Retinal Image Analysis & Hybrid Classifier for Screening of Diabetic Retinopathy

Abstract

Medical image analysis is very popular research area these days in which digital images are analyzed for the diagnosis and screening of different medical problems. Diabetic retinopathy (DR) is an eye disease caused by the increase of insulin in blood and may cause blindness if not treated in time. Healthy retina contains blood vessels, optic disc and macula as main components but abnormal retina may contain other components and signs as well. An au- tomated system for early detection of DR can save patient’s vision and can also help the ophthalmologists in screening of DR. In this thesis, we develop algorithms for retinal image analysis based on image processing and pattern classification. Image processing techniques are used for retinal image enhancement and pattern recognition is used for classification of DR stages. The proposed system consists of different stages such as preprocessing, compo- nent extraction, candidate region detection, feature extraction and finally the classification. The first phase consists of input retinal image enhancement, noise removal, extraction of main retinal components and candidate lesions detection. We apply Gabor wavelets and Gabor filter banks for lesion detection. The system then extracts features from candidate lesions using four main properties, i.e. shape, color, gray level and statistical. Finally the classifier takes the feature vectors as inputs and grades the input retinal image into dif- ferent stages of DR. We present a hybrid classifier which combines the Gaussian Mixture Model (GMM), Support Vector Machine (SVM) and an extension of multimodel mediod based modeling approach in an ensemble to improve the accuracy of classification. The im- plemented algorithms are tested and evaluated on publicly available retinal image databases using performance parameters such as sensitivity, specificity, positive predictive value and accuracy. The performance improvement of our proposed system is demonstrated by com- paring them with recently proposed and published methods.