Advanced Machine Learning Techniques for Retinal Lesion Segmentation: A Comprehensive Review Across Ophthalmic Pathologies

Automated segmentation of retinal lesions using machine learning methods represents a cru- cial tool for diagnosing and monitoring ophthalmic diseases such as diabetic retinopathy (DR), retinopathy of prematurity (ROP), pathological myopia, and age-related macular degeneration (AMD). This review systematically evaluates the current state of machine learning- and deep learning-based segmentation methods for pathological retinal lesions in fundus imaging. A total of 86 publications meeting PRISMA criteria were analyzed, covering various architectures of convolutional neural networks (CNN), transformers, hybrid models, and generative adversarial networks (GAN). Variants of the U-Net architecture were found to be the most frequently used models, with hybrid approaches integrating CNNs and transformers demonstrating increasing potential. While lesion segmentation in DR dominates the available literature, the area of ROP remains underexplored, largely due to the limited availability of annotated datasets. The review identifies key challenges such as model generalization across different imaging platforms and patient populations, emphasizing the need for further research aimed at dynamic lesion quantification, especially in clinically underserved domains like ROP. The findings indicate that advanced deep learning-based models achieve high segmentation accuracy, offering substantial potential for enhancing diagnostic and therapeutic procedures in ophthalmology.