An ensemble approach to detect exudates in digital fundus images

Abstract

Fundus Image analysis is a major concern with respect to various disease detection. Diabetic retinopathy (DR) is seen in patients suffering from diabetes mellitus type 2 which leads to blindness. Fundus images are used to identify abnormalities like microaneurysms, haemorrhages, cotton wool spots, exudates, venous beading, and optic disc oedema that cause DR. Automated diagnosis of DR gives first-hand information about the disease presence, and save diabetic patients from vision loss. This paper presents a novel ensemble approach to automatically detect exudates in the fundus images. Normal background features are removed initially. Morphological operations combined with logical operations is the ensemble approach that has enhanced the detection and marking of exudates. Publicly available standard database DIARETDB1 and images of Forus Health is used to experiment the algorithm. 89.6% of specificity, 100% of sensitivity is obtained and evaluated with logistic regression classifier. Also, 89.13% of positive predictive value and 100% negative predictive value is obtained with this approach. The AUC of ROC plot obtained is 0.969.