CT And Radiomics For Differentiation Of Serous Cystic Neoplasms And Mucin-producing Pancreatic Cystic Neoplasms

Purposes:Radiomics can reflect the biological nature of neoplasm to a certain extent.Based on CT images,this study created and validated a nomogram model combining imaging predictors and radiomics labels to explore the identification of pancreatic serous neoplasm(SCN)and pancreatic mucinous cystic neoplasm before surgery.Materials and Method:Retrospectively collected imaging data of 122 patients with pancreatic SCN,MCN and IPMN who underwent preoperative imaging examinations from October 2012 to October 2018.First we explore the differences between pancreatic SCN and mucinous cystic neoplasm from general data and CT features(location of tumor,morphology,margins,realm of the capsule,whether there is calcification,whether there is distal pancreatic atrophy,pancreatic duct diameter,whether the tumor communicates with the main pancreatic duct,the number of capsules,the intrasacral separation,the thickness of the capsule wall,Wall nodules,plain scans,and enhancement of solid tumor density).For continous parameters,we used independent sample t test and Mann-Whitney U test analysis.For categorical variables,we apply ?2 test or Fisher exact test analysis.We performed multi-factor logistic regression on the statistically significant features and single-factor logistic regression to evaluate the AUC value,sensitivity,specificity,and accuracy of the model and each feature.A comprehensive analysis selected the best feature as Independent predictors.Then,116 pancreatic SCN,MCN and IPMN patients included in the radiographics analysis were randomly divided into a training group(n=83)and a verification group(n=33)according to 7:3.This study extracts radiomics features from CT images of plain scan,pancreatic parenchymal phase,and portal phase.Then,several radioics labels were identified and radioics scores were calculated.From the perspectives of AUC value,sensitivity,specificity,and accuracy,the prediction performance of each radioics label in the training group and the verification group is evaluated,and the best radioics label is selected by comprehensive analysis.Finally,this study combined general data or CT independent predictors and the best radioics labels to construct radioics nomograms,and used correction curves and decision curves to evaluate the correction effect and clinical value of the nomograms.Result:General data and CT characteristics in univariate analysis revealed gender,number of cysts,diameter of main pancreatic duct,whether the lesion is in communication with the main pancreatic duct,and enhanced scanning of solid part density of the pancreatic parenchymal tumor to identify pancreatic SCN and mucinous cystic tumor There was statistical significance(P values were:P=0.001,P=0.008,P<0.001,P=0.001,P<0.001).In multivariate logistic regression,it was found that the number of capsules and the diameter of the main pancreatic duct were less than 0.05 in the model,which were 0.003 and 0.019.The AUC of the five characteristics of the number of capsules and the density of the solid part of the pancreas parenchyma were 0.660 and 0.667 and 0.630 and 0.681 and 0.545,respectively.Based on the data from multivariate and univariate logistic regression analysis,we used the number of capsules as an independent predictor[OR value 0.344,96%CI(0.536-0.724),p=0.001].In the radiomics analysis,the ICCs of 708 radiomics parameters after applying the repeatability analysis were>0.80.The results of dimensionality reduction analysis found that radiology parameters were selected based on plain scan,pancreatic parenchymal phase,plain scan combined with pancreatic parenchymal phase,plain scan combined with portal vein phase,pancreatic parenchymal phase with portal vein phase,and three-phase combination,and radioics labels were created separately.Based on the performance of the model in the training and validation groups,we selected the radiomic feature in plain scan+venous to develop a mixed model as the best radioics label.The results show that the AUCs of the mixed model in the training group and the validation group are 0.951(0.910-0.993)and 0.815(0.651-0.978),respectively.The accuracy,sensitivity,and specificity in the training group were 0.892,0.881,and 0.917,respectively,and the accuracy,sensitivity,and specificity in the verification group were 0.848,0.917,and 0.667,respectively.The Hosmer-Lemeshow test indicates that the joint model has good consistency in the training group and the verification group,and there is no statistical difference(P values are 0.9334 and 0.6924,respectively).The decision curve shows that the model has good clinical application value.Conclusion:Gender,the number of cysts,the diameter of the main pancreatic duct,whether the lesion is in communication with the main pancreatic duct,and enhanced scanning of the solid part density of the pancreatic parenchymal tumor have statistical significance in identifying pancreatic SCN and mucinous cystic neoplasm.We have created a model alignment chart combining the number of cysts and the best imaging omics label as an effective tool to identify pancreatic SCN and mucinous cystic neoplasm before surgery,which is helpful for the clinical diagnosis and decision-making of patients with pancreatic cystic neoplasm.