Classification of Abnormal Blood Vessels in Diabetic Retinopathy Using Neural Network

Abstract

Neovascularization is a serious sight-threatening complication of proliferative diabetic retinopathy (PDR) occurring in the diabetes mellitus persons, which causes progressive damage to the retina through the growth of new abnormal blood vessels. Preprocessing technique primarily extracts and normalizes the green plane of fundus image used to increase the level of contrast, the change in contrast level has been analyzed using Pair-wise Euclidean distance method. Normalized green plane image is subjected into the two-stage approach: detecting neovascularization region using compactness classifier and classifying neovascularization vessels using neural network. Compactness classifier with morphology-based operator and thresholding techniques are used to detect the neovascularization region. A function matrix box is added to categorize the neovascularization from normal blood vessels. Then, the feed-forward back-propagation neural network for the extracted features like number of segments, gray level, gradient, gradient variation, gray-level coefficient is attempted in Neovascularization region to get an indicative accuracy of classification. The proposed method is tested on images from online datasets and from two hospital eye clinic real-time images with varying quality and image resolution, achieves sensitivity and specificity of 80 and 100% respectively and with an accuracy of 90% gives encouraging abnormal blood vessels classification.