FedDPI-SH: A Quality and Similarity Aware Federated Learning Framework for Medical Image Analysis

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence
• Data Privacy

Background:

• Federated learning (FL) enables decentralized medical image analysis, but struggles with data heterogeneity and inter-client feature similarity.
• Existing FL methods often yield suboptimal performance due to these challenges.

Purpose of the Study:

• To propose FedDPI-SH, a framework addressing data quality heterogeneity and inter-client feature similarity in FL for medical imaging.
• To improve the performance of decentralized medical image analysis.

Main Methods:

• Developed FedDPI-SH, incorporating a Data Performance Index (DPI) and Client Representation Similarity Matrix (CRSM) for weighted aggregation.
• DPI quantifies client reliability using dataset size, image resolution, label balance, duplication, and generalization accuracy.
• CRSM measures inter-client feature alignment via cosine similarity.

Main Results:

• FedDPI-SH demonstrated significant improvements across three medical imaging datasets (Medical MNIST, PathMNIST, HAM10000) under non-IID conditions.
• On HAM10000, FedDPI-SH achieved 81.24% balanced accuracy, substantially outperforming MOON, FedAvg, FedAvgM, and FedProx.
• The framework effectively handles data quality heterogeneity, including duplication, label balance, and resolution variations.

Conclusions:

• FedDPI-SH offers a robust solution for federated medical image analysis by effectively managing data heterogeneity and similarity.
• The proposed quality and similarity-aware weighted aggregation significantly enhances model performance and reliability in decentralized settings.