In this paper, we propose FedCLAM, a novel method for improving the performance of medical image segmentation in federated learning. To address the problem that existing federated learning methods suffer from performance degradation due to feature differences between institutions, FedCLAM introduces client-adaptive momentum and personalized dampening factor to prevent overfitting, and intensity alignment loss to address the difference in intensity profiles of heterogeneous medical images. Experimental results on two datasets show that FedCLAM outperforms eight existing state-of-the-art methods. The source code is open source.
