Privacy-Preserving Federated Anomaly Detection on Chest X-ray Images

We propose a federated learning (FL) framework for automated chest X-ray diagnosis using the NIH ChestX-ray14 dataset, comprising 112,120 frontal radiographs labeled with 14 thoracic pathologies. Leveraging the Flower library, our approach simulates cross-institutional collaborative training without centralizing sensitive patient data. We evaluate both a simple CNN and ResNet-18 architecture in federated and centralized settings. Results show that the federated ResNet-18 achieves over 98 percent of the AUC performance of a centrally trained model across most disease labels, surpassing models trained on isolated datasets. Our FL implementation, based on the FedAvg algorithm, reaches radiologist-level performance in detecting common conditions like pneumonia while maintaining data privacy. We also investigate the impact of key FL hyperparameters, such as the number of clients and local training epochs, on model convergence. To ensure reproducibility, we detail our FL setup, data processing pipeline, and training configuration. Lastly, we discuss privacy considerations, current limitations—including simulation constraints and class imbalance—and future directions for improving privacy-preserving chest X-ray analysis.