Tuberculosis (TB) causes over 1.5 million deaths annually, disproportionately affecting developing nations where trained radiologists are critically scarce. Chest X-ray screening is the most accessible diagnostic tool in such settings, yet accurate interpretation demands specialist expertise that rural populations rarely have access to. This paper proposes a hybrid deep learning architecture combining DenseNet121 — motivated by its proven radiologist-level performance in the CheXNet study on chest X-ray classification — as a convolutional feature extractor with custom Transformer encoder blocks for global context modelling. Training was performed on a balanced multi-source dataset of 2,388 chest X-rays (1,194 TB and 1,194 Normal) merged from the Shenzhen, Montgomery, and TBX11K collections. Contrast Limited Adaptive Histogram Equalisation (CLAHE) in the LAB colour space standardised contrast variation across imaging sources. Two model versions were evaluated under identical experimental conditions: V1.3 using ResNet50 and V2.0 using DenseNet121 as the CNN backbone. The final V2.0 model achieved 95.54% accuracy and 0.9914 AUC on a held-out test set of 359 images, reducing false negatives from 16 to 12 compared to the V1.3 baseline. Gradient-weighted Class Activation Mapping (Grad-CAM) confirmed that the model attends to clinically consistent anatomical regions. The system is deployed as an interpretable Flask web application with uncertainty-aware predictions for clinical demonstration in low-resource healthcare settings.
Mohammad et al. (Sat,) studied this question.