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

Multidrug-resistant tuberculosis (MDR-TB) poses a significant global health threat, with persistently low treatment success rates despite prolonged, individualized therapies. Traditional clinical and microbiological predictors show limited performance in forecasting treatment outcomes. This study aimed to develop a deep learning-based system that integrates longitudinal chest CT imaging to predict MDR-TB treatment outcomes. We conducted a retrospective multicenter cohort study involving 265 patients with multidrug-resistant tuberculosis from three hospitals in China. Each patient underwent chest CT imaging at baseline, at two months, and at six months after the initiation of treatment. Handcrafted radiomic features and deep learning-derived features were extracted from manually segmented lesions and whole-lung regions. A total of 64 predictive models were developed by combining different radiomics strategies with various temporal modeling approaches. The optimal model, named MTBTOPS, was selected based on its performance in the internal and external validation cohorts. Model performance was assessed using classification metrics, receiver operating characteristic curves, calibration curves, and decision curve analysis. MTBTOPS achieved superior predictive accuracy in both binary (favorable vs. unfavorable outcomes) and multiclass (treatment success, treatment completion, and treatment failure) tasks. Within the external validation cohort, the best-performing models yielded AUCs of 0.838 (MTBTOPS-FU) and 0.849 (MTBTOPS-SCF), significantly outperforming clinical and microbiological indicators, such as 6-month sputum culture conversion. The integration of three-time-point CT data and GRU-based temporal modeling significantly improved predictive performance compared with static imaging or clinical variables alone. This study demonstrates the potential of longitudinal CT-based deep learning for non-invasive risk stratification of MDR-TB treatment outcomes at the 6-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.