The Application Of Quantitative CT Features In Predicting Postoperative Pancreatic Fistula In Patients With Pancreaticoduodenectomy

Objects: The purpose of this study is to train and test a model that can accurately predict the(postoperative pancreatic fistula,POPF)of pancreatic fistula after pancreatic surgery using preoperative enhanced CT images by combining CT imaging and machine learning,and compare it with the pancreatic fistula scoring system.Methods: 127 consecutive cases were analyzed retrospectively and randomly divided into training cohort(n=88)and test cohort(n=39).First,the 3D ROI was manually drawn on the preoperative enhanced CT pancreatic parenchyma images of the subjects in the training cohort,and the CT radiomics features were extracted,then the features that passed the consistency test(intra-group correlation coefficient,ICC > 0.75)were screened and dimensionally reduced,and finally four supervised machine learning models,Random forest(random forest,RF),support vector machine(support vector machine,SVM),XBGOOST,logical regression(logistic regression,LR),were constructed of these selected features.We choose the optimal model,establish a scoring system through it,and compare it with FRS.Results: The performance of RF model was the best in all machine learning models(3-fold cross-validation results showed that the AUC value of RF model was all greater than 0.85).Moreover,the performance of RF model was better than FRS in both training cohort and test cohort(RF model AUC=0.95,95%CI=0.91-0.97,FRS model AUC=0.74,95%CI=0.65-0.86 in training cohort;RF model AUC=0.89,95%CI=0.81-0.96,FRS model AUC=0.69,95%CI=0.58-0.80 in test cohort).Conclusion: Preoperative enhanced CT combined with machine learning can accurately predict POPF,and its performance is better than that of FRS.