Nationwide federated learning for histopathology: secure deployment across Germany behind firewalls

npj Digital Medicine volume 9 , Article number: 581 ( 2026 ) Cite this article

Federated Learning (FL) enables collaborative training across institutions without sharing sensitive data, a solution for privacy-preserving AI in medical imaging. However, hospital deployment remains challenging due to strict data protection regulations, heterogeneous infrastructures, and limited network accessibility behind firewalls. We introduce TheODen, an open-source framework for Federated training on histopathology Whole Slide Imaging (WSI). It requires no open client-side ports, enabling training through firewalls via a secure reverse-proxy architecture. We conducted, to our knowledge, the first nationwide FL study for histopathology segmentation of colorectal cancer across three German university hospitals, using breast and colorectal cancer datasets without opening firewall ports. TheODen achieves robust segmentation, with global average dice scores of 0.764 on BCSS and 0.754 on SemiCOL despite data heterogeneity and network constraints. These findings underline TheODen’s potential to facilitate secure, large-scale collaborations between medical institutions and to accelerate clinical translation of AI models under real-world infrastructure constraints, providing a privacy-preserving-by-design architecture for future collaborations.

Segmentation models for histopathology data are emerging as important methods to support pathologists and mitigate the shortage of pathologists 1 . However, these models depend on the extensive availability of training data 2 . The amount of required training cases and their variability typically requires extending the training to multiple hospitals 3 , 4 . Due to privacy regulations such as General Data Protection Regulation (GDPR) in Europe or Health Insurance Portability and Accountability Act (HIPAA) in the US, it is often infeasible to share the actual training data between multiple hospitals. 5 Federated Learning 6 has emerged as a paradigm to address this issue 7 as it allows privacy-aware, decentralized training with each hospital training a local model on their data and only sharing the resulting model parameters with a central server instance instead of the potentially privacy-critical training data 8 .

Existing work has simulated Federated Learning in the context of histopathology segmentation and shown that it is able to achieve performance similar to the privacy-critical centralized training 9 , 10 , 11 . However, these works were usually just simulated locally instead of actually running spatially distributed over the internet across hundreds of kilometers and covering an area of 10,500 km 2 . These approaches would not be securely applicable to real-world training over the internet because they require a direct connection to the clients and therefore opening ports of the firewall. This is typically not feasible as it causes serious security implications and would therefore usually violate IT security policies of the hospitals. Consequently, most Federated Learning frameworks demonstrate promising proof-of-concept results but still avoid the key technical challenges of real-world deployment.

In this work, we propose TheODen, a framework for Federated Learning, specialized to histopathology tasks, that allows simple and secure deployment on hospital infrastructure over the internet without requiring to open any ports at the firewalls. This simplifies the deployment because it does not require any coordination with the hospital IT administrations or additional security risks from firewall modifications. We show that it is possible to set up a training on a client with 2 lines of client code plus a 5-line configuration file. The source code is publicly available on GitHub: https://github.com/MECLabTUDA/theoden .

We show in a nationwide deployment between three German university hospitals, that is the first of its kind, that TheODen can be trained in such a real-world scenario. We evaluate the performance on the breast cancer segmentation dataset (BCSS) 12 leading to a dice score of 0.764. Moreover, we evaluate the behavior of the training under distribution shifts as they typically occur when using data from different institutions. Therefore, we additionally train on the multi-site dataset SemiCOL 13 that contains WSIs of colorectal cancer from different hospitals and distribute them between the clients. Despite this scenario of Client Drift, we still achieve an average dice score of 0.754 on SemiCOL using TheODen. These results highlight the potential of TheODen to leverage data availability between hospitals through an architecture designed to preserve patient privacy by keeping data local. Figure 1 provides an overview of the structure of TheODen and its main features.

We train on two distinct histopathology datasets, BCSS and SemiCOL. To achieve this, we deploy our own open-source framework, TheODen. This framework employs a reverse proxy positioned outside the server’s firewall, enabling training through closed firewalls. This significantly enhances security and simplifies deployment, as there is no need to modify the firewall configurations of the hospitals. The reverse proxy achieves this by forwarding all communication to server and clients that are all located behind their firewalls.

The public code repository contains detailed tutorials and example code, allowing clinicians to reproduce the results and deploy their own training on medical imaging data, making TheODen a valuable foundation for larger distributed training collaborations even including international deployments in big consortia.

Existing work aims to deploy federated learning but is either limited, as it is just simulating it locally, or does not allow training without opening ports at the clients, which is both a significant limitation for clinical real-world deployment. Rieke et al. 8 introduce Federated Learning 14 as a promising approach for decentralized, privacy preserving training on distributed medical imaging data. However, there are significant challenges for real-world deployment of Federated Learning such as data distribution 15 and deployment in the infrastructure that need to be addressed by a framework.

