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

Abstract

Federated learning (FL) enables decentralized medical image analysis while preserving data privacy. However, conventional methods overlook client data heterogeneity and inter-client feature similarity, resulting in suboptimal performance. In this paper, our proposed FedDPI-SH framework addresses these limitations through quality and similarity-aware weighted aggregation. The framework introduces a Data Performance Index (DPI) quantifying client reliability through dataset size, image resolution, label distribution balance, duplication levels, and cross-client generalization accuracy. Client Representation Similarity Matrix (CRSM) measures inter-client feature alignment via cosine similarity. FedDPI-SH combines these components to compute aggregation weights, prioritizing high-quality clients during feature extractor updates while maintaining personalized classifiers. Evaluation across three medical imaging datasets (Medical MNIST, PathMNIST, HAM10000) under severe non-IID conditions demonstrates improvements, with HAM10000 achieving 81.24% balanced accuracy, outperforming MOON (58.85%), FedAvg (45.22%), FedAvgM (17.66%), and FedProx (12.64%).The framework addresses data quality heterogeneity through explicit assessment of duplication, label balance, and resolution in federated medical imaging.