FedCGP: Cluster-Based Gradual Personalization for Federated Medical Image Segmentation

Federated learning empowers the privacy-preserving training of a global model in decentralized medical scenarios. Subsequently, personalized and clustered federated learning have been respectively proposed to address data heterogeneity across clients. However, the limited performance gains for clients with diverse data distributions and the increased complexity of personalized federated learning, along with the potential for suboptimal assignments in clustered federated learning necessitates further exploration. To address these limitations, this paper develops a cluster-based gradual personalization framework for federated learning (FedCGP) to integrate personalization and clustering in medical image segmentation. FedCGP accomplishes personalization by dynamically adjusting the aggregation strategy during iterative clustering among clients. Initially, a typical homogeneous aggregation strategy is employed to extract common features from the clients' data before the cluster convergence. Once a cluster converges, shared and personalized layers are identified in the client models within the cluster based on layer-wised parameter similarity. Shared layers are then aggregated across client model layers within the cluster, while personalized layers are retained within each client to better adapt to their unique data features. Experiments are implemented on two human tissues imaged by seven datasets to validate the advantage of the proposed algorithm. The benefits of FedCGP are demonstrated in a comparison of segmentation tasks, and the effectiveness of FedCGP is interpreted in setting analysis and ablation studies.