FedPC: An Efficient Prototype-Based Clustered Federated Learning on Medical Imaging.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Science

Background:

• Federated learning (FL) enables collaborative training while preserving data privacy.
• Data heterogeneity in FL leads to performance degradation.
• Existing clustered federated learning (CFL) methods struggle with accurate client representation and clustering.

Purpose of the Study:

• To propose an efficient prototype-based CFL framework (FedPC) to address data heterogeneity.
• To improve client clustering accuracy and overall cluster performance in FL.
• To reduce communication overhead in federated learning systems.

Main Methods:

• Introduced a dual-prototype strategy (specific and generalized prototypes) for capturing client class representations.
• Implemented a prototype-contrastive training mechanism to enhance intra-cluster prototype consistency.
• Evaluated the framework on medical imaging datasets (BloodMNIST and DermaMNIST).

Main Results:

• FedPC outperformed nine state-of-the-art (SOTA) CFL methods on BloodMNIST and DermaMNIST, with average accuracy improvements of 2.17% and 3.47%.
• The FedPC framework reduced communication overhead by 3.33 to 5.68 times compared to SOTA methods.
• Demonstrated superior performance and efficiency in handling data heterogeneity in FL.

Conclusions:

• The proposed FedPC framework effectively addresses data heterogeneity in FL through advanced prototype learning.
• FedPC offers significant improvements in clustering accuracy and model performance.
• The framework presents a practical and efficient solution for real-world federated learning applications, especially in medical imaging.