The Value Of CT Radiomics In Predicting The Rupture Risk Of Cirrhosis-related Esophageal-gastric Varices

[Objectives]To explore the ability of nomogram combining CT imaging radiomics and clinical high risk index to determine the risk of esophageal and gastric varicose rupture in cirrhosis.[Materials and Methods]This retrospective study was reviewed and approved by the Ethics Committee of Qilu Hospital of Shandong University(ethics number:KYLL-2019-2-019),and the patients were exempted from their informed consent.A total of 281 patients were included,including 178 males and 103 females.All patients underwent abdominal CT scan and upper gastrointestinal endoscopy,and the patients were classified into low risk of rupture bleeding(n=96)and high risk of rupture bleeding(n=185)according to the endoscopic results.According to CT scan phase,all cases were divided into non-contrast enhanced group(n=93),contrast-enhanced arterial phase group(n=116),portal phase group(n=116),enhanced double combination group(n=116),and portal/splenic vascular target group(n=116)to construct the radiomics model,the samples were randomly divided in 7:3 ratio.In all cases,93 for non-contrast enhanced data and 116 for arterial and venous data respectively,meanwhile the other 72 patients was regarded as an external validation data set underging non-contrast enhanced and enhanced arteriovenous-phase CT scanning.Using a radiomics cloud platform(Radcloud platform,Huiying medical technology co,LTD.),the liver,spleen,portal vein and its mian branches,and splenic vein was drawn as the region of interest(ROI)in different phases to extract the imaging radiomics characteristic values,some characteristics related to varicose vein rupture bleeding were screened through the variance threshold method(variance threshold),SelectKBest algorithm and minimum absolute shrinkage and selection operator(LASSO)method,and then the efficiency were verified using the internal validation model.After the above process,17 imaging radiomics models were obtained,and their receiver operating characteristic curves(ROC)and the area under the curves(AUC)were observed.Then the two better models was screened including the right portal branch support vector machine(support vector machine,SVM)and the liver plus spleen and the combined SVM classifier model of the non-contrast enhanced group.External validation data were input into the best model to further verify its diagnostic efficacy.The platelet was regarded as an independent clinically high risk bleeding-related factor through unit and multiple regression analysis,and combined with imaging radiomics model together to build imaging radiomics nomogram model of rupture risk of esophageal and gastric varicose veins.The calibration curve and decision analysis curve(decision curve analysis,DCA)were also drawn to evaluate the calibration and clinical effectiveness of the developed radiomics nomogram.[Results]Among the five different radiomics models,the portal/splenic vein-target group and non-contrast enhanced group had the highest overall prediction efficacy.Among them,the AUC area of the right portal vein,the portal vein and splenic vein were all over 0.9,which showed good efficacy.However,the SVM classifier model for the right branch of the portal vein was optimal,with the lower AUC area of 0.981 and 0.975 in the training and validation sets,respectively,and area under the AUC was 0.866 after substituting the external validation set data into the model.In non-contrast enhanced group the area under the AUC of the liver and spleen combined with the SVM classifier model training set and validation sets in the non-contrast enhanced group was 0.927 and 0.909,respectively,and area of the external validation set data was 0.850 under the AUC into the model,and both of the above models showed good predictive efficiency.The area under the AUC of the external validation set of the nomogram constructed by combining the portal/splenic vein-target and non-contrast enhanced model with the clinical high-risk bleeding factor model was 0.907 and 0.914,respectively,higher than the AUC value of the single radiomics model,the comparison results showed that the nomogram constructed by radiomics combined with clinical high-risk factors had better prediction performance.[Conclusions]The nomogram based on the right portal vein model of venous phase group and the combined liver and spleen model of the non-contrast enhanced group in multiphase CT imaging histology combined with clinical high-risk factors has a high predictive effect on the risk of variceal bleeding caused by cirrhosis.