Applications of Fairness Training in Artificial Intelligence to Ophthalmic Disease Diagnosis using External Eye Photographs

LLaVA on 40,687 external eye photographs from 9,878 patients at Massachusetts Eye and Ear achieved 85.4% AUC for vision-threatening diabetic retinopathy, with fairness gaps partially mitigated after training.

• Why it matters: It remains unclear whether external eye photos contain usable phenotypic signals for posterior segment diseases, and demographic disparity mitigation for such models is underexplored.
• What they did: Deep learning and vision-language models (WResNet, EfficientNet, ViT, and fine-tuned LLaVA) were evaluated for VTDR, advanced AMD, and glaucoma using AUC and equity-scaled AUC (ES-AUC), then five fairness-training strategies were applied to the best models.
• The result: Fairness training raised VTDR and advanced AMD AUCs to 86.3% and 85.9% and partially reduced sex, race, and ethnicity disparities, while glaucoma remained substantially worse at 67–74% AUC.

AI summary of the abstract · may contain errors · not medical advice · read the paper.