Objective To investigate the value of PET/CT radiomics based on the primary tumor and peritumoral region on ¹⁸F-fluorodeoxyglucose (¹⁸F-FDG) PET/CT in predicting lymph node metastasis in lung adenocarcinoma, and to compare the predictive performance of different machine learning algorithms.

Methods A total of 224 patients with lung adenocarcinoma from our hospital were included as the internal dataset, and an additional 85 patients were selected from the public Cancer Imaging Archive (TCIA) database as an independent external validation set. PET/CT radiomic features were extracted from the primary tumor and the 5-mm peritumoral region. Nine core features, including six from the tumor region and three from the peritumoral region, were selected using the intraclass correlation coefficient (ICC), Spearman rank correlation analysis, and least absolute shrinkage and selection operator (LASSO) regression. For each of the tumor-region, peritumoral-region, and combined tumor-peritumoral feature sets, five models were constructed using logistic regression (LR), support vector machine (SVM), random forest (RF), K-nearest neighbors (KNN), and extreme gradient boosting (XGBoost). Receiver operating characteristic (ROC) curves and decision curve analysis (DCA) were used to compare and evaluate the predictive performance and clinical benefit of the models.

Results In the internal test set, the XGBoost model based on the combined feature set achieved the highest area under the curve (AUC), at 0.899 (95% CI: 0.828-0.970). After Bonferroni correction, the difference was statistically significant only between the XGBoost and KNN models (P = 0.009), while no statistically significant differences were observed between the XGBoost model and the SVM, LR, or RF models (all P ＞ 0.013). In the external validation cohort, the XGBoost model based on the combined feature set achieved an AUC of 0.887 (95% CI: 0.825-0.965) and remained superior to the tumor-region model (AUC = 0.810) and the peritumoral-region model (AUC = 0.779) after Bonferroni correction, with both differences being statistically significant (both P ＜ 0.017). DCA showed that the XGBoost model based on the combined feature set yielded a higher clinical net benefit across all cohorts.

Conclusions Integrating PET/CT radiomic features from the primary tumor and peritumoral region can improve the predictive performance for lymph node metastasis. The XGBoost model based on the combined feature set demonstrates good accuracy and generalizability and can provide decision support for clinical preoperative assessment.