Client-Specific Normalization with Selective Proximal Regularization for Heterogeneous Federated Medical Image Classification

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Accurate and reliable automated medical image classification is important for computer-aided clinical decision support. However, clinical data are often distributed across institutions and cannot be centrally shared because of privacy, governance, and regulatory constraints. Federated Learning (FL) enables collaborative training while keeping data local, but statistical heterogeneity may cause client drift and performance degradation. We propose FedPBN, a personalized FL approach combining client-specific Batch Normalization (BN) with FedProx-based proximal regularization. BN parameters and running statistics remain local, while non-BN parameters are optimized and aggregated according to client data sizes. FedPBN was evaluated on five clients using PathMNIST and BloodMNIST under IID, Dirichlet (α = 0.5), and pathological Label-Skew distributions with ResNet18, DenseNet121, and MobileNetV2, against FedAvg, FedProx, and FedBN. Personalization and generalization were examined using PFPV, GRV, and CPAV. Results show that preserving client-specific BN is critical under severe Label-Skew, whereas GRV can degrade performance; CPAV recovers much of this loss but requires centralized labeled data. MobileNetV2-FedPBN achieved 97.88% accuracy and 96.79% Macro F1 on PathMNIST. Jetson Orin Nano Super deployment further demonstrated practical hospital-side edge inference, with MobileNetV2 offering the best efficiency–performance balance and DenseNet121 the strongest BloodMNIST Client-0 accuracy.