Objective: To investigate the ability of preoperative CT-based radiomics signature to be used as prognostic indicators of unresectable pancreatic ductal adenocarcinoma (PDAC) after chemotherapy and develop radiomics-based prediction models.
Methods: Clinicopathological data and radiomics features of 146 pancreatic cancer patients (102 in the training cohort and 44 in the testing cohort) were retrospectively reviewed. Radiomics features were selected by the lasso cox regression model. Multivariate Cox regression analysis was used to establish the radiomics model, clinicopathologic model, and combined model in the training cohort to predict survival of PDAC after chemotherapy, and then verified in the testing cohort.
Results: Of 855 extracted radiomics features in portal venous CT images, 15 most stable features were selected. Multivariate Cox regression analysis identified radiomics model riskscore, diabetes, common hepatic artery invasion, and splenic vein invasion to construct a combined model for predicting overall survival (OS) of unresectable PDAC (with AUC of 0.949 in the validation set; 95% CI 0.889-1.000). While the AUC of the radiomics model and clinicopathologic model were 0.901 (95% CI, 0.812-0.990) and 0.556 (95% CI, 0.382-0.730), respectively.
Conclusion: CT radiomics signature is a powerful predictor associated with OS in patients with unresectable PDAC after chemotherapy. The combined model performed well in predicting survival.