Foundation models for ophthalmic imaging

Area of Science:

• Ophthalmic Artificial Intelligence
• Machine Learning in Healthcare
• Medical Imaging Analysis

Background:

• Foundation models offer transferable feature learning from large, unlabeled ophthalmic datasets.
• These models are crucial for advancing artificial intelligence in ophthalmology.
• They enable flexible application across various downstream diagnostic and analytical tasks.

Purpose of the Study:

• To systematically analyze the evolution of ophthalmic foundation models from 2022 to July 2025.
• To examine key advancements in model architecture, pre-training, and supervision strategies.
• To provide a comprehensive overview of current trends and future directions in the field.

Main Methods:

• Systematic review of 12 distinct ophthalmic foundation models.
• Analysis of modality integration (unimodal, multimodal, vision-language).
• Evaluation of pretraining objectives (generative vs. contrastive) and supervision (image/text-guided).

Main Results:

• Observed a shift from domain-specific unimodal models to modality-agnostic models guided by clinical text.
• Emerging techniques like modality-agnostic encoders and synthetic data augmentation enhance performance and generalizability.
• Advances in pretraining and supervision strategies are key drivers of progress.

Conclusions:

• Ophthalmic foundation models are increasingly sophisticated, moving towards broader applicability.
• Future research should focus on wider modality integration, higher-dimensional data, and standardized benchmarks.
• This review provides foundational knowledge for developing and applying these AI models in ophthalmology.