Prediction of active drug-resistant pulmonary tuberculosis based on CT radiomics: construction and validation of independent models and combined models for residual pulmonary parenchyma

Background Drug-resistant tuberculosis (DR-TB) is a severe public health threat and burden worldwide. This study seeks to develop and validate both independent and combined radiomic models using pulmonary cavity (PC), tree-in-bud sign (TIB), total lung lesions (TLL), and residual pulmonary parenchyma (RPP) to evaluate their effectiveness in predicting DR-TB. Methods We recruited 306 confirmed active pulmonary tuberculosis cases from two hospitals, comprising 142 drug-resistant and 164 drug-sensitive cases. Patients were assigned to five training and testing cohorts: PC ( n = 109, 47), TIB ( n = 214, 92), TLL ( n = 214, 92), RPP ( n = 214, 92), and their combination ( n = 109, 47). Radiomic features were extracted using variance thresholding, K-best, and LASSO techniques. We developed four separate radiomic models with random forest (RF) for DR-TB prediction and created a combined model integrating all features from the four indicators. Model performance was validated using ROC curves. Results We extracted 10, 2, 10, 3, and 9 radiomic features from PC, TIB, TLL, RPP, and the combined model, respectively. The combined model achieved AUC values of 0.886 (95% CI: 0.827–0.945) in the training set and 0.865 (95% CI: 0.764–0.966) in the testing set. It slightly surpassed the PC model in the training set (0.886 vs. 0.850, p < 0.05) and was comparable in the testing set (0.865 vs. 0.850, p > 0.05). The combined model showed similar performance to the TIB, TLL, and RPP models in both sets ( p > 0.05). Conclusion The newly defined and developed RPP model and the combined model demonstrated robust performance in identifying DR-TB, highlighting the potential of CT-based radiomic models as effective non-invasive tools for DR-TB prediction.