International Journal of Clinical Practice · Published 2026-01-01 · DOI 10.1155/ijcp/8814983
Accurately distinguishing between active and inactive pulmonary tuberculosis is essential for effective treatment and controlling the spread of the disease. Current diagnostic methods rely on manual identification, which is time-consuming and labor-intensive. Deep learning has emerged as a powerful tool for lung CT image classification, but inconsistencies in the number of slices in 3D medical images caused by different scanning devices and parameters pose challenges. Traditional methods like manual slice selection or interpolation introduce data loss and subjectivity. To address this, we propose HATB-Net, a novel deep learning model for classifying active and inactive pulmonary tuberculosis. HATB-Net integrates two branches—one based on a Transformer and the other on a convolutional neural network (CNN)—which interact and fuse at multiple levels through the branch interaction fusion (BIF) module, mitigating the inherent limitations of convolutional operations. Additionally, we developed the deep slice prioritizer (DSP) module, which adaptively selects diagnostically valuable slices, eliminating the need for manual input dimension adjustments. This reduces computational costs by avoiding irrelevant slice calculations and enhances model interpretability. Our method was evaluated on the Guizhou Provincial People’s Hospital TB dataset (GPPH-TB dataset), which includes 221 active and 298 inactive tuberculosis cases. HATB-Net achieved an AUC of 0.949 ± 0.013 and an accuracy of 0.894 ± 0.019. In addition, in an independent external test conducted at Zhijin County People’s Hospital, the model attained an AUC of 0.901. Furthermore, in additional experiments on the public pneumonia datasets MosMedData and CC-19, HATB-Net achieved an AUC of 0.9364, demonstrating strong classification performance and general applicability to other lung image classification tasks. The HATB-Net model offers a promising tool for assisting radiologists in analyzing lung images and addressing pulmonary disease classification.
Abstract from DOAJ. Public domain (CC0 1.0).
Read the article at the publisher →