Deep Learning Approaches for Diabetic Retinopathy: A Review with Empirical Insights

Diabetic Retinopathy is a hazardous eye state that affects the diabetic patients. Without early detection, it may impact the retina and potentially result in permanent loss of vision. Although, it is very difficult to diagnose DR and it requires a well experienced ophthalmologist. An advancement in AI known as Deep learning can help ophthalmologist with an option with respect to the classification of the DR by utilizing an independent classifier. The paper concentrates on the exhibition of various exceptionally proficient and versatile CAD frameworks designed especially for DR Classification using AI, ML and DL models and to compare their performances. To evaluate the performance, various performance-based metrics such as, True & False Positive Rates, Sensitivity, Specificity, Accuracy, and so forth are used. In this study we analyzed 12 deep learning (DL) and 8 machine learning (ML) models across Kaggle, MESSIDOR-2, and e-Ophtha datasets. The primary goal of this review is to compare different CAD models implemented for DR detection in early stages. To overcome the limitations of those CAD models, Transfer learning and Federated learning can be used. The future scope can be to develop a modified TL model to further develop classification precision.