HistoFL

Deep learning has shown strong performance in computational pathology but typically requires large, diverse, and carefully curated datasets, which are difficult to share across institutions due to privacy and logistical constraints. This paper introduces a privacy-preserving federated learning framework using weakly supervised multiple-instance learning and differential privacy to train models on distributed gigapixel whole-slide images without sharing raw data. The approach is evaluated on diagnostic and survival-prediction tasks using thousands of slide-level–labeled images and demonstrates that accurate models can be learned while maintaining patient privacy.