Diabetic Retinopathy Using Deep Learning

Abstract

According to a recent survey by WHO, every fourth person is diabetic. It can be due to lifestyle, pollution, heredity or many other reasons. A diabetic person with the time has some threats in life. One of these threats is Retinopathy which is a well-known microvascular consequence of diabetes mellitus that poses a risk to vision is diabetic retinopathy. Worldwide, 93 million people have diabetic retinopathy at this time. Studies on the prevalence of diabetic retinopathy and risk factors are lacking. People who have diabetes may have more chances to develop diabetic retinopathy, an eye disorder that can lead to blindness. The retina's blood vessels are impacted (the light-sensitive layer of tissue in the back of your eye). It's crucial to undergo a thorough dilated eye exam at least once a year if you have diabetes. In this paper, we have used the deep learning method by taking a single snapshot of the human fundus for detecting the stage of diabetic retinopathy. We have also suggested the multistage transfer learning method, which uses comparable datasets with various labelling. With a sensitivity and specificity of 0.99, the proposed method can be used as early diagnosis of diabetic retinopathy on APTOS 2019 dataset for Detecting Blindness (13000 images). We employed a deep learning model Shapley Additive exPlanations (SHAP) for game theoretic strategies to research the diagnosis of diabetic retinopathy. In order to associate optimal credit allocation with local explanations, it makes use of the conventional Shapley values from game theory and their related extensions. To understand the predictions of our model and how to further improve its performance, in this study we will train a base ResNet50 model, evaluate it, and use the SHAP model's explain ability technique.

Topics & Concepts