FedCLAM: Client Adaptive Momentum with Foreground Intensity Matching for Federated Medical Image Segmentation

Vasilis Siomos, Jonathan Passerat-Palmbach, Giacomo Tarroni

City St George's, University of London

We propose FedCLAM, a federated learning method that addresses data heterogeneity in medical image segmentation through client-adaptive momentum and foreground intensity matching. Our approach achieves state-of-the-art performance on cardiac MRI and abdominal CT segmentation tasks.

MICCAI 2025 / Spotlight (Top 11%)

Abstract

Federated Learning (FL) enables collaborative training across institutions without sharing sensitive medical data. However, heterogeneous data distributions across clients remain a major challenge, particularly in medical image segmentation where intensity variations and class imbalances are common.

We introduce FedCLAM (Federated Client-Adaptive Momentum), a novel aggregation strategy that adapts momentum coefficients based on client-specific data characteristics. Our method incorporates foreground intensity matching to align feature distributions across clients, improving segmentation consistency without compromising privacy.

Key Contributions

Our work makes the following contributions:

Results

We evaluate FedCLAM on two challenging medical image segmentation benchmarks: cardiac MRI (M&Ms dataset) and abdominal CT (BTCV dataset). Our method consistently outperforms existing federated learning approaches including FedAvg, FedProx, and SCAFFOLD.

On cardiac MRI segmentation with 4 heterogeneous clients, FedCLAM achieves 87.2% Dice score compared to 83.4% for FedAvg, representing a significant improvement in cross-institutional performance.

Citation

If you find this work useful, please cite our paper:

@inproceedings{siomos2025fedclam,
    title={FedCLAM: Client Adaptive Momentum with Foreground Intensity 
           Matching for Federated Medical Image Segmentation},
    author={Siomos, Vasilis and Passerat-Palmbach, Jonathan and Tarroni, Giacomo},
    booktitle={International Conference on Medical Image Computing and 
               Computer-Assisted Intervention (MICCAI)},
    year={2025}
}