Abstract
BACKGROUND: In non-small cell lung cancer (NSCLC), accurate lymph node staging is vital for prognosis and treatment planning. However, positron emission tomography/computed tomography (PET/CT) is limited by false positives, and morphology-based criteria lack reliability. This study aimed to develop and validate a PET/CT-based deep learning radiomics (DLR) approach to distinguish benign from malignant lymph nodes. METHODS: A total of 217 hypermetabolic lymph nodes from 185 NSCLC patients were retrospectively analyzed. Radiomics and DenseNet121-based deep network features were extracted from PET/CT images. Clinical and imaging variables were selected using logistic regression (LR), correlation analysis, and recursive feature elimination (RFE). Nine machine learning models were trained and externally validated; diagnostic performance was evaluated by the area under the receiver operating characteristic (ROC) curve (AUC), accuracy, sensitivity, specificity, and F1 score. RESULTS: Our study demonstrated that the artificial neural network (ANN) and extra trees (ET) models exhibited superior diagnostic performance in identifying suspected malignant lymph nodes in NSCLC patients. Specifically, the ANN achieved an AUC of 0.865, sensitivity of 66.7%, and accuracy of 82.0% on the test set, while the ET model performed best in the external validation set with an AUC of 0.865, sensitivity of 76.9%, and accuracy of 80.4%. Tumor location, lymph node long-to-short axis (L/S) ratio, and bilateral hilar (18)F-fluorodeoxyglucose (FDG) uptake were significant predictors of nodal status. Correlation analysis showed that deep learning and radiomics features are complementary, suggesting their integration can significantly improve lung cancer diagnostic accuracy. CONCLUSIONS: The PET/CT-based DLR model accurately differentiates benign from malignant lymph nodes, outperforming conventional methods. Combining DenseNet121-derived features with radiomics improves staging accuracy and aids personalized treatment planning.