Automatic Detection and Classification of Diabetic Lesions for Grading Diabetic Retinopathy Using Fuzzy Rule-Based Classification System

Abstract

Diabetic Retinopathy (DR) is a chronic, progressive retinal disease which is the most common cause of legal blindness. The disease is threatening to eyes as it shows no signs of visual abnormality at the initial stage. It gradually decreases patient’s eye-sight and drives into blindness in future. Hence, the early detection of DR is vital to prevent the complete vision loss of diabetes patients. Traditional diagnosing system of DR requires quite trained ophthalmologists for monitoring the retina periodically. Moreover, several physical tests like fluorescein angiography, visual acuity test, and ocular coherence tomography are involved to detect DR which also require a lot of time to process. In this paper, a fuzzy rule-based classification technique is proposed for automatic detection and classification of retinal lesions for grading DR. The proposed technique consists of preprocessing of fundus images, extraction of candidate retinal lesions, formulation of feature set, and classification of DR. In the preprocessing phase, the technique eliminates background noises and extracts optic disc from the retinal fundus image. Four leading lesions; blood vessels, microaneurysms, haemorrhages, and exudates are extracted using different image processing techniques and two textural features; contrast and homogeneity are calculated in the detection phase. Then, these six input features; blood vessels area, microaneurysms count, haemorrhages area, exudates area, contrast and homogeneity are fed to fuzzy if-then rule-based classifier for predicting normal, mild NPDR, moderate NPDR, severe NPDR and PDR stages of DR.. A total of 520 retinal fundus images are collected from four public database; STARE, DIARETDB0, DIARETDB1 and MESSIDOR and the images are successfully classified by the fuzzy rule-based classifier with accuracy up to 92.42%. The sensitivity and specificity of the classifier are 92.44% and 94.29% respectively. The simulation result on publicly standard image datasets exhibits that the intended technique gives promising results in identifying retinal lesions and it has better capability of classifying several stages of DR compared with other existing automatic diagnosing system.