Beyond convolutions and supervised learning with transformers and representation learning for retinal image analysis

Retinal image analysis has enjoyed groundbreaking advances in the last ten years due to seismic improvements in image analysis techniques from the field of computer vision. Previous reviews in deep learning and artificial intelligence (AI) (Schmidt-Erfurth et al., 2018; Ting et al., 2019) have either focused on supervised learning, where labels are curated or manually created, or concentrated on the application of AI in specific image modalities and retina diseases (Hormel et al., 2021; Li et al., 2024a (Hormel et al., 2021; Li et al., 2024a)). In this review, we sought to summarize the advances in the field with the shift towards label-free approaches using representational learning and the emergence of vision transformers as alternatives to convolutional neural networks for image analysis. These advances include semi-supervised learning, self-supervised learning and directly led to the advent of foundation models, vision-language models, and multi-modal models.