Deep learning reads serial chest CT scans to forecast MDR-TB treatment outcomes

Researchers in China have built a deep learning system that combines chest CT scans taken at three points during therapy to predict how patients with multidrug-resistant tuberculosis will fare, according to a retrospective multicentre study published in BioData Mining on 21 September 2026.

The study, led by Jie Li and colleagues with Maoshui Wang among the senior authors, enrolled 265 patients with multidrug-resistant tuberculosis (MDR-TB) treated at three hospitals in China. Each patient underwent chest CT imaging at baseline, at two months, and at six months after the start of treatment, giving the models a longitudinal view of how lung disease evolved during therapy.

MDR-TB remains a serious global health problem, with persistently low treatment success rates despite prolonged, individualized regimens. Traditional clinical and microbiological predictors, such as sputum culture status, show limited performance in forecasting treatment outcomes, which motivated the search for imaging-based risk tools.

To build the system, the team extracted both handcrafted radiomic features and deep learning-derived features from manually segmented lesions and from whole-lung regions. They then combined different radiomics strategies with various temporal modelling approaches, producing 64 candidate predictive models in total.

The best-performing model, named MTBTOPS, was selected on the basis of its performance in internal and external validation cohorts. Performance was assessed using classification metrics, receiver operating characteristic curves, calibration curves, and decision curve analysis.

MTBTOPS achieved superior predictive accuracy in both a binary task, distinguishing favourable from unfavourable outcomes, and a multiclass task separating treatment success, treatment completion, and treatment failure. In the external validation cohort, the best-performing models yielded AUCs of 0.838 for MTBTOPS-FU and 0.849 for MTBTOPS-SCF, significantly outperforming clinical and microbiological indicators such as six-month sputum culture conversion.

The analysis also showed that integrating the three time points of CT data with GRU-based temporal modelling significantly improved predictive performance compared with static imaging or clinical variables alone. In other words, tracking how lesions changed over the first six months of therapy carried predictive information that no single scan or clinical variable provided.

The authors conclude that the study demonstrates the potential of longitudinal CT-based deep learning for non-invasive risk stratification of MDR-TB treatment outcomes at the six-month treatment assessment point. By enabling individualized risk stratification and timely intervention, the MTBTOPS framework may inform precision treatment strategies and resource allocation in high-burden settings. As a retrospective study of 265 patients from three Chinese hospitals, the findings will require further validation before any clinical deployment, and the published version is a peer-reviewed accepted manuscript shared early, subject to final edits.

Subject of Research: Deep learning on longitudinal chest CT imaging to predict treatment outcomes in multidrug-resistant tuberculosis

Article Title: Deep learning on longitudinal chest CT predicts treatment outcomes in multidrug-resistant tuberculosis: a multicentre retrospective study

Article References: Li, J., Zhang, J., Xie, T., Xu, M., Wang, X., Zhou, Y., Zhang, Q., Kong, X., Qu, H., & Wang, M. (2026). Deep learning on longitudinal chest CT predicts treatment outcomes in multidrug-resistant tuberculosis: a multicentre retrospective study. BioData Mining. https://doi.org/10.1186/s13040-026-00603-8

Image Credits: AI Generated

DOI: 10.1186/s13040-026-00603-8

Keywords: multidrug-resistant tuberculosis, deep learning, chest CT, radiomics, treatment outcomes, longitudinal imaging, GRU, risk stratification, multicentre retrospective study, sputum culture conversion, BioData Mining, Deep