Journal of Biomedical Advancement Scientific Research

Abstract

Tuberculosis (TB) remains a leading infectious disease globally, accounting for approximately 10 million new cases and 1.6 million deaths annually. Manual chest X-ray (CXR) interpretation, still the primary TB screening tool in many low- and middle-income countries, is frequently compromised by high subjectivity and inter-observer variability ranging from 15% to 25%. This study addresses these diagnostic limitations through a comprehensive comparative analysis of six deep learning architectures for automated lung segmentation classical U-Net, U-Net+ResNet, RNGU-Net, Attention U-Net, EfficientNetV2-UNet, and MobileNetV2-UNet using the Shenzhen Hospital CXR dataset (n=662) and 5-fold cross-validation, with more than 30 performance metrics evaluated for clinical reliability and computational efficiency, together with statistical significance testing. Attention U-Net achieved the most balanced performance, with 97.42% accuracy (95% CI: 97.18–97.66%), 93.02% recall (95% CI: 92.74–93.30%), and 94.66% F1-Score (95% CI: 94.42–94.90%), while requiring only moderate computational resources (11.36 minutes training time, 5.3M parameters). Its false negative rate of 6.98% - the lowest among all evaluated models indicates a low likelihood of missed lung-region segmentation, an outcome of particular importance in TB screening, where missed lesions carry severe clinical consequences. Threshold sensitivity analysis across the 0.3–0.7 range confirmed robust performance, with variation below 0.6%. Uncertainty quantification further revealed that high-confidence predictions (≥ 0.90) correspond to a clinician override rate of only 1.7%, indicating strong potential clinician acceptance. These findings provide evidence-based, statistically grounded recommendations for deploying AI-assisted TB lung segmentation in clinical decision support systems, particularly in resource-limited healthcare settings, and offer preliminary evidence to inform the development and staged deployment of such tools through calibrated model confidence and robust performance across challenging anatomical presentations.