Non-invasive Prediction of Lymph Node Metastasis in NSCLC Using Clinical, Radiomics, and Deep Learning Features From 18F-FDG PET/CT Based on Interpretable Machine Learning

Purpose

This study aimed to develop and evaluate a machine learning model combining clinical, radiomics, and deep learning features derived from PET/CT imaging to predict lymph node metastasis (LNM) in patients with non-small cell lung cancer (NSCLC). The model's interpretability was enhanced using Shapley additive explanations (SHAP).

Methods

A total of 248 NSCLC patients who underwent preoperative PET/CT scans were included and divided into training, test, and external validation sets. Radiomics features were extracted from segmented tumor regions on PET/CT images, and deep learning features were generated using the ResNet50 architecture. Feature selection was performed using minimum-redundancy maximum-relevance (mRMR), and the least absolute shrinkage and selection operator (LASSO) algorithm. Four models—clinical, radiomics, deep learning radiomics (DL_radiomics), and combined model—were constructed using the XGBoost algorithm and evaluated based on diagnostic performance metrics, including area under the receiver operating characteristic curve (AUC), accuracy, F1 score, sensitivity, and specificity. Shapley Additive exPlanations (SHAP) was used for model interpretability.

Results

The combined model achieved the highest AUC in the test set (AUC = 0.853), outperforming the clinical (AUC = 0.758), radiomics (AUC = 0.831), and DL_radiomics (AUC = 0.834) models. Decision curve analysis (DCA) demonstrated that the combined model offered greater clinical net benefits. SHAP was used for global interpretation, and the summary plot indicated that the features ct_original_glrlm_LongRunHighGrayLevelEmphasis, and pet_gradient_glcm_lmc1 were the most important for the model’s predictions.

Conclusion

The combined model, combining clinical, radiomics, and deep learning features from PET/CT, significantly improved the accuracy of LNM prediction in NSCLC patients. SHAP-based interpretability provided valuable insights into the model's decision-making process, enhancing its potential clinical application for preoperative decision-making in NSCLC.