Deep Learning Approaches for Multi-Class Retinal Disease Classification: A Comprehensive Survey

Abstract

Retinal conditions like Diabetic Retinopathy (DR), glaucoma, and cataract are global leading causes of visual disability. Manual screening techniques are time-consuming and prone to inter-observer variation, rendering them unsuitable for large-scale diagnosis. Automated retinal image analysis has proven to be highly promising with the development of deep learning to enhance diagnostic accuracy and accessibility. The following paper discusses a thorough review of recent deep learning-based techniques for fundus image classification, including CNN, transformer, and hybrid texture-aware models. Although various models are proven to work well, issues like poor generalizability, susceptibility to artifacts and noise, class imbalance, and lack of interpretability remain. Overcoming these shortcomings, recent developments point towards the importance of lightweight and interpretable models combining deep features and handcrafted texture descriptors. Such models become stronger and more efficient and are fit for deployment in real-world settings, especially low-resource clinical settings.