ObjectiveTo build a radiomics model based on preoperative B mode ultrasound (BMUS) and shear wave elastography (SWE) characteristics for predicting clinically relevant postoperative pancreatic fistula (CR-POPF) following radical pancreatectomy.MethodsA total of 138 patients who underwent radical surgery for pancreatic cancer at Zhongshan Hospital Affiliated to Fudan University from February 2021 to February 2022 and received pancreatic ultrasound examination and ultrasound SWE elasticity assessment within one week before surgery were included in this retrospective study. There were 60 cases of pancreaticoduodenectomy and 78 cases of distal pancreatectomy. The patients were divided into CR-POPF positive group (39 cases) and CR-POPF negative group (99 cases). Based on preoperative BMUS images and quantitative shear wave velocity (SWV) values of the pancreatic lesion and the surrounding normal pancreatic parenchyma in ultrasound elastography, the radiomics features were extracted using the minimum redundancy maximum relevance algorithm (mRMR) and the modified logistic regression algorithm with the selection operator (LASSO). A radiomics model for prediction of CR-POPF was constructed through multivariate logistic regression analysis. The diagnostic efficacy of the clinical prediction model (CM) based on potential clinical risk factors of CR-POPF, the radiomics model (lesRS) based on two-dimensional gray-scale ultrasound images of pancreatic lesions, the radiomics model (panRS) based on two-dimensional gray-scale ultrasound images of normal pancreatic parenchyma, the radiomics model (LT-RS) based on ultrasound images of pancreatic lesions and normal parenchyma, and the CR-POPF prediction model (C-LT-RS) combining clinical parameters and radiomics features were evaluated and compared through ROC curve analysis.ResultsThe SWV value of normal pancreatic tissue in the CR-POPF positive group was significantly lower than that in the CR-POPF negative group ([0.94±0.28] m/s vs. [1.18±0.50] m/s, P<0.001). The SWV ratio of the lesion center to normal pancreatic tissue in the CR-POPF positive group was significantly higher than that in the CR-POPF negative group ([1.60±0.71] m/s vs. [1.29±0.57] m/s, P=0.016). The results of univariate logistic regression analysis showed that the quantitative SWV value of normal pancreatic tissue by ultrasound elastography was significantly negatively correlated with the occurrence of CR-POPF after surgery (OR=－1.686, 95%CI: －3.288 to －0.085, P=0.039). The results of multivariate logistic regression analysis showed that the SWV value of normal pancreatic tissue was an independent risk factor for the occurrence of CR-POPF after surgery (OR=－1.892, 95%CI: －3.956 to 0.115, P=0.021). In the training set of CM, the accuracy rate of preoperative prediction of CR-POPF by the SWV value of normal pancreatic tissue was 50.54%, with a sensitivity of 0.885 and a specificity of 0.358; in the test set, the accuracy rate was 68.29%, with a sensitivity of 0.833 and a specificity of 0.621. In the training set of lesRS, the accuracy rate of preoperative prediction of CR-POPF was 67.74%, with a sensitivity of 0.615 and a specificity of 0.702; in the test set, the accuracy rate was 70.73%, with a sensitivity of 0.833 and a specificity of 0.966. In the training set of panRS, the accuracy rate of preoperative prediction of CR-POPF was 68.65%, with a sensitivity of 0.525 and a specificity of 0.751; in the test set, the accuracy rate was 70.36%, with a sensitivity of 0.833 and a specificity of 0.602. In the training set of LT-RS, the accuracy rate of preoperative prediction of CR-POPF was 69.89%, with a sensitivity of 0.577 and a specificity of 0.746; in the test set, the accuracy rate was 68.29%, with a sensitivity of 0.734 and a specificity of 0.621. In the training set and test set of C-LT-RS, the accuracy rate of preoperative prediction of CR-POPF was 62.37% and 80.49%, respectively, with a sensitivity of 0.962 and 0.667, and a specificity of 0.493 and 0.862. In the training set and test set of C-LT-RS, the area under the ROC (AUROC) for predicting the occurrence of CR-POPF after surgery was higher than those of CM, lesRS, panRS and LT-RS.ConclusionModified prediction model based on BMUS radiomics characteristics and quantitative SWE values of pancreas displays potential value in non-invasive and quantitative assessment of CR-POPF risk before radical surgery for pancreatic cancer.