Federated Learning for Computational Pathology: A Pilot Study Across Four Institutions

Federated Learning (FL) offers a solution for collaborative machine learning in healthcare by enabling institutions to train shared models without exchanging sensitive patient data. FL is particularly crucial in clinical domains like digital pathology, where ethical and regulatory barriers often hinder centralized data sharing. Yet, the practical feasibility and effectiveness of FL in this domain remain largely untested. In this study, we report a pilot FL deployment for tumor segmentation in histopathology, conducted across four pathology centers. Each institution contributed whole slide images of oral cavity cancer, encompassing natural and data variations. An FL pipeline was established in which models were trained locally at each site and periodically synchronized via a central coordination server. Final model evaluation was conducted on a completely unseen, held-out test set to assess generalization and the results compared to centralized learning (CL). Our findings reveal that FL can achieve near-parity with CL, with a statistically insignificant difference between segmentation Dice coefficients. Segmentation Dice scores reached 0.825 ± 0.0289 in the centralized setup, and 0.820 ± 0.0153 and 0.810 ± 0.0399 using FL methods FedProx and FedAvg, respectively. These results demonstrate that federated approaches can maintain high segmentation accuracy while preserving data privacy. This work provides critical evidence that FL is not only feasible but also highly effective for real-world, multi-institutional collaboration in computational pathology laying the groundwork for future large-scale, privacy-aware AI development in clinical environments.