Explainable Deep Learning for Referable Diabetic Retinopathy

Abstract

Diabetic Retinopathy (DR) is a retinal lesion due to diabetes. Through blood leaks and excess glucose in the blood vessels, pathological lesions including hemorrhages, exudates and microaneurysms (HM, EX, MA) develop in the eye, which may lead to blindness if not timely treated. In this paper, we propose a deep Convolutional Neural Network (CNN) architecture trained to identify Referable Diabetic Retinopathy (RDR) lesions from retinal fundus images. The model uses a pre-trained network with fine-tuned layers, cosine learning rate decay, and warm up. The efficiency of the proposed architecture has been evaluated on eight public datasets. The results show that the proposed architecture obtains state-of-the-art performance using only publicly available datasets. An explainability algorithm was also developed to show the efficiency of the model in detecting RDR signs.