Deep Learning for Effective Tuberculosis Detection from Chest Radiographs: A Comparative Study

This paper presents a comparative study to assess the performance of the visual transformer (ViT) alongside several convolutional-based models in distinguishing tuberculosis (TB) from healthy and various other lung conditions using chest radiographs. Seven deep learning models including Xception, InceptionNetV3, VGG16, ResNet50, DenseNet201, EfficientNetV2 and the ViT model are selected in this study for comparison on the TBX11K dataset. Empirical results revealed the superiority of established convolutional models, especially ResNet50, over the ViT model in TB recognition by reaching 99.22%, 98.27%, 99.08%, and 99.62% for accuracy, recall, precision and specificity, respectively. This study contributes to the medical field by emphasizing the crucial role of deep learning (DL) model selection in medical image analysis for automatic TB diagnosis from chest x-rays.