Abstract:

Objective: To develop an effective imaging-based prediction tool for reducing overmedication in pulmonary tuberculosis patients. Methods: We collected longitudinal chest CT imaging data from 279 pulmonary tuberculosis patients who received treatment at the First People’s Hospital of Kashgar between January 2020 and January 2024. The dataset included both pre-treatment and post-treatment CT scans. A total of 3386 radiomic features were extracted to construct five distinct predictive models: (1) a pre-treatment CT-based model, (2) a post-treatment CT-based model, (3) a combined pre- and post-treatment CT radiomics model, (4) a delta radiomics model based on temporal changes in imaging features between pre- and post-treatment scans, (5) a deep learning model utilizing ResNet18 architecture incorporating both pre- and post-treatment CT images. Model performance was evaluated using metrics including the area under the receiver operating characteristic curve (AUC). Additionally, SHapley Additive exPlanations (SHAP) and Gradient-weighted Class Activation Mapping (Grad-CAM) techniques were employed for model interpretation and visualization. Results: The combined model utilizing both pre- and post-treatment CT data achieved AUCs of 0.845 and 0.770 in the training and test sets, respectively. The deep learning model demonstrated superior performance with AUCs of 0.883 and 0.858 in the training and validation sets, respectively. Conclusion: The deep learning model based on longitudinal chest CT imaging demonstrated robust performance in predicting tuberculosis treatment outcomes, offering valuable support for personalized treatment strategies and resource optimization.

Key words: Tuberculosis, pulmonary, Treatment outcome, Prediction, Artificial intelligence, Radiography