This is the PyTorch implemention of our paper Fair federated learning for weakly supervised nuclei segmentation via feature disentanglement by Qian, Yi and Pan, Xipeng and Yu Hang and Wen, Yimin and Bian, Xinjun
In medical image analysis, nuclei segmentation development is hindered by high acquisition costs and privacy regulations, whereas federated learning (FL), a privacy-preserving distributed method, shows great potential to address these challenges. However, domain skew remains a significant issue, manifested in two key problems: 1) Feature Variations among Clients: Variations in data collection environments yield domain-specific and domaingeneral features, whose aggregation into the global model can degrade performance on other clients. 2) Model Aggregation Bias: Domain diversity across clients requires the federated model to handle multiple challenges, but the fixed aggregation weights used by most methods cannot cover all scenarios and thus bias model aggregation. To address these challenges, we propose a novel federated learning framework comprising Disentangled Orthogonal Network (DONet) and Federated Profit Estimation (FedPE). DONet employs a dual-branch design to separate cross-client and local features, leveraging an orth-attention mechanism to minimize the interference of domain-specific features on the global model. FedPE is a fairness-based strategy that uses an auxiliary model to evaluate client participation benefits and adaptively adjust aggregation weights, thereby correcting global model biases and ensuring fairer updates across domains. Extensive experiments show that our method surpasses stateof-the-art FL approaches in performance and reliability.
python train.pyIf you find the code and dataset useful, please cite our paper.
@article{QIAN2026113293,
title = {Fair Federated Learning for Weakly Supervised Nuclei Segmentation via Feature Disentanglement},
journal = {Pattern Recognition},
pages = {113293},
year = {2026},
issn = {0031-3203},
doi = {https://doi.org/10.1016/j.patcog.2026.113293}