Evaluating a Foundation Artificial Intelligence Model for Glaucoma Detection Using Color Fundus Photographs.

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Diagnostics

Background:

• Glaucoma is a leading cause of irreversible blindness worldwide.
• Early detection and accurate diagnosis are crucial for effective glaucoma management.
• Automated diagnostic tools can improve screening efficiency and accessibility.

Purpose of the Study:

• To evaluate the diagnostic accuracy of the RETFound artificial intelligence (AI) model for glaucoma detection using optic disc photographs.
• To assess the model's performance across diverse demographic groups (race, age) and disease severities.
• To determine the impact of training dataset size and training cycles on the AI model's accuracy.

Main Methods:

• The study utilized a diverse dataset of 9787 color fundus photographs (CFPs) from 2329 participants.
• RETFound, a self-supervised learning model, was employed for binary glaucoma classification.
• Model performance was evaluated by varying dataset sample sizes (50-2000 images) and training epochs (5-50), stratified by race, age, and glaucoma severity.

Main Results:

• The area under the receiver operating characteristic curve (AUC) for glaucoma detection improved from 0.52 (50 images, 5 epochs) to 0.86 (2000 images, 50 epochs).
• Performance remained consistent across different races (Black AUC=0.87 vs. other AUC=0.86) and age groups (<60 years AUC=0.87 vs. >60 years AUC=0.84).
• The model demonstrated higher accuracy in detecting moderate to severe glaucoma (AUC=0.95) compared to mild glaucoma (AUC=0.84).

Conclusions:

• RETFound demonstrated good performance in detecting glaucoma, even with relatively small fine-tuning datasets.
• The AI model's accuracy was consistent across various races and age groups.
• RETFound shows promise as a tool for automated glaucoma detection in diverse clinical settings.