PathoFL: Federated Learning for Digital Pathology Images Across Hospitals

PathoFL is an open science project aimed at developing and validating an AI-based framework for the analysis of digital pathology images using federated learning (FL). By training machine learning models on publicly available whole-slide images (WSIs) — including Hematoxylin and Eosin (H&E) and immunohistochemistry (IHC) stained samples — PathoFL enables AI-based cancer prognosis without the need to share sensitive patient data across institutions. The project embraces the FAIR principles by making both the trained models and training workflows openly accessible, interoperable, and reusable via the European Open Science Cloud (EOSC).

Traditional AI development in digital pathology requires centralizing large volumes of sensitive patient data from multiple hospitals and research institutions, which raises serious ethical, legal, and logistical concerns, especially under GDPR. Additionally, few tools exist to enable privacy- preserving, collaborative AI training in pathology, and many publicly available WSIs are underutilized due to lack of standardized preprocessing and annotation protocols. Finally, many small research groups lack the computational infrastructure or resources to participate in this domain, creating barriers to both adoption and reproducibility.

PathoFL proposes a novel, lightweight, and scalable federated learning pipeline tailored for digital pathology. The project will source H&E and IHC WSIs from existing public datasets, preprocess them using reproducible workflows (e.g., QuPath scripts, U-Net segmentation), and train deep learning models in a simulated federated setting. The pipeline will include automated region-of-interest (ROI) detection, cancer grading, and prognostic score prediction. All code, trained models, and documentation will be released openly and packaged following EOSC- compatible standards (e.g., RO-Crate, metadata schema.org), allowing other groups to replicate and extend the work without requiring centralized data collection (Figure 1).

In terms of infrastructure, PathoFL will utilize a local high-performance workstation (Dell Precision with NVIDIA A2000 GPU) and EOSC cloud resources to demonstrate model training across multiple "nodes" representing different virtual institutions. The project will also include outreach activities such as a public workshop to foster adoption among digital pathology researchers, especially those in smaller labs or countries with limited access to data-sharing agreements.

PathoFL advances open science in digital pathology by demonstrating the feasibility and effectiveness of federated learning on heterogeneous, publicly available WSIs — a domain where data privacy and reproducibility are especially challenging. The project will reduce the technical and ethical barriers for training and deploying AI models across distributed pathology data, helping democratize access to precision oncology tools. Furthermore, PathoFL directly supports EOSC objectives by contributing fully documented, open-source pipelines and training resources that can be adopted by other research groups or integrated into existing European science clusters.

By increasing the transparency, reusability, and privacy compliance of digital pathology AI models, PathoFL has the potential to improve diagnostic workflows, support precision cancer medicine, and foster cross-border collaboration without compromising patient confidentiality. The approach can also serve as a blueprint for federated learning in other biomedical imaging domains.

PathoFL aims to bridge digital pathology with federated learning by simulating decentralized model training on public datasets. The main scientific and innovation objective of the PathoFL project is to develop a federated learning (FL) framework tailored for digital pathology, using publicly available whole-slide images (WSIs) for cancer prognosis without requiring direct data sharing. The project aims to leverage open-source tools and cloud resources to simulate federated learning across institutions, enabling the training of deep learning models in a GDPR- compliant manner. Through this approach, PathoFL will significantly advance the application of open science in digital pathology, a field reproducibility, and privacy-preserving collaboration.