TreeFedDG: Alleviating Global Drift in Federated Domain Generalization for Medical Image Segmentation

Yucheng Song, Chenxi Li, Haokang Ding, Zhining Liao, Zhifang LiaoarXiv 2025 · DOI · arXiv

Introduction

Training robust deep learning models for medical image segmentation requires large, diverse datasets. However, strict data privacy regulations like HIPAA and GDPR often prevent the centralized collection of patient data. Federated Learning (FL) has emerged as a solution, allowing models to be trained across multiple institutions without sharing raw data. Despite its promise, standard FL faces a significant challenge in medical contexts: Global Drift (GD).

Domain heterogeneity—caused by variations in scanner hardware, imaging protocols, and patient populations across different hospitals—leads to divergent local model updates. When a central server averages these disparate updates, the resulting global model often "drifts" away from an optimal solution for any single domain, particularly failing on unseen target domains. This paper introduces TreeFedDG, a hierarchical framework designed to mitigate Global Drift and improve Federated Domain Generalization (FedDG).

Methodology: The Tree-Structured Framework

Unlike traditional "star" topologies that use a single global server, TreeFedDG organizes client models into a hierarchical tree structure based on their parameter similarity.

Overview of the TreeFedDG framework, featuring hierarchical parameter aggregation, FedStyle mixing, and similarity-guided inference.

Hierarchical Parameter Aggregation

The framework clusters similar client models into nodes within a tree. Aggregation proceeds bottom-up, where models within a cluster are averaged to form a parent node. This multi-granularity approach ensures that the model captures both domain-specific nuances at the leaves and broader clinical commonalities at the root, preventing the "confusion" inherent in flat, one-size-fits-all averaging.

FedStyle: Targeted Style Mixing

To enhance robustness against domain shift, the authors propose FedStyle. This technique identifies the most dissimilar client pairs in the parameter space and mixes their image "styles" (feature statistics) during training. By exposing the model to extreme variations between the most divergent sites, it learns to extract features that are invariant to the specific "style" of any single hospital.

Similarity-Guided Inference

At deployment, TreeFedDG does not simply provide the root global model to a new, unseen hospital. Instead, it extracts features from the new site's images and identifies the most similar source hospital in the tree. It then utilizes a "model chain"—a weighted combination of models from that specific leaf up to the root—to perform the final segmentation.

Experimental Evaluation

The researchers evaluated TreeFedDG on two benchmarks: a multi-site Fundus Dataset for retinal optic disc/cup segmentation and a multi-site Prostate MRI Dataset.

Performance and Precision

TreeFedDG consistently outperformed state-of-the-art (SOTA) methods like FedAvg, FedDG, and FedEvi. On the Fundus dataset, it achieved a Dice score of 88.29%, a significant improvement over previous benchmarks. In prostate segmentation, it reached a Dice score of 90.35% and reduced the 95% Hausdorff Distance (HD95) to 7.43, indicating much sharper and more accurate anatomical boundaries.

Qualitative comparison of segmentation results across different federated learning methods on fundus and prostate datasets.

Stability and Reliability

A critical finding was the reduction in performance variance. In clinical settings, consistency is as important as average accuracy. TreeFedDG significantly lowered the Standard Deviation (STD) of Dice scores across different sites. For instance, in optic cup segmentation, the STD dropped to 3.51, compared to 6.58 for standard FedAvg. This suggests that TreeFedDG is more reliable when deployed to new, unseen clinical environments.

Conclusion

TreeFedDG addresses the fundamental issue of Global Drift in federated medical imaging. By replacing flat aggregation with a hierarchical tree and introducing targeted style mixing, the framework achieves superior generalization and stability. While the hierarchical structure introduces some computational complexity, the gains in segmentation accuracy and cross-site reliability represent a significant step toward deployable federated AI in healthcare. Future work will focus on scaling this approach to larger networks and optimizing the feature extraction process for diverse medical niches.

Summary updated

Abstract

In medical image segmentation tasks, Domain Generalization (DG) under the Federated Learning (FL) framework is crucial for addressing challenges related to privacy protection and data heterogeneity. However, traditional federated learning methods fail to account for the imbalance in information aggregation across clients in cross-domain scenarios, leading to the Global Drift (GD) problem and a consequent decline in model generalization performance. This motivates us to delve deeper and define a new critical issue: global drift in federated domain generalization for medical imaging (FedDG-GD). In this paper, we propose a novel tree topology framework called TreeFedDG. First, starting from the distributed characteristics of medical images, we design a hierarchical parameter aggregation method based on a tree-structured topology to suppress deviations in the global model direction. Second, we introduce a parameter difference-based style mixing method (FedStyle), which enforces mixing among clients with maximum parameter differences to enhance robustness against drift. Third, we develop a a progressive personalized fusion strategy during model distribution, ensuring a balance between knowledge transfer and personalized features. Finally, during the inference phase, we use feature similarity to guide the retrieval of the most relevant model chain from the tree structure for ensemble decision-making, thereby fully leveraging the advantages of hierarchical knowledge. We conducted extensive experiments on two publicly available datasets. The results demonstrate that our method outperforms other state-of-the-art domain generalization approaches in these challenging tasks and achieves better balance in cross-domain performance.
BibTeX
@misc{song2025treefeddg,
  author = {Yucheng Song and Chenxi Li and Haokang Ding and Zhining Liao and Zhifang Liao},
  title = {{TreeFedDG: Alleviating Global Drift in Federated Domain Generalization for Medical Image Segmentation}},
  year = {2025},
  eprint = {2510.18268},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2510.18268},
}
Figure