Foundation models for ophthalmic imaging

Foundation models represent a new frontier in ophthalmic artificial intelligence, enabling learning of transferable features from large unlabelled imaging datasets for flexible application to varying downstream tasks. We systematically analyze the evolution of ophthalmic foundation models across 12 distinct models developed between 2022 and July, 2025. We examine advances in modality integration (unimodal, multimodal to vision-language), pretraining objectives (generative versus contrastive approaches) and supervision strategies (image and text guided). Emerging techniques such as imaging modality agnostic encoders, synthetic data augmentation, and computationally efficient architectures improved model performance and generalisability. Overall, we observed a clear shift from domain-specific unimodal approaches towards modality-agnostic foundation models guided by clinical text. Future directions include wider modality integration, higher dimensional inputs (spatially and temporally), diverse pretraining, and standardised benchmark datasets. In synthesizing these trends, this review offers the conceptual and technical grounding to support both clinicians and researchers in ophthalmic foundation model design, selection, and application.