Abstract

Respiratory diseases such as Pneumonia, Pulmonary Fibrosis, and Pleural Effusion remain leading causes of illness worldwide, yet diagnosis still depends heavily on radiologist interpretation, which is time-consuming and prone to inter-observer disagreement, especially in resource-limited settings. Centralized deep learning models can automate this task but require pooling sensitive patient scans in one location, raising serious privacy and regulatory concerns for multi-hospital deployment. This study proposes a Federated Bayesian Convolutional Neural Network that extracts disease-specific radiographic features through convolutional layers and produces calibrated, uncertainty-aware predictions through a Bayesian layer using Monte Carlo dropout, while federated averaging allows multiple institutions to jointly train the model by sharing only weights, never raw images. The framework is evaluated on the VinDr-CXR Chest X-ray Abnormalities Detection dataset, covering four diagnostic categories including Normal, Pneumonia, Pulmonary Fibrosis, and Pleural Effusion, collected from two independent hospitals. The proposed framework achieves 96.4% accuracy, an AUC of 0.982, and a Brier score of 0.029, outperforming centralized CNN, standard federated CNN, and standalone Bayesian CNN baselines by 2.6, 2.4, and 1.5 percentage points respectively, and exceeding several published baseline models on the same task. The main novelty is the joint integration of privacy-preserving federated learning with clinically meaningful uncertainty quantification in a single pipeline, rather than treating these as separate design goals. This combination supports trustworthy, scalable, and interpretable respiratory disease screening across multiple healthcare institutions, offering a practical pathway toward privacy-compliant clinical decision support in resource-varied hospital settings and multi-institutional diagnostic networks.