS SYSTEM

To accomplish the above mentioned aims and objectives a Fundus

Image Analysis system is developed in this thesis which is as shown in

Fig. 1.5. The input colour retinal image is analysed automatically and

a comprehensive assessment of the severity of diabetic retinopathy

and macular edema is derived after analysis. The proposed fundus

image analysis system consists of six main components. The binary

vasculature extracted by the blood vessel segmentation component is

used in optic disk detection, fovea detection and red lesion detection.

The optic disk localization and contour detection component finds the

location and boundary of the optic disk. The location and the

boundary of the optic disk are used to detect fovea and bright lesions.

Detection of

Microaneurysms

and Hemorrhages

Fovea

Detection

Optic Disc Localization

and Contour Detection

Blood Vessel

Segmentation

Detection of

Exudates and

Cottonwool spots

Diagnosis of

Diabetic Retinopathy

Input: Colour Fundus Image

Output

Fig.1.5. Fundus Image Analysis System

The location of optic disk is employed to remove any spurious bright

lesion detections on the optic disk. The fovea detection component

detects vascular arcade, macula and fovea. Based on the location of

fovea a fundal coordinate system is set up that is used to find the

severity of diabetic retinopathy. The bright lesion detection component

detects exudates and cottonwool spots. The red lesion detection

component identifies microaneurysms and hemorrhages. The

outcomes of these two components are used in diagnosis of diabetic

retinopathy component. The fundus image analysis system grades

diabetic retinopathy and macular edema based on the detection of

these lesions and this system also provides the spatial distribution of

abnormalities based on fovea such that an ophthalmologist can make

a detailed diagnosis.

1.6. ORGANIZATION OF THE THESIS

The thesis is organized in nine chapters including Introduction in

Chapter 1. In Chapter 2, the literature corresponding to retinal image

analysis systems and algorithms is reviewed.

Chapter 3 describes the proposed method for automated

segmentation of blood vessels in fundus images. A method based on

relative local entropy is presented to segment the blood vessels.

Chapter 4 presents the proposed methods for localization and

contour detection of the optic disk in a fundus image. Two methods

are investigated to localize the optic disk. First method is based on

finding vessel branch which has maximum number of blood vessels

and the second method employs principal component analysis to

localize the optic disk. Geometric active contour model with new

variational formulation is applied to detect a more accurate optic disk

boundary.

Chapter 5 explains the proposed approach used for localization of

macula and fovea. Using the segmented vessel network as input,

vascular arcade and horizontal raphe of the retina are determined.

Then macula and centre of macula (fovea) are localized from the

horizontal raphe. Finally, a fundal coordinate system is established

centered on fovea.

An automatic bright lesion detection system is proposed in

Chapter 6. The colour retinal image is preprocessed using local

contrast enhancement technique to enhance the contrast of the

retinal image. As optic disk appears similar to hard exudates it is

eliminated using the entropy feature. Spatially Weighted Fuzzy C-

Means (SWFCM) clustering is applied to extract candidate bright

lesions. Then true bright lesions are classified from bright non-lesions

using K-Nearest Neighbourhood (KNN) and Support Vector Machine

(SVM) classifiers. A blood vessel segmentation algorithm based on

SWFCM clustering is also presented in this chapter.

Chapter 7 presents the proposed automatic red lesion detection

system. The candidate red lesions are extracted based on a hybrid

approach combining mathematical morphology based segmentation

and detection using matched filtering and relative local entropy based

thresholding. Next, true red lesions are separated from red non-

lesions using KNN and SVM classifiers.

In Chapter 8, the proposed fundus image analysis system is

presented. This system is evaluated and compared to existing diabetic

retinopathy screening systems. Finally, Chapter 9 summarizes the

principal contributions of this research work. Suggestions for future

work are also presented in the concluding remarks.