A Comprehensive Review of Deep Learning-Based Retinal Layer Integrity Assessment in Optical Coherence Tomography (OCT)

Abstract

This comprehensive review investigates recent advances in deep learning-based methods for assessing the integrity of retinal layers - specifically the External Limiting Membrane (ELM) and Ellipsoid Zone (EZ) - in Optical Coherence Tomography (OCT) imaging. A structured search was conducted on PubMed, IEEE Xplore, ScienceDirect and Google Scholar to identify peer-reviewed studies published between 2015 and 2025 that applied deep learning for ELM/EZ segmentation or integrity assessment with quantifiable outcomes such as dice, intersection over union (IoU) or volume metrics. Forty-three eligible studies were synthesized, revealing the dominance of U-Net-based models, attention mechanisms, and emerging transformer architectures. The review highlights advances in model performance, dataset diversity, and clinical applicability while identifying persistent challenges in generalization, smoothness quantification, and real-time deployment. To address these gaps, we recommend standardized annotation protocols, publicly available benchmarks, multicenter and multidevice validation, and prospective studies linking automated ELM/EZ metrics with clinical outcomes. Clearer reporting of statistical uncertainty, improved model explainability, and optimized lightweight architectures for clinical use will be essential to translate automated ELM / EZ analysis into reliable and actionable tools for diagnosis, prognosis, and treatment monitoring.