VCC-Fed: A Multi-task Federated Learning Paradigm with Versatile Collaborative Clients

Yue Hua, Yi Li, Xin Zheng, Ming Yang, Haiyan Fu, Alan Wee-Chung Liew, Yanqing Guo

Abstract: Federated learning has achieved significant success in handling medical images while ensuring patient privacy and security. However, the limited availability of medical data significantly impacts the performance of federated medical image processing models. Therefore, to fully exploit and capture the information from limited data sources, in this work, we propose a novel multi-task FEDerated learning paradigm with Versatile Collaborative Clients, dubbed VCC-Fed, enabling an effective and privacy-guaranteed federated medical image learning method. Concretely, the proposed VCC-Fed incorporates comprehensive learning abilities from two versatile clients: classification clients for diagnosing the condition, and segmentation clients for dividing the lesion. Different network architectures are utilized for different clients to capture diverse and informative medical image representations. Furthermore, VCC-Fed involves an effective federated aggregation mechanism, where local models update versatile clients and collaboratively share partial informative representations with global model aggregation. Such indirect distribution of private patient images across clients ensures robust privacy protection, highlighting the expressive privacy preservation capabilities of the proposed VCC-Fed approach. Extensive experiments on public medical image datasets, ISIC2017 and COVID-19, demonstrate the effectiveness of the proposed VCC-Fed, with enhanced individual client performance under the powerful multi-task federated learning paradigm.

External IDs:dblp:conf/pakdd/HuaLZYFLG25