Differentiation of pulmonary tuberculosis from non-tuberculous solid lung lesions using radiomics and clinical-semantic features on contrast-enhanced CT

This study evaluated the feasibility of a combined model integrating radiomics and clinical-semantic features on contrast-enhanced CT for differentiating pulmonary tuberculosis (PTB) from non-tuberculous solid lung lesions. A total of 900 patients enrolled before October 2016 were randomly partitioned 3:1 into training and internal validation sets, while patients recruited through October 2017 formed a temporal validation set. Clinical-semantic features were selected via univariate and multivariate analysis, while radiomics features were identified using ANOVA, Spearman correlation, and LASSO regression; binary logistic regression was used to construct three models. The combined model achieved average precision (AP) of 0.91, 0.85, and 0.62 in the training, internal validation, and temporal validation sets, respectively, outperforming both the clinical-semantic model (AP: 0.64/0.61/0.41) and the radiomics-only model (AP: 0.88/0.82/0.45). Decision curve analysis further confirmed the combined model's superior net benefit across threshold probabilities, supporting its potential for non-invasive CT-based differentiation of PTB. Jaeyoun Yi (이재연) of Coreline Soft is listed as a co-author on this study.