Detection and Prognosis Evaluation of Diabetic Retinopathy using Ensemble Deep Convolutional Neural Networks

Abstract

Diabetic Retinopathy is a condition that occurs in the eye as a result of diabetes in patients. Due to uncontrolled blood sugar levels in patients, there would be a lack of blood flow and oxygen to the retina. This causes strain on blood vessels some extent without invasive treatment and when detected in its early stages. When the strain in the blood vessels increases, it may cause leakage of fluids from blood vessels and loss of proper vision in the eye. This system implements a deep learning model using ResNet to determine the performance for the detection of the various stages of the condition in individuals. Individual submodels are built using ResNet to detect the presence of Diabetic Retinopathy and are ensembled together using the AdaBoost Classifier. Multiclass classification ResNet models are built and stacked together to detect the prognosis of Diabetic Retinopathy. The implemented models showed a performance accuracy of 78.88% to detect the presence and 61.9% to evaluate the prognosis of Diabetic Retinopathy. The performance of the trained models is visualised with a Grad-CAM and the results are analysed.