The Role of Deep Learning Approaches in the Classification of Diabetic Retinopathy

Over the past couple of decades, there has been a tremendous rise in the global population suffering from mellitus. Intergenerational individuals are being affected. DR is a disorder that causes damage to the optic nerve and occurs in people who have had diabetes for a long period. Such complications as blindness can be avoided if there is a system of detection using innovative technology for early detection. While machine learning strategies in general, and Deep Learning, in particular, have become more sophisticated, DL-based algorithms are the most common choice for the development of DR identification mechanisms. For this purpose, the paper aimed to review the body of research on deep learning-powered diabetic retinopathy evaluation using fundus images and provide a review of the most novel methods that have been so far utilized by experts in the domain. Some of the commonly used datasets are then enumerated in the present study. The performance evaluation of those evaluated approaches is then reported with respect to a couple of metrics that are commonly used for machine vision applications. The study results show that the enhanced diabetic retinopathy detection method reaches <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{9 8. 6 5 \%}$</tex> accuracy with RetNet-10 and 96.85 % accuracy with hybrid CNN models built from InceptionV3 and ResNet50. The APTOS 2019 and Messidor datasets prove the most remarkable since they delivered a 95% accuracy. The hybrid models built with EfficientNet and Swin Transformer demonstrate 97 % accuracy but the 2D-SWT method with feature extraction and CPSO-kNN and RNN-LSTM achieves <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{9 7. 7 0 \%}$</tex> accuracy.