18F-FDG PET/CT radiomics for differentiating peripheral tuberculosis from lung cancer: development and validation

This study aimed to develop and validate a multidimensional radiomics signature utilizing fluorine-18 fluorodeoxyglucose positron emission tomography/computed tomography (¹⁸F-FDG PET/CT) and machine learning to accurately differentiate peripheral pulmonary tuberculosis from primary peripheral lung cancer. This single-center retrospective study included 346 patients (144 peripheral pulmonary tuberculosis, 202 lung cancer). Radiomic features were extracted from pretreatment PET/CT scans. Patients were randomly split 7:3 into training ( n = 242) and internal validation ( n = 104) sets. Due to the short 3-year enrollment period with stable protocols, random splitting was the primary validation method. A temporal split (first 70% vs. most recent 30%) was also performed as a sensitivity analysis to assess temporal robustness. Three classifiers (random forest, support vector machine, gradient boosting machine) were compared. The optimal model was evaluated using the area under the receiver operating characteristic curve (AUC), calibration curves, and decision curve analysis. The final model incorporated six key predictors: age, maximum standardized uptake value (SUVmax), metabolic heterogeneity index (CV%), PET skewness, PET gray-level co-occurrence matrix energy (PET-GLCM Energy), and shape sphericity. All were identified as independent predictors via multivariate logistic regression (all P < 0.05). The random forest model demonstrated superior and stable performance, achieving an AUC of 0.861 (95% CI: 0.805–0.917) in training and 0.833 (95% CI: 0.732–0.934) in internal validation set. Calibration curves indicated good agreement between predictions and observations. Furthermore, decision curve analysis confirmed the model’s clinical net benefit across a wide threshold probability range. A machine learning model based on multidimensional ¹⁸F-FDG PET/CT radiomics features was developed and internally validated. This model demonstrates robust discriminatory efficacy, good calibration, and clinical utility for differentiating peripheral pulmonary tuberculosis from lung cancer.