Flexible and Scalable Federated Learning with Deep Feature Prompts for Digital Pathology

Collaborative learning across medical institutions is considered essential for developing robust and generalisable models in digital pathology. Federated learning (FL) offers a promising framework for enabling collaboration without the need to centralise data. Still, its practical adoption remains limited due to challenges such as high communication overhead, model heterogeneity across institutions, and privacy concerns. To address these challenges, we propose Federated Deep Feature Prompting (FedDFP), a novel and efficient FL paradigm designed to operate in real-world, heterogeneous clinical environments. FedDFP introduces a lightweight, learnable, client-specific prompt applied to patch-level embeddings extracted from whole slide images. By sharing only these compact prompts across clients ( i.e. , participating centres), FedDFP drastically reduces communication costs by over 99.9% compared to standard FL frameworks, with improved classification accuracy across clients. Through comprehensive evaluations on three public datasets (TCGA-IDH, CAMELYON16, and CAMELYON17), which include breast cancer and glioma classification tasks, FedDFP consistently outperforms both standard and personalised FL baselines. It maintains strong performance across diverse feature extractors and heterogeneous MIL classifiers, demonstrating flexibility, scalability, and robustness. These results highlight FedDFP as a promising solution for efficient, accurate, and privacy-conscious federated learning in digital pathology.