Fedharmo: Harmonizing Global Aggregation and Local Alignment in Cross-Domain Federated Medical Image Segmentation

Federated learning has emerged as the mainstream technique for breaking data silos in the healthcare domain by enabling collaborative training of machine learning models among different clients while ensuring data privacy. However, due to the discrepancy in the data distribution of each client, the aggregated model may ultimately be shifted towards certain clients. Consequently, we propose FedHarmo, a federated learning segmentation framework that employs an adaptive aggregation weight updating strategy to prevent model aggregation shifted towards certain clients. Additionally, we utilize the distribution measurement metric to minimize the distance between the global model and local models. Comprehensive experiments on ten medical image datasets demonstrate the effectiveness of our proposed method. Our code are available at https://github.com/GGbond-study/FedHarmo.