Development and validation of a parsimonious clinical-texture signature for differentiating primary lung cancer from pulmonary metastases on ¹⁸F-FDG PET/CT: a retrospective diagnostic study

Distinguishing primary lung cancer (PLC) from pulmonary metastases (PM) is a pivotal yet elusive challenge in thoracic oncology, directly dictating surgical strategy from lobectomy to lung-sparing mastectomy. Conventional CT morphology and standard metabolic parameters (SUVmax) frequently lack the specificity required for accurate characterization, particularly in solitary nodules, leading to potential misclassification and treatment delays. We developed and internally validated a machine learning framework integrating ¹⁸F-FDG PET/CT radiomics with clinical variables to non-invasively differentiate PLC from PM. In this retrospective study, 161 multimodal features were extracted from 79 pathologically confirmed lesions (30 PLC, 49 PM). A rigorous two-stage feature selection pipeline (Spearman correlation + Least Absolute Shrinkage and Selection Operator (LASSO) regression) identified the most predictive signature. Ten supervised classifiers were trained and evaluated on the held-out internal test set, with comprehensive assessment via AUC, calibration curves, and Decision Curve Analysis (DCA). The final parsimonious signature comprised only four features: age and three higher-order texture features (GLSZM), notably excluding conventional SUV metrics which failed to retain predictive value. Logistic Regression achieved superior discrimination (AUC = 0.90), while XGBoost demonstrated optimal calibration (Brier score = 0.119) and the highest net clinical benefit across relevant threshold probabilities. Feature selection consistently improved model stability and interpretability by reducing redundant variables while preserving diagnostic performance across the evaluated machine-learning algorithms. This study validates a concise, biologically meaningful radiomics-clinical model that significantly outperforms conventional metabolic metrics in distinguishing PLC from PM. By delivering accurate, well-calibrated predictions, this framework offers tangible clinical utility for optimizing surgical planning, personalizing treatment decisions, and reducing diagnostic uncertainty in patients with pulmonary lesions.