18F-fluorodeoxyglucose PET/computed tomography–derived radiomics and 2.5D deep learning for noninvasive prediction of lymph node metastasis in pancreatic ductal adenocarcinoma
Jiyuan Jin, Dongxue Wang, Furui Duan, Yunfei Cai, Yu Zhao, Yong Wan, Wei Yuan, Huazhen Liu, Hongquan Li, Ping LiBackground
Accurate preoperative prediction of lymph node metastasis (LNM) guides treatment and prognostic stratification for pancreatic ductal adenocarcinoma (PDAC). This study established an interpretable PET/computed tomography (CT)–based artificial intelligence model combining clinical indicators, radiomic and 2.5D deep learning features to predict preoperative LNM.
Methods
A total of 173 pathologically confirmed PDAC patients with preoperative 18 F-fluorodeoxyglucose PET/CT scans were retrospectively enrolled and randomly split into training and test sets at a 7 : 3 ratio. Clinical indicators, PET/CT radiomic features and ResNet50-based 2.5D deep learning features were extracted. Feature selection was implemented via, Spearman correlation analysis and logistic regression. Seven machine learning algorithms were built and compared: support vector machine, logistic regression, random forest, extra trees, gradient boosting machine (GBM), k-nearest neighbors and linear discriminant analysis (LDA). Model performance was assessed by receiver operating characteristics curve, DeLong test, calibration curve, Hosmer–Lemeshow test and decision curve analysis, and Shapley Additive Explanations (SHAP) were used for model interpretability.
Results
Among 173 patients, 105 were LNM-positive and 68 were LNM-negative. The GBM model achieved the highest area under the curve (AUC) of 0.970 in the training set with suspected overfitting. On the independent test set, LDA yielded optimal performance (AUC = 0.871, accuracy = 0.792, sensitivity = 0.719, specificity = 0.905,
Conclusion
This interpretable PET/CT-based model provides a noninvasive tool for individualized preoperative LNM risk assessment in PDAC.