An Explainable AI-Based Computer Aided Detection System for Diabetic Retinopathy Using Retinal Fundus Images

Abstract

Diabetic patients have a high risk of developing diabetic retinopathy (DR), which is one of the major causes of blindness. With early detection and the right treatment patients may be spared from losing their vision. We propose a computer-aided detection system, which uses retinal fundus images as input and it detects all types of lesions that define diabetic retinopathy. The aim of our system is to assist eye specialists by automatically detecting the healthy retinas and referring the images of the unhealthy ones. For the latter cases, the system offers an interactive tool where the doctor can examine the local lesions that our system marks as suspicious. The final decision remains in the hands of the ophthalmologists. Our approach consists of a multi-class detector, that is able to locate and recognize all candidate DR-defining lesions. If the system detects at least one lesion, then the image is marked as unhealthy. The lesion detector is built on the faster R-CNN ResNet 101 architecture, which we train by transfer learning. We evaluate our approach on three benchmark data sets, namely Messidor-2, IDRiD, and E-Ophtha by measuring the sensitivity (SE) and specificity (SP) based on the binary classification of healthy and unhealthy images. The results that we obtain for Messidor-2 and IDRiD are (SE: 0.965, SP: 0.843), and (SE: 0.83, SP: 0.94), respectively. For the E-Ophtha data set we follow the literature and perform two experiments, one where we detect only lesions of the type micro aneurysms (SE: 0.939, SP: 0.82) and the other when we detect only exudates (SE: 0.851, SP: 0.971). Besides the high effectiveness that we achieve, the other important contribution of our work is the interactive tool, which we offer to the medical experts, highlighting all suspicious lesions detected by the proposed system.