Clinically grounded retinal representation learning from minimal supervision

Early detection of retinal abnormalities is critical for preventing avoidable vision loss, yet large-scale screening programs face challenges including limited specialist availability, heterogeneous imaging conditions, and sparse disease annotations. In routine practice, the primary screening decision is not disease subtyping but whether a fundus image is normal or abnormal, suggesting that binary supervision may provide a clinically grounded and scalable training signal for retinal representation learning. We developed M-FunFound, a clinically grounded retinal representation learning framework pretrained on binary abnormality labels from routine ophthalmic screening. The model was pretrained on 113,645 real-world fundus images and evaluated across three downstream tasks. Abnormality classification was evaluated on both internal and external datasets (JSIEC and RFMiD), while multi-label disease classification and vessel segmentation were evaluated on our internal multi-label dataset and FIVES dataset, respectively. For each task, M-FunFound was compared with models pretrained on ImageNet, FLAIR, and RETFound. M-FunFound achieved AUCs of 0.945 (internal), 0.958 (JSIEC), and 0.829 (RFMiD) for binary abnormality classification, the highest weighted F1-score (0.725) across eight disease categories in multi-label classification, and the highest Dice score for vessel segmentation (0.853). These findings suggest that binary abnormality supervision can support robust and transferable retinal representation learning while substantially reducing annotation burden in real-world screening settings.