Preoperative Extrapancreatic Extension Prediction in Patients with Pancreatic Cancer Using Multiparameter MRI and Machine Learning-Based Radiomics Model

Rationale and Objectives Pancreatic cancer is a common malignant tumor with a dismal prognosis. Preoperative differentiation of extrapancreatic extension (EPE) based on radiomics will facilitate treatment decision-making. Materials and Methods This research retrospectively recruited 156 patients from two medical centers. 122 patients from the center A were randomly divided into the training set and the internal test set in a 4:1 ratio. Additionally, 34 patients from the center B served as the external test set. Radiomics features were extracted from multiparametric MRI (MP-MRI), containing axial T2 weighted imaging (T2WI), diffusion weighted imaging (DWI), and dynamic contrast enhancement (DCE) sequences. The three-step method was used for feature extraction: SelecteKBest, least absolute shrinkage and selection operator (LASSO) algorithm, and recursive feature elimination based on random forest (RFE-RF). The model was constructed using six classifiers based on machine learning, and the classifier with the best performance was chosen. Finally, clinical factors associated with EPE were incorporated into the combined model. Results The classifier with the best performance was XGBoost, which obtained area under curve (AUC) values of 0.853 and 0.848 in the internal and external test sets, respectively. Through SelectKBest, the most relevant clinical factor for EPE was determined to be platelet, which was then added to the combined model, yielding AUC values of 0.880 and 0.848 in the internal and external test sets, respectively. Conclusion Radiomics models had the potential to noninvasively and accurately predict EPE before surgery. Additionally, it would add value to personalized precision treatment. Pancreatic cancer is a common malignant tumor with a dismal prognosis. Preoperative differentiation of extrapancreatic extension (EPE) based on radiomics will facilitate treatment decision-making. This research retrospectively recruited 156 patients from two medical centers. 122 patients from the center A were randomly divided into the training set and the internal test set in a 4:1 ratio. Additionally, 34 patients from the center B served as the external test set. Radiomics features were extracted from multiparametric MRI (MP-MRI), containing axial T2 weighted imaging (T2WI), diffusion weighted imaging (DWI), and dynamic contrast enhancement (DCE) sequences. The three-step method was used for feature extraction: SelecteKBest, least absolute shrinkage and selection operator (LASSO) algorithm, and recursive feature elimination based on random forest (RFE-RF). The model was constructed using six classifiers based on machine learning, and the classifier with the best performance was chosen. Finally, clinical factors associated with EPE were incorporated into the combined model. The classifier with the best performance was XGBoost, which obtained area under curve (AUC) values of 0.853 and 0.848 in the internal and external test sets, respectively. Through SelectKBest, the most relevant clinical factor for EPE was determined to be platelet, which was then added to the combined model, yielding AUC values of 0.880 and 0.848 in the internal and external test sets, respectively. Radiomics models had the potential to noninvasively and accurately predict EPE before surgery. Additionally, it would add value to personalized precision treatment.