From Fundus Images to Clinical Decisions: A comprehensive Review on Robust Multi-Retinal Disease Classification

Some of the common causes of preventable blindness across the world include retinal diseases, such as Diabetic Retinopathy (DR), Glaucoma, Age-Related Macular Degeneration (AMD), hypertensive retinopathy, and myopia. There is no question of the vitality of early diagnosis, but clinical screening is limited by the insufficient number of experts and inter-observer differences, especially in low-resource conditions. Recent developments in deep learning have shown a robust opportunity in the field of spontaneous retina image analysis, but single convolutional neural networks (CNN) models tend to be limited in generalization and unreliable in practice in multi-disease conditions. The current paper provides a detailed survey of the deep learning strategies of the robust multi-retinal disease classification based on fundus and OCT images. This review critically reviews different strategies, including soft voting, stacking, and bagging; compares the performance of the strategies among the different CNN architectures; discusses the use of explainable artificial intelligence (XAI) in improving clinical trust; and speaks about the possibility of implementing lightweight diagnostic systems in mobile and resource-constrained environments. This review can be used to fill the existing gap between the algorithmic developments and clinical decision-making, as it synthesizes the existing literature and practical considerations related to system design.