<![CDATA[Machine learning has achieved diagnostic performance comparable to clinical experts on medical imaging, yet centralized training paradigms necessitate patient data aggregation, risking violations of privacy regulations such as GDPR and HIPAA. In 2023, 1,853 healthcare data breaches were reported in the United States, compromising over 133 million medical records, rendering raw inter-institutional data exchange increasingly unsustainable. Federated Learning (FL) offers a viable solution by enabling collaborative model training without data transfer. However, prior studies predominantly evaluate single algorithms and often neglect non-IID Dirichlet-distributed conditions and probabilistic calibration metrics like log-loss. This study rigorously compares FedAvg, FedProx, FedSVRG, and FedAtt across three MedMNIST v2 datasets—PneumoniaMNIST (binary), DermaMNIST, and BloodMNIST (multi-class)—using three clients under non-IID Dirichlet partitioning (α=0.1) over 50 communication rounds. FedProx demonstrates the most consistent performance and stability, achieving accuracy of 0.9521 and log-loss of 0.1850 on PneumoniaMNIST; 0.8595 and 0.4066 on BloodMNIST; and 0.5747 and 1.5996 on DermaMNIST. It also exhibits fastest convergence and superior probability calibration. Thus, FedProx’s proximal regularization enhances FL robustness against extreme clinical heterogeneity, establishing it as a scalable, privacy-preserving framework for cross-institutional medical image diagnostics.]]>
Copyrights © 2026
