Privacy-aware federated learning for pneumonia detection in chest X-Ray images using MobileNetV2

Abstract

Pneumonia is the main cause of morbidity and mortality on the planet, and correct diagnosis based on chest radiographs is vital for timely intervention. Nonetheless, medical imaging deep learning models have commonly been limited by strict privacy policies that do not allow the central aggregation of sensitive patient information. This paper presents a privacy-preserving federated learning (FL) Python framework for pneumonia detection in the Kaggle Chest X-Ray dataset that seeks to address data governance, institutional autonomy, and non-IID data heterogeneity issues. The lightweight MobileNetV2 network was initialized with ImageNet weights and executed on four simulated hospital clients in the Flower (FLWR) framework. All clients had two epochs of local training per round, and all knowledge was synchronized through the Federated Averaging (FedAvg) algorithm. The data were divided according to a Dirichlet non-IID distribution to simulate realistic cross-hospital variations. The results of the global model over ten communication rounds reached 0.9924 accuracy, 0.9949 F1-score, and 0.9998 AUC, which indicates convergence and resistance to non-homogeneous client distributions. The findings validate the hypothesis that federated learning can perform diagnostic tasks as well as centralized training while maintaining privacy and minimizing patients' computational requirements. The research provides an experimentally verified FL pipeline that can be applied to low-resource clinical environments and provides a basis for the future integration of differential privacy, secure anonymization, and multi-task medical imaging.