Ophthalmology foundation models for clinically significant age macular degeneration detection

Area of Science:

• Ophthalmology
• Computer Vision
• Artificial Intelligence

Background:

• Self-supervised learning (SSL) enables Vision Transformers (ViTs) to learn robust image representations.
• Foundation models in retinal imaging show promise, but the necessity of in-domain pretraining is unclear.
• Age-related macular degeneration (AMD) identification is crucial in digital fundus imaging (DFI).

Purpose of the Study:

• To benchmark SSL-pretrained ViTs for AMD identification in DFIs.
• To evaluate the generalization capabilities of models pretrained on natural versus domain-specific data.
• To investigate the necessity of in-domain pretraining for AMD detection.

Main Methods:

• Benchmarking six SSL-pretrained ViTs on seven DFI datasets (70,000 images).
• Evaluating model performance on moderate-to-late stage AMD identification.
• Comparing out-of-distribution generalization of natural vs. domain-specific pretraining.

Main Results:

• iBOT, pretrained on natural images, achieved superior out-of-distribution generalization (AUROCs 0.80-0.97).
• Natural image pretraining outperformed domain-specific models (AUROCs 0.78-0.96) and a non-pretrained baseline (AUROC 0.68-0.91).
• Foundation models significantly improve AMD identification accuracy.

Conclusions:

• Natural image pretraining for ViTs is highly effective for AMD identification, challenging the need for in-domain pretraining.
• Foundation models offer significant value in enhancing AMD detection in DFI.
• The study introduces BRAMD, a new open-access dataset for AMD research.