Bayesian Deep Learning Approaches for Chest X-Ray Pathology Detection: Comparative Analysis of Calibration and Discrimination
Summary
No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.
Summarise with VIRAT Free to try with your Science Catalogs account.Abstract
Deep learning models have demonstrated strong performance in automated chest radiograph interpretation; however, their deployment in clinical practice is hindered by severe class imbalance, label uncertainty, and poorly calibrated predictive confidence. In this work, we present a comprehensive uncertainty-aware analysis of thoracic pathology classification on the CheXpert dataset using Bayesian deep learning methods, including Monte-Carlo Dropout, Mean-Field Variational Inference, Stochastic Weight Averaging Gaussian (SWAG), and Bayesian Deep Ensembles built on modern convolutional backbones. We systematically evaluate these approaches across 13 clinically relevant pathologies using discriminative performance (AUROC), probabilistic calibration (Expected Calibration Error and Brier score), and a clinically motivated risk metric that jointly accounts for annotation uncertainty and predictive error. Our results show that Bayesian deep ensembles consistently achieve the highest AUROC, while Mean-Field Variational Inference and MC Dropout often yield better-calibrated predictions. We further demonstrate that pathologies characterized by high annotation ambiguity, such as pneumonia and atelectasis, exhibit elevated clinical risk across methods, underscoring the necessity of uncertainty-aware evaluation. Together, these findings highlight the trade-offs between accuracy and calibration in Bayesian inference strategies and emphasize the importance of reliable uncertainty estimation for safe and clinically meaningful deployment of deep learning systems in medical imaging.
Fields of Study & Subjects
COVID-19 diagnosis using AI
Artificial Intelligence in Healthcare and Education
Radiomics and Machine Learning in Medical Imaging
Publication details
- Journal / source
- Zenodo (CERN European Organization for Nuclear Research)
- Publication date
- 1 January 2026
- Language
- EN