Nvidia FLARE 16 and Flower 17 are frameworks for Federated Learning that allow distributed training between multiple machines, but they do not include any adaptations to the reality of medical imaging such as training on histopathology WSIs.

Li et al. 18 introduce a framework for Federated Learning for classification of histopathology images. They simulate the training locally, showing the potential of Federated Learning for histopathology. Lu et al. 19 simulate Federated Learning for histopathology classification. However, these approaches are just simulating Federated Learning locally without actually conducting distributed training. Consequently, they do not cover the challenges connected to deployment in the real world over network infrastructure.

Lutnick et al. 11 train a decentralized model between multiple physical clients. However, these are located in the same network, still not covering deployment on clinical infrastructure over the internet. Roth et al. 20 deploy Federated Learning to the real-world for the classification of breast density from BI-RAD radiology images. Bujotzek et al. 21 deploy Federated Learning in the context of a nationwide radiology project and summarize pitfalls. Sheller et al. 22 are training on MRI data between 10 hospitals. Tolle et al. 23 trains over the internet on CT images.

These works show the potential of Federated Learning for radiology images, while Ogier et al. 24 train using Federated Learning for histopathology classification. They show that their federated model surpasses the classification performance of all local models. However, there is no public code available, hindering deployment for similar projects. To the best of our knowledge, there are no papers with public code that deploy Federated Learning for histopathology segmentation to real clinical infrastructure in a nationwide consortium over the internet. Consequently, this work aims to bridge this gap by providing a framework optimized for simple and guided real-world deployment to clinical infrastructure for distributed federated histopathology segmentation, bringing secure real-world deployment of segmentation models for histopathology one step closer.

As a proof of concept that TheODen enables distributed training on clinical infrastructure over the internet, we deploy TheODen on clients at the three German University Hospitals Cologne, Essen and Leipzig on the histopathology datasets BCSS for breast cancer segmentation and SemiCOL for colorectal cancer segmentation. The methods section provides further details regarding the setup of the experiment including federated learning configurations, hardware configurations and datasets.

In the following, we will analyze the resulting segmentation performance in different scenarios, compare them to centralized and distributed baselines and discuss the implications of the results.

As visualized in Fig. 2 , the distributed trainings achieved a global average test dice score of 0.764 ± 0.097 on BCSS and 0.754 ± 0.066 on SemiCOL, but there are significant differences between the clients:

a (left) Breast cancer dataset (BCSS) and b (right) colorectal cancer dataset (SemiCOL). The plots display the number of samples per location and the global average at the central server in Darmstadt, providing a proof of concept for training on real-world clinical infrastructure.

In case of BCSS, the training dice score and cross entropy loss is very similar between all clients at the end of the training. However, the resulting test performance differs more with a test dice score ranging from 0.661 in Cologne to 0.854 in Leipzig. This can be explained with the different distribution of training and test data even within datasets, such as through different acquisition methods as described by Stacke et al. 25 . Our performance is comparable to that of Weng et al. 26 , suggesting successful training.

To also cover the important aspect of handling multi-site datasets, we train clients on different subsets of SemiCOL that contain patches obtained from different hospitals. Therefore, the performance difference between clients is larger. For example, the test dice score of Cologne and Essen is higher with 0.822 and 0.748 compared to Leipzig with only 0.691. This performance distribution makes sense as Leipzig was training on data obtained from Munich, unlike Cologne and Essen that trained on the SemiCOL subset from Cologne. A possible explanation for this degradation in segmentation performance at the client in Leipzig is Client Drift 15 that causes a performance degradation through heterogeneous data distribution between clients. As the client in Leipzig is training on a different subset of histopathology slides from Munich, these statistically differ from the slides obtained from Cologne that Essen and Cologne train on. Consequently, this different distribution causes Client Drift, potentially explaining the difference in segmentation performance between clients. Using a different aggregation method than FedAvg 14 could potentially mitigate this issue and even further improve performance but would also require more tuning which might not be feasible for our goal of a simple real-world deployment. However, even with FedAvg the average segmentation performance resulted in an acceptable dice score of 0.754 despite the challenges from the diverse data distribution of SemiCOL.

For simplicity we only show and discuss the average dice scores across all classes. However, we provide the detailed per-class results in Supplementary Table 1 and Supplementary Table 2 .

Figure 3 shows the training time in comparison to the dice scores on the test set for the local models of the clients in Cologne, Essen and Leipzig and thei