A foundation model for generalizable disease detection from retinal images

• Published in: Nature (London) Vol. 622; no. 7981; pp. 156 - 163
• Main Authors: Zhou, Yukun, Chia, Mark A., Wagner, Siegfried K., Ayhan, Murat S., Williamson, Dominic J., Struyven, Robbert R., Liu, Timing, Xu, Moucheng, Lozano, Mateo G., Woodward-Court, Peter, Kihara, Yuka, Altmann, Andre, Lee, Aaron Y., Topol, Eric J., Denniston, Alastair K., Alexander, Daniel C., Keane, Pearse A.
• Format: Journal Article
• Language: English
• Published: London Nature Publishing Group UK 05.10.2023; Nature Publishing Group
• Subjects: 692/308/575; 692/699/3161/3175; 692/699/75; 692/700/1421; 692/700/1750; Annotations; Artificial Intelligence; Clinical medicine; Congestive heart failure; Datasets; Diabetes; Diabetic retinopathy; Diagnosis; Disease detection; Efficiency; Eye; Eye diseases; Eye Diseases - complications; Eye Diseases - diagnostic imaging; Heart attacks; Heart failure; Heart Failure - complications; Heart Failure - diagnosis; Hospitals; Humanities and Social Sciences; Humans; Labels; Macular degeneration; Medical diagnosis; Medical imaging; Modelling; multidisciplinary; Myocardial infarction; Myocardial Infarction - complications; Myocardial Infarction - diagnosis; Retina; Retina - diagnostic imaging; Retinal images; Science; Science (multidisciplinary); Self-supervised learning; Supervised Machine Learning; Systemic diseases; United Kingdom > UK
• ISSN: 0028-0836; 1476-4687
• DOI: 10.1038/s41586-023-06555-x

Medical artificial intelligence (AI) offers great potential for recognizing signs of health conditions in retinal images and expediting the diagnosis of eye diseases and systemic disorders 1 . However, the development of AI models requires substantial annotation and models are usually task-specific with limited generalizability to different clinical applications 2 . Here, we present RETFound, a foundation model for retinal images that learns generalizable representations from unlabelled retinal images and provides a basis for label-efficient model adaptation in several applications. Specifically, RETFound is trained on 1.6 million unlabelled retinal images by means of self-supervised learning and then adapted to disease detection tasks with explicit labels. We show that adapted RETFound consistently outperforms several comparison models in the diagnosis and prognosis of sight-threatening eye diseases, as well as incident prediction of complex systemic disorders such as heart failure and myocardial infarction with fewer labelled data. RETFound provides a generalizable solution to improve model performance and alleviate the annotation workload of experts to enable broad clinical AI applications from retinal imaging. RETFound, a foundation model for retinal images that learns generalizable representations from unlabelled images, is trained on 1.6 million unlabelled images by self-supervised learning and then adapted to disease detection tasks with explicit labels.