How Foundational Is the Retina Foundation Model? Estimating RETFound's Label Efficiency on Binary Classification of...

Area of Science:

• Ophthalmology
• Medical Imaging
• Machine Learning

Background:

• Machine learning progress is often limited by the scarcity of labeled medical data due to privacy and cost constraints.
• Self-supervised learning, particularly pretraining on large unlabeled datasets, shows promise for improving model performance on downstream tasks.
• The RETFound model, a large vision transformer pretrained on 1.6 million retinal images, represents a significant advancement in ophthalmology AI.

Purpose of the Study:

• To evaluate the label efficiency of the RETFound model for identifying normal versus abnormal Optical Coherence Tomography (OCT) B-scans.
• To compare the performance of RETFound against established models (ResNet-50, ViT-L) in a primary care diabetic retinopathy screening context.

Main Methods:

• 1150 OCT B-scans from 647 patients were randomly allocated into training, validation, and testing sets.
• Three models (ResNet-50, ViT-L, RETFound) were fine-tuned on varying dataset sizes (915 to 50 OCT B-scans).
• Model performance was assessed using accuracy, AUROC, AUPRC, F1 score, precision, and recall.

Main Results:

• RETFound consistently outperformed ResNet-50 and ViT-L across all metrics and dataset sizes.
• ResNet-50 and ViT-L showed comparable performance on larger datasets but degraded significantly with smaller subsets.
• The study highlights RETFound's superior ability to learn from limited labeled data, emphasizing the benefit of retina-specific pretraining.

Conclusions:

• The findings validate the effectiveness of the RETFound model's specialized pretraining for ophthalmology tasks.
• RETFound demonstrates strong label efficiency, making it suitable for applications with limited labeled OCT B-scan data.
• Further research is recommended to optimize fine-tuning strategies for RETFound in various downstream applications.