Customized Federated Learning for Multi-Source Decentralized Medical Image Classification.

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Machine Learning

Background:

• Deep networks in medical imaging face challenges due to limited, privacy-sensitive data.
• Federated learning (FL) enables collaborative training without data sharing but often yields suboptimal models for local data characteristics.

Purpose of the Study:

• To introduce Customized FL (CusFL), an approach that trains client-specific models using a federated global model.
• To improve personalized model performance by enabling clients to selectively learn from the federated model.

Main Methods:

• CusFL trains private models iteratively, guided by a federated global model aggregated from previous private models.
• The federated model focuses on feature alignment using only feature extraction layers.
• The federated feature extractor guides the training of each client's personalized model.

Main Results:

• CusFL demonstrated superior performance compared to standard federated learning.
• The method effectively aligns features across diverse datasets.
• Personalized models achieved improved accuracy in medical image analysis tasks.

Conclusions:

• CusFL offers an effective strategy for privacy-preserving medical image analysis.
• The approach enhances diagnostic accuracy by leveraging collaborative learning while preserving data individuality.
• CusFL shows promise for applications in prostate cancer identification and skin lesion classification.