A federated foundation model framework for collaborative histopathological intelligence in multicenter digital pathology

Scalable AI in digital pathology continues to face three closely related challenges. Patient confidentiality regulations restrict the transfer of whole‑slide images (WSIs) between healthcare institutions, making direct cross-institutional data sharing difficult. At the same time, widely used federated aggregation methods, including FedAvg, generally assume that participating institutions have comparable reliability and similarly distributed data. Such assumptions are often unrealistic in multihospital environments, where substantial heterogeneity may exist across institutions. A further limitation is the lack of federated frameworks that combine a pathology-specific self-supervised foundation model for pre-training with transferable morphological representation learning. This study introduces Federated Pathology Foundation Learning (FPFL), a five-layer framework for privacy-preserving collaborative analysis of gigapixel whole-slide images (WSIs). Each participating hospital performs local self-supervised contrastive learning using its WSI archive and a ViT-B/16 backbone. Only differentially private gradient updates are transmitted to the central aggregator, while WSIs remain within institutional boundaries. Institution-aware adaptive aggregation accounts for variations in local data distributions. FPFL is evaluated on CAMELYON16, CAMELYON17, PANDA, and TCGA-LUAD, achieving 93.7% accuracy, 96.4% AUC, and 93.1% F1. Privacy is controlled using ε = 3.2 and δ = 10−5.