Detection of Tuberculosis Using Deep Learning Models

Tuberculosis is a major international concern, with approximately 10 million new emerging cases and 1 million deaths reported annually. Developing a semiautomated system for identifying TB through scan images is crucial for improving diagnostic accuracy. Current deep learning approaches face challenges such as lowquality datasets, limited model generalization, and high computational costs. Our paper addresses these issues by evaluating three deep learning models: SqueezeNet, ResNet, and DenseNet-121. SqueezeNet and ResNet were trained on 4,200 trillion bytes images of chest X-ray from the TB Chest X-ray Database, achieving test accuracies of 90% and 83%, respectively. DenseNet-121 was trained on the Shenzhen Chest X-ray Dataset, achieving an accuracy of 85%. After conducting a comparative study and implementing these models, we concluded that SqueezeNet provides the highest accuracy at 90%. Therefore, SqueezeNet is the most suitable model for accurately detecting tuberculosis and can be utilized for this purpose moving forward.rculosis and can be utilized for this purpose moving forward.