Deep Learning based Optimized Diabetic Retinopathy Detection Platform from Fundus Photographs

Abstract

Diabetic Retinopathy (DR) is primarily an eye disease caused on by persistently high levels of blood glucose. If left untreated, DR can result in lifelong blindness. The fundus images are typically examined with the aid of highly skilled professionals to diagnose these devastating disorders. The manual diagnosis of these disorders, which is often carried out by clinicians, is exceedingly time-consuming, expensive and prone to error. Machine learning can evaluate medical images and has demonstrated its proficiency in evaluating the retinal fundus images. The implementation of deep learning techniques through algorithms has greatly assisted the earlier detection of diabetic retinopathy (DR). In this study, we used the EyePACS dataset, which contains high-resolution annotated images in five distinct classes. The images are produced under a range of imaging conditions and in a variety of dimensions. To detect the stage of diabetic retinopathy, we have experimented with a wide variety of deep learning models. And, we have achieved 89% testing F1score in our own model RevampResNet18. We have also developed a web platform to detect the stage of DR from fundus images. Previously, it took weeks to clinically diagnose the stage of DR. But, with our platform, users can find the stage of DR within seconds. People can access the website from anywhere in the world using an active internet connection in the PC or smartphone and a web browser.