General-purpose versus retina-specific foundation models for multilabel retinal disease classification: effects of adaptation and class weighting

Whether a retina-specific or a general-purpose foundation model performs better for retinal disease classification can depend on how the model is adapted and how class imbalance is handled during training. These factors have not been examined jointly for this model pair. In this retrospective comparative study, RETFound-MAE and DINOv2 were compared on patient-disjoint BRSET splits as validation-selected linear probes and as fully fine-tuned models, with supervised ResNet-50 as a reference. External testing used ODIR-5K, mBRSET, and APTOS-2019, and acquisition shift was evaluated between overlap-free Canon and Nikon subsets. DINOv2 had higher macro AUROC than RETFound-MAE under linear probing (0.897 vs 0.822) and after fine-tuning (0.937 vs 0.911), and fine-tuning narrowed the difference. Class-weighted training produced substantial calibration error. Correcting the weighting-induced probability offset reduced macro ECE from 0.152 to 0.007 for RETFound-MAE and from 0.066 to 0.012 for DINOv2, and matched unweighted training improved both discrimination and calibration. Fine-tuned DINOv2 had the highest observed AUROC on all three external cohorts, and its frozen probe showed a smaller Canon-to-Nikon decrease than the RETFound-MAE probe. Under the evaluated recipes, backbone choice, adaptation regime, and loss formulation each influenced performance.