Convolutional neural networks for automated tuberculosis detection in chest X-ray screening programs.

Tuberculosis remains a leading cause of morbidity worldwide, particularly in resource-constrained regions where radiological expertise is scarce. Chest X-ray screening is a cost-effective method for population-level TB detection, yet subtle lesions and class imbalance challenge automated analysis. This paper proposes a novel framework that combines generative data augmentation with a shallow convolutional neural network to improve accuracy and interpretability. The publicly available TB Chest X-ray dataset contains 4200 images, of which only 700 are TB-positive; class imbalance and data scarcity hinder conventional deep networks. To mitigate these issues, an auxiliary-classifier generative adversarial network synthesizes realistic TB images and blends them with classical augmentations. A lightweight CNN architecture with four convolutional layers is trained on the balanced data, and interpretability is achieved using class activation maps and LIME. Five-fold cross-validation shows that the proposed method achieves 95 % accuracy and 92.5 % recall, improving recall by 6.9 % and F1 by 5.4 % over a baseline shallow network. Qualitative analyses demonstrate that the attention maps focus on pathological regions rather than background structures. This work illustrates how generative augmentation and explainable models can enhance TB screening accuracy without sacrificing transparency.