From generalization to precision: A large domain-specific pretrained model for specialized medical tasks.

Foundation models are recognized for their broad generalizability across diverse tasks, which has driven their widespread adoption. In medicine, however, precision is often paramount, particularly in the era of precision medicine, where diagnostic and therapeutic decisions demand exceptional accuracy. In this study, we investigate whether a large domain-specific pretrained model can outperform general-purpose foundation models in specialized medical tasks based on ultra-widefield fundus (UWF) imaging. We develop USPEC, a large-scale domain-specific model pretrained on 875,947 unlabeled UWF images using self-supervised learning, and subsequently adapt it to downstream tasks with explicit labels. USPEC consistently outperforms the foundation models (RETFound, RETFound-DE, VisionFM, RetiZero, and EyeFound) across multiple UWF-based diagnostic and few-shot learning tasks, while also demonstrating superior label and training efficiency. Its robustness is further confirmed on independent datasets from different countries. These findings support a large domain-specific pretrained model as a powerful paradigm for advancing artificial intelligence (AI) performance in specialized medical fields.