Uncertainty-Calibrated Deep Learning for Reliable Pneumonia Screening from Chest Radiographs

Abstract

Deep learning systems have demonstrated considerable potential in medical image interpretation, yet unreliable confidence estimates continue to limit their practical applicability in clinical screening. This paper presents an uncertainty-calibrated framework for automated pneumonia detection from chest radiographs. The approach integrates a convolutional vision backbone with predictive uncertainty estimation and a confidence-based referral mechanism, allowing ambiguous cases to be prioritized for human review. A retrospective experimental design is constructed around 16,400 chest radiographs representing pneumonia-positive and pneumonia-negative cases from three simulated clinical data sources. The model achieves an area under the receiver operating characteristic curve of 0.936, compared with 0.907 for a standard convolutional classifier. Calibration analysis demonstrates a reduction in expected calibration error from 0.084 to 0.029. When 15% of the most uncertain predictions are referred for additional review, the remaining automated predictions reach 96.1% accuracy. Performance is also assessed across image acquisition settings to examine robustness to distribution shifts. The study illustrates how uncertainty-aware learning may support more transparent and reliable AI-assisted medical screening, while emphasizing the need for prospective clinical validation.