Building on AI Foundations
Systematic review examines how foundation models could potentially reshape clinical decision-making and ophthalmic diagnostics
Objective:
To evaluate the potential of vision and vision-language foundation models in ophthalmology and identify challenges to their adoption.
Approach:
- Systematic Review: Analyzed 10 studies published between 2023 and 2025 on AI foundation models in retinal disease, glaucoma, and ocular oncology.
Key Findings:
- Foundation models can adapt across multiple diagnostic tasks with limited labeled data.
- Ophthalmology is well-suited for AI due to its reliance on image-based diagnostics.
- Models demonstrated performance nearing or exceeding that of experienced clinicians.
- RETFound achieved an AUC of 0.94 for diabetic retinopathy detection.
- VisionFM reached AUC values of 0.974 for age-related macular degeneration and 0.945 for diabetic retinopathy.
- Models targeting glaucoma detection reported AUC values ranging from 0.721 to 0.913.
- The ocular surface tumor model OSPM achieved AUC scores as high as 0.993.
- Models showed strong results in few-shot and zero-shot scenarios, beneficial for rare conditions.
Interpretation:
Significant challenges remain regarding bias, interpretability, and integration into clinical practice.
Limitations:
- Most studies relied on retrospective datasets and convenience sampling.
- Concerns about generalizability across different populations and imaging devices.
- Algorithmic bias and computational demands pose challenges.
- Fragmented EHR interoperability affects model deployment.
- The 'black-box' nature of deep learning complicates understanding model predictions.
Conclusion:
Foundation models may enable scalable, multimodal AI systems for diagnosis and decision-making.
Sources:
