The Value Of Multiphase CT Imaging Omics In The Differentiation Of Focal-Autoimmune Pancreatitis And Pancreatic Cancer

Objective:To explore the value of the joint model of CT signs and multiphase omics features in the differential diagnosis of autoimmune pancreatitis(f-AIP)and pancreatic cancer(PC).Methods:73 cases of f-AIP patients and 74 cases of PC patients in 2012-2020,with clinical information and CT images,were collected respectively.Firstly,the CT signs of f-AIP and PC were evaluated and identified through imaging data by two radiologists.Secondly,all levels of three-phase lesions were uploaded to the Darwin Medical Research Platform,then the ROI on all level of the lesion were outlined and extracted radiomics features.By using the variance selection method and least absolute shrinkage and selection operator(LASSO)to screen out the radiomics characteristics,three phases of Support Vector Machine(SVM)were established.Through evaluated each model using ROC,the best phase diagnosis model was obtained and compared with artificial diagnosis results.Finally,morphological signs were merged to establish a joint diagnostic model,which was compared with the three-phase omics models.Results:Among the CT signs,the incidence of duct-penetrating sign and capsule-like rim in f-AIP was significantly higher than that of the PC,and the peripancreatic strands and vascular involvement were easier to appear in the PC(P<0.05).The venous phase CT value of around 10 mm of the center level of the pancreatic parenchyma lesion in the f-AIP group(84.5 HU)was lower than that in the PC group(93.7 HU)(P<0.01).The AUC was0.850(0.776 ～ 0.924),the sensitivity and specificity were 70% and 90% respectively in the established CT sign model.In the radiomics training group,the radiomics prediction model basing on the venous phase had the highest diagnostic efficiency,accuracy and the Youden index(AUC value was 0.93),which was higher than that based on unenhanced and arterial period(AUC value was 0.88 and 0.89 respectively).The unenhanced phase model showed the highest specificity(about 0.96),and the venous and arterial phase models had a similar sensitivity(about 0.92).The AUC value,sensitivity and specificity(0.97,0.96,and 0.90 respectively)of the venous period model were higher than the other periods in the verification group.In the training group of joint model,the AUC values of the joint radiomics model basing on the arterial and venous period were higher than the simple omics model.In the verification group of joint model,the AUC values of the joint model were higher than the simple omics model.Conclusion:In the identification of f-AIP and PC,the artificial model has a certain diagnostic ability,but when the signs are not typical,the misdiagnosis rate of manual diagnosis is high.By using the machine learning algorithm SVM to build a prediction model,the result showed that the three-phase radiological model was better than the CT signs model in the identification of f-AIP and PC.The venous phase model was higher than the unenhanced and arterial phase models,and the joint model of each phase omics characteristics and the CT symbol had a higher ability to identify.