Independent Evaluation of RETFound Foundation Model's Performance on Optic Nerve Analysis Using Fundus Photography

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Accurate prediction of optic nerve metrics is crucial for diagnosing and monitoring eye diseases.
• Foundation models offer potential for extracting meaningful features from medical images.

Purpose of the Study:

• To evaluate the RETFound foundation model as a feature extractor for predicting optic nerve metrics (CDR and RNFL thickness).
• To assess the performance of RETFound models against baseline models using independent clinical data.

Main Methods:

• Retrospective observational study using fundus images and RNFL OCT data.
• Extracted latent features from fundus images using RETFound.
• Trained linear regression models (Ridge, Lasso, Elastic Net, OLS) with RETFound features.
• Compared RETFound models with VGG16 and Vision Transformer (ViT) baseline models.

Main Results:

• RETFound models achieved high accuracy for single-output tasks (CDR: R² 0.706-0.898; Avg RNFL: R² 0.855-0.961).
• Performance was less robust for multi-output tasks (RNFL clock-hour: R² 0.583; RNFL quadrant: R² 0.811).
• RETFound models outperformed VGG16 and ViT models in predicting RNFL thickness and CDR.

Conclusions:

• Machine learning models using the RETFound foundation model can accurately predict CDR and average RNFL thickness from fundus photos.
• RETFound effectively overcomes small dataset limitations for specialized ophthalmic applications.
• The study highlights the potential of foundation models in ophthalmology for optic nerve evaluation.