A foundation model for generalizable disease detection from retinal images

AI Generated Summary*

RETFound, a publicly released self-supervised retinal foundation model, was pretrained with a masked autoencoder on 1.6 million unlabelled images (904,170 colour fundus photographs and 736,442 OCT scans), mostly from 37,401 Moorfields diabetic patients, then fine-tuned for classification tasks. Against ImageNet-pretrained and retinal-only baselines it achieved higher AUROC in most tasks, including 0.943 for diabetic retinopathy on APTOS-2019. Using fundus photographs, AUROC was 0.862 for one-year fellow-eye conversion to wet AMD and 0.794, 0.737, 0.754 and 0.669 for three-year incidence of heart failure, myocardial infarction, ischaemic stroke and Parkinson's disease in the AlzEye cohort. With 10% of the labelled data, heart failure prediction still exceeded comparators. Performance dropped on external UK Biobank validation, where it led most tasks.

*This summary was generated by AI and is published unedited. Oku does not alter these summaries. It may contain errors or omissions and is provided for general informational purposes only. Oku does not guarantee its accuracy, completeness, or reliability. For authoritative information, please refer to the original, peer-reviewed article.