Red Lesion Detection in Color Fundus Images for Diabetic Retinopathy Detection

Abstract

Diabetic retinopathy (DR) is a chronic disease in the eye due to blood leakages which causes vision impairment and can be identified on the surface of the retina. Diabetic patients are the most common subject to this disease, and ignorance to it can result in permanent visual damage and eventual blindness. DR falls in two categories: proliferative diabetic retinopathy (PDR) and non-proliferative diabetic retinopathy (NPDR). Non-proliferative DR is the most common form of DR, whereas proliferative DR is the severe stage of DR which causes the blood vessels to close off. DR can be detected in its early stages by the red lesions, i.e., microaneurysms and hemorrhages. In this paper, we implement a method of detecting red lesions for detection of NPDR form. The proposed methodology uses fundus images as its input and employs modified approach to extraction of retinal blood vessels and median filtering. To train the model, multiclass support vector machine classifier is implemented using the extracted features. The method is tested on 1928 fundus images from ‘KAGGLE APTOS’ and 103 images from ‘IDRiD’ dataset. The performance of the proposed methodology is as follows: sensitivity 75.6% and accuracy 94.5% on ‘KAGGLE APTOS’ and 78.5% sensitivity and 93.3% accuracy on ‘IDRiD’ dataset.