Abstract: Background　Surufatinib has demonstrated favorable efficacy in advanced pancreatic neuroendocrine tumors (pNETs). However, accurately and noninvasively identifying patients who will benefit from surufatinib remains challenging. Objective　To develop a radiomics-based machine learning (ML) model to noninvasively predict treatment response to surufatinib in patients with pNETs liver metastases and interpret the key imaging features associated with response. Methods　Clinical data and portal venous - phase contrast-enhanced CT images of patients with pNETs liver metastases from seven centers were retrospectively collected. A machine learning (ML) model was constructed using a stacking ensemble algorithm within a nested cross-validation framework. Support vector machine (SVM), random forest (RF), and eXtreme gradient boosting (XGBoost) served as base classifiers, and a multilayer perceptron (MLP) was used as the meta-classifier for fusion to predict response to surufatinib treatment (response defined as ＞15% reduction in the target lesion size). In parallel, a clinical model was built using multivariable logistic regression based on clinical variables and compared with the ML model. Model performance was evaluated using the area under the curve (AUC), specificity, sensitivity, accuracy, calibration curves, and decision curve analysis. Feature coefficients derived from least absolute shrinkage and selection operator (LASSO) regression were analyzed to evaluate the association between radiomic features and treatment response. Results　A total of 102 patients were ultimately included in this study, with 51 patients in the responder group and 51 patients in the non-responder group. The machine learning model constructed using five core radiomics features selected by LASSO regression and frequently retained across Bootstrap resampling iterations achieved a mean AUC of 0.805 , an accuracy of 76.47% and a sensitivity of 84.31%, with overall performance superior to that of the corresponding clinical model (AUC=0.623, accuracy of 68.40%, and sensitivity of 74.62%). In the overall cohort, the model-predicted responder group had a significantly longer progression-free survival (PFS) than the non-responder group (HR=0.573, 95% CI: 0.349 - 0.941, P＜0.05). The LASSO-selected key imaging features suggested that patients with lower pretreatment tumor burden, richer blood supply, and higher tumor homogeneity were more likely to benefit from surufatinib. Conclusion　A CT radiomics - based machine learning model can effectively predict the efficacy of surufatinib in patients with pNETs liver metastases and identify interpretable radiomic biomarkers with prognostic significance.