Lightweight deep learning for chest X-rays tuberculosis detection

Early tuberculosis detection is vital for reducing transmission, yet diagnostic resources are often limited in low-resource settings. This study evaluates lightweight deep learning models for detecting tuberculosis from chest X-rays, including strictly lightweight networks, parameter-efficient classical models, and a TB-specific design. Results show that LightTBNet achieved the highest accuracy (0.9125) and sensitivity (0.9241), while MobileNetV3-Small offered the best trade-off, combining competitive accuracy with the fastest inference speed. Transfer learning improved performance across all models, with the largest gains for weaker baselines. These findings underscore the potential of lightweight architectures, particularly LightTBNet and MobileNetV3-Small, to enable efficient, reliable, and accessible TB screening in real-world healthcare environments.