Foundation models could reshape AI in eye care

By AI, Created 12:26 PM UTC, June 03, 2026, /AGP/ – A new systematic review published Oct. 24, 2025, says vision and vision-language foundation models are showing strong performance across retinal disease, glaucoma, ocular surface tumors and rare eye conditions. The findings suggest ophthalmology may be moving beyond single-task AI tools toward more flexible systems, but the review says clinical validation, fairness and interpretability still need work before broad adoption.

Why it matters: - Foundation models could make ophthalmic AI more adaptable across diseases, devices and patient populations. - The review says these systems may improve diagnosis, screening and clinical decision support by learning reusable patterns from large image and text datasets. - If validated in real-world settings, the models could support earlier diagnosis, better referral decisions and broader access to specialist-level eye care.

What happened: - Researchers from Zhejiang University Eye Hospital and collaborators across China, the United States, the United Kingdom, Australia, Singapore, Poland and Hong Kong published a systematic review in Advances in Ophthalmology Practice and Research on Oct. 24, 2025. - The review examined vision and vision-language foundation models in ophthalmology, with a focus on diagnostic performance, interpretability, fairness, deployment barriers and future directions. - The paper assessed ten studies published between January 2020 and July 2025.

The details: - The review covered representative models including RETFound, FLAIR, VisionFM, EyeCLIP, FMUE, MetaGP, MINIM, RETFound-DE, RetiZero and OSPM. - Retinal disease was the most common use case, especially diabetic retinopathy, age-related macular degeneration and diabetic macular edema. - RETFound reached an AUC of 0.94 for diabetic retinopathy detection on EyePACS. - VisionFM reached an AUC of 0.974 for age-related macular degeneration in external validation. - RETFound-DE reached an AUC of 0.902 on REFUGE-2 for glaucoma. - EyeCLIP showed promising performance across several external datasets. - OSPM produced AUC values of about 0.986 to 0.993 for ocular surface tumors. - RetiZero recognized more than 400 rare fundus diseases with top-five accuracy of 75.6%. - Several models demonstrated few-shot and zero-shot learning, which means they could adapt to new diagnostic tasks with limited labeled data. - The authors said foundation models may help clinicians extract more value from routine eye-care data. - The authors also said strong results on research datasets are only a starting point and that clinical use still requires transparency, careful validation and support for clinical judgment.

Between the lines: - Traditional ophthalmic AI often targets one disease, one dataset or one task, which can limit performance across hospitals, imaging devices and patient groups. - Foundation models are positioned as a more flexible alternative because they can combine images, clinical language and patient records. - The review suggests the main barrier is no longer whether these models can perform well in studies, but whether they can be trusted, explained and deployed safely in practice. - The paper flags limited data diversity, algorithmic bias, overfitting, high computational demands, weak interpretability, EHR interoperability problems and insufficient clinical validation as major obstacles.

What’s next: - Future work should focus on larger and more representative datasets. - The authors call for explainable AI tools such as saliency maps, SHAP and counterfactual reasoning. - Post-deployment monitoring for fairness and performance drift will be important if the models move into routine care. - The review argues that the next step is not just higher accuracy, but safer and more scalable AI-assisted ophthalmic workflows.

The bottom line: - Ophthalmic foundation models are showing broad promise, but clinical adoption will depend on validation, transparency and fairness, not just benchmark scores. - The review suggests the field is shifting from narrow diagnostic tools to multimodal systems designed for real-world eye care.