A comprehensive review of deep learning in OCT image segmentation and classification

Abstract Optical Coherence Tomography (OCT) has transformed ophthalmology by enabling high-resolution imaging essential for diagnosing a wide range of ocular diseases. In recent years, Deep Learning (DL) techniques have really integrated well with OCT, bringing a significant boost in the accuracy and efficiency of automatic disease detection. This review is meant as a thorough and full-scale exploration of DL-based OCT segmentation and classification. It discusses the progress and points out key issues which need to be addressed in future research. The paper first looks at OCT datasets, their diversity, representational diseases and capacity to form reliable training sets for DL models. Then it delves into analysis with segmentation methods, compares their performance and identifies problems with existing approaches. The review surveys current classification techniques, contrasting deep learning models of various architectures which are capable of recognizing and distinguishing retinal diseases. It also focuses on the clinical significance of these models–details precisely what ocular conditions they analyze, and how well they can diagnose disease. In addition to synthesizing existing achievements, the review makes clear the major highpoints of current research as well as future directions. It identifies problems such as inadequate dataset diversity, model generality irregulatity, interpretability and computation efficiency. It makes concrete proposals for improvement, including real-world image collection and algorithm optimizations to fill in gaps in databases or increase model episode performance remarkably. This kind-of study should eventually help to provide a clear view of the present state as well as future prospects for deep learning in ophthalmic diagnosis by OCT. • Deep learning significantly improves OCT-based segmentation and classification. • Most OCT datasets are limited in diversity, affecting model generalization. • Challenges include model interpretability, dataset bias, and computational efficiency. • Future research should focus on dataset expansion and model optimization. • This review identifies research gaps and proposes solutions for clinical adoption.