DG-UCTB: A Domain-Generalized Uncertainty-Calibrated Framework for Tuberculosis Detection in Chest Radiographs

Abstract

Tuberculosis remains a major global health burden, creating a need for accurate and generalizable computer-aided screening methods based on chest radiographs. Deep learning models trained on a single imaging cohort can experience substantial performance degradation when applied to data acquired from different institutions. This study proposes DG-UCTB, a domain-generalized and uncertainty-calibrated framework for automated tuberculosis detection in chest radiographs. The framework combines an EfficientNet-B0 backbone with learnable spatial attention, appearance-based domain randomization, prediction-consistency regularization, temperature scaling, and test-time augmentation to improve cross-domain robustness while maintaining computational efficiency. The Shenzhen chest radiograph dataset was divided into training, validation, and independent internal test subsets, while the Montgomery dataset was retained exclusively for external validation. On the Shenzhen internal test set, DG-UCTB achieved an AUROC of 0.9480 with a 95 percent confidence interval of 0.9040 to 0.9811, an AUPRC of 0.9544, an accuracy of 0.8500, an F1-score of 0.8352, and a Matthews correlation coefficient of 0.7187. External evaluation on the Montgomery cohort achieved an AUROC of 0.8067 with a 95 percent confidence interval of 0.7329 to 0.8797 and an AUPRC of 0.7612, demonstrating preservation of discriminatory capability under cross-dataset variation. External specificity reached 0.9750, while sensitivity was 0.1897, indicating that operating-point transfer across heterogeneous imaging cohorts remains challenging. The findings demonstrate the potential of domain-oriented augmentation, consistency learning, uncertainty estimation, and calibration to improve the robustness of lightweight tuberculosis screening models across imaging cohorts.