Ziying Lin1, Bingjun Tang2, Jinxiu Cai1, Xiangpeng Wang3, Changxin Li3, Xiaodong Tian2, Yinmo Yang4, Xiaoying Wang5. 1. Department of Radiology, Peking University First Hospital, Beijing, 100034, China. 2. Department of General Surgery, Peking University First Hospital, Beijing, 100034, China. 3. Beijing Smart Tree Medical Technology Co. Ltd., Beijing, 100011, China. 4. Department of General Surgery, Peking University First Hospital, Beijing, 100034, China. Electronic address: yangyinmo@263.net. 5. Department of Radiology, Peking University First Hospital, Beijing, 100034, China. Electronic address: wangxiaoying@bjmu.edu.cn.
Abstract
OBJECTIVES: To develop a radiomics model and a combined model for preoperative prediction of clinically relevant postoperative pancreatic fistula (CR-POPF) in patients undergoing pancreaticoduodenectomy and to compare the predictive performance of the two models with the traditional Fistula Risk Score system. METHODS: A total of 250 patients who underwent pancreaticoduodenectomy (PD) with preoperative computed tomography (CT) were divided into a training set (n = 175) and validation set (n = 75). The pancreatic area was automatically segmented on the portal venous phase CT images using a 3D U-Net segmentation model. A radiomics model was developed using radiomics features extracted from the volume of interest (VOI) and a combined model was developed using radiomics features, demographic information and radiological features. The FRS was also used to predict POPF. The predictive performance of the prediction models was assessed using receiver operating characteristic (ROC) curves, calibration curves and decision curve analysis (DCA). RESULTS: Eleven and 18 features were extracted for the radiomics model and combined model, respectively. The combined model showed excellent predictive value, with an AUC of 0.871 (95 %CI 0.816,0.926) and 0.869 (95 %CI 0.779,0.958) in the training cohort and validation cohort, respectively. Calibration curves and DCA showed that the combined model outperformed the traditional FRS system and radiomics model. CONCLUSION: The combined model exhibited excellent predictive performance and outperformed the traditional FRS system and radiomics model in the preoperative prediction of CR-POPF.
OBJECTIVES: To develop a radiomics model and a combined model for preoperative prediction of clinically relevant postoperative pancreatic fistula (CR-POPF) in patients undergoing pancreaticoduodenectomy and to compare the predictive performance of the two models with the traditional Fistula Risk Score system. METHODS: A total of 250 patients who underwent pancreaticoduodenectomy (PD) with preoperative computed tomography (CT) were divided into a training set (n = 175) and validation set (n = 75). The pancreatic area was automatically segmented on the portal venous phase CT images using a 3D U-Net segmentation model. A radiomics model was developed using radiomics features extracted from the volume of interest (VOI) and a combined model was developed using radiomics features, demographic information and radiological features. The FRS was also used to predict POPF. The predictive performance of the prediction models was assessed using receiver operating characteristic (ROC) curves, calibration curves and decision curve analysis (DCA). RESULTS: Eleven and 18 features were extracted for the radiomics model and combined model, respectively. The combined model showed excellent predictive value, with an AUC of 0.871 (95 %CI 0.816,0.926) and 0.869 (95 %CI 0.779,0.958) in the training cohort and validation cohort, respectively. Calibration curves and DCA showed that the combined model outperformed the traditional FRS system and radiomics model. CONCLUSION: The combined model exhibited excellent predictive performance and outperformed the traditional FRS system and radiomics model in the preoperative prediction of CR-POPF.