Title : A deep learning approach to identify active and drug-resistant tuberculosis

Background: Tuberculosis (TB) remains a critical global health threat, with accurate differentiation between active and inactive TB, as well as early detection of drug resistance, being essential for effective treatment and transmission control. Current diagnostic methods are often limited by time, cost, or generalizability across diverse clinical settings.

Methods: This study employs artificial intelligence methods to diagnose and differentiate active and inactive pulmonary tuberculosis, as well as drug-resistant and non-drug-resistant tuberculosis. We would like to introduces two integrated deep learning frameworks developed to address these challenges. The first model, DMTB-Net, employs a Multi-Scale Convolutional Fusion Block (MSCB) and a Gradient Reversal Unit (GRU) to enhance feature extraction and domain generalization in multi-center CT imaging data. The second model, DFAMIL (Dual Path Fusion Attention Multi-Instance Learning), incorporates attention mechanisms and multi-instance learning to automatically localize discriminative regions in 3D images for predicting drug-resistant TB. 

Results: DMTB-Net was validated on a multi-center dataset from three hospitals in Guizhou Province, achieving an AUC of up to 0.924 across external test sets, demonstrating robust generalization. DFAMIL attained an ACC of 85.71% and AUC of 88.46% on internal tests, and maintained strong performance on external validation (ACC 82.56%, AUC 84.00%). Both models show significant improvements over conventional approaches in accuracy and adaptability. 

Conclusion: These AI-driven models offer scalable, efficient, and accurate tools for supporting clinical decisions in TB diagnosis and drug resistance prediction. Their ability to generalize across diverse datasets highlights the potential for integration into real-world diagnostic workflows, contributing to personalized treatment and strengthened public health responses.

Dr. Rongpin Wang is an radiologist specializing in chest imaging, and artificial intelligence-based research for infectious diseases and tumor. He earned his MD from Zunyi Medical College (1995), PhD from Southern Medical University (2010), and completed postdoctoral training at Boston Children’s Hospital, Harvard Medical School (2015). He has 30 years of clinical experience in CT and MRI diagnosis. He is currently the Director of Department of Radiology (2015) at Guizhou Provincial People’s Hospital, China. In the past five years, he has led 8 scientific research projects, published over 30 SCI papers.