Introduction: OCT is essential for imaging retinal microstructures, with precise segmentation of retinal layers being crucial for disease diagnosis and monitoring. Despite the multitude of proposed strategies, discrepancies in image quality, device kinds, and annotation standards result in variable performance among studies. Methods: The study systematically analyzes traditional image-processing techniques, machinelearning techniques, and deep-learning architectures employed for OCT retinal layer segmentation. A comparative study of publicly available datasets, assessment metrics, and methodological aspects was conducted using information derived from the current literature and presented in a dedicated table. Results: Studies indicate that deep-learning models, particularly U-Net variants and multi-scale encoder-decoder architectures, often outperform classical and machine learning-based methods in terms of segmentation accuracy. Most methods demonstrate diminished generalization when utilized on datasets with varying imaging characteristics. Discussion: This review highlights several significant limitations in existing research, including inconsistent evaluation metrics, insufficient clinical validation, and susceptibility to noise and artifacts. This study presents a comparative analysis of methodologies, highlighting the necessity for standardized datasets and clinically relevant metrics (such as boundary error and thickness deviation). This review is distinguished from previous studies by these critical insights. Conclusion: This study consolidates classical, machine learning, and deep learning segmentation methodologies, presenting a systematic comparison of datasets, metrics, and problems, thereby serving as a thorough and analytically robust resource for advancing OCT segmentation work.
Yadav et al. (Mon,) studied this question.