Machine Learning Methods Based on CT Features Differentiate G1/G2 From G3 Pancreatic Neuroendocrine Tumors

Abstract

To identify CT features for distinguishing grade 1 (G1)/grade 2 (G2) from grade 3 (G3) pancreatic neuroendocrine tumors (PNETs) using different machine learning (ML) methods. A total of 147 patients with 155 lesions confirmed by pathology were retrospectively included. Clinical-demographic and radiological CT features was collected. The entire cohort was separated into training and validation groups at a 7:3 ratio. Least absolute shrinkage and selection operator (LASSO) algorithm and principal component analysis (PCA) were used to select features. Three ML methods, namely logistic regression (LR), support vector machine (SVM), and K-nearest neighbor (KNN) were used to build a differential model. Receiver operating characteristic (ROC) curves and precision-recall curves for each ML method were generated. The area under the curve (AUC), accuracy rate, sensitivity, and specificity were calculated. G3 PNETs were more likely to present with invasive behaviors and lower enhancement than G1/G2 PNETs. The LR classifier yielded the highest AUC of 0.964 (95% confidence interval [CI]: 0.930, 0.972), with 95.4% accuracy rate, 95.7% sensitivity, and 92.9% specificity, followed by SVM (AUC: 0.957) and KNN (AUC: 0.893) in the training group. In the validation group, the SVM classier reached the highest AUC of 0.952 (95% CI: 0.860, 0.981), with 91.5% accuracy rate, 97.3% sensitivity, and 70% specificity, followed by LR (AUC: 0.949) and KNN (AUC: 0.923). The LR and SVM classifiers had the best performance in the training group and validation group, respectively. ML method could be helpful in differentiating between G1/G2 and G3 PNETs.