STPFL: Stable teacher-guided personalized federated learning for heterogeneous multi-institution medical imaging

Abstract Multi-institutional medical imaging data are distributed across sites and cannot be pooled for centralized training due to privacy and governance constraints. Federated learning enables collaborative training without sharing raw data; however, under non-IID client distributions, it often faces unstable optimization and weak global-local knowledge transfer. Personalized federated learning mitigates this issue through client-specific adaptation, but many regularization-, meta-learning-, or distillation-based methods still degrade under highly skewed partitions. To address this, we propose STPFL, a stable teacher-guided personalized federated learning framework. STPFL builds an EMA-based aggregated teacher to accumulate historical global knowledge and provide consistent guidance. Each client optimizes a unified objective combining supervised learning with temperature-scaled KL distillation from the teacher model. On the server side, STPFL applies validation-driven performance weighting as a reliability signal and performs Adam-style aggregation to improve convergence under cross-silo/multi-institutional client distributions. We evaluate STPFL under controlled data heterogeneity (α = 0–1) and cross-silo client-count scalability (1–10 clients) on three benchmark datasets (PathMNIST, BloodMNIST, TissueMNIST) and three real-world medical imaging datasets (brain tumour MRI, blood cell cancer, and Kermany2018 OCT). Compared with FedAvg, FedProx, MetaFed, and FedLFP, STPFL achieves 4–23% and 6–29% improvements in global accuracy and F1-score, while personalized models improve by 4–25% and 4–32%, respectively. Under cross-silo client-count scalability, STPFL further improves global accuracy and F1-score by 5–30% and 5–40%, with corresponding personalized gains of up to 13–33% and 17–47%.