CT-and MRI-based Models For Non-invasive Diagnosis Of Portal Hypertension In Patients With Cirrhosis

Clinically significant portal hypertension(CSPH)indicates an increased risk of decompensation in patients with cirrhosis.Early identification and early intervention of CSPH are keys to the management of cirrhosis.However,the gold standard for CSPH,hepatic venous pressure gradient(HVPG)measurement,is invasive and expensive,and has certain technical requirements for the operators.The aim of this study is to explore the value of CT-and MRI-based non-invasive models for diagnosis of CSPH(defined as an HVPG≥10 mmHg)in patients with cirrhosis.This study was divided into three parts,as shown below.1.CT-based radiomics model for non-invasive diagnosis of CSPHThis study prospectively involved 385 patients with cirrhosis from five hospitals in China between August 2016 and September 2017.The radiomics model,which we termed rHVPG,was developed based on CT images in a cohort 222 patients from one of the five centers.The diagnostic performance of rHVPG was assessed in patients from four other involved centers,which were served as four external validation cohorts(n=136).In the training cohort,rHVPG showed a high diagnostic accuracy for detection of CSPH with the area under the receiver operating characteristic curve(AUC),sensitivity,and specificity of 0.849,78.7%,and 76.9%,respectively.Application of rHVPG in four external prospective validation cohorts still had excellent performance with the AUCs of 0.889,0.800,0.917,and 0.827,respectively.Conclusion:rHVPG was accurate for detection of CSPH in patients with cirrhosis.In resource-limited settings where the HVPG measurement was not ready.rHVPG could serve as an auxiliary parameter for detecting CSPH.2.MRI-based liver surface nodularity quantification for non-invasive diagnosis of portal hypertensionThis study prospectively recruited 139 patients with cirrhosis from five hospitals with four in China and one in Turkey between December 2018 and April 2019.Besides,ten patients was retrospectively recruited from a hospital in Italy between March 2015 and November 2017.All the included patients had HVPG measurements and upper-abdominal MRI scans.Patients from four Chinese centers were served as the training cohort(n=125)while patients from two international centers were considered as the validation cohort(n=25).The proposed model,CHESS-DIS score was calculated based on liver surface nodularity quantification using non-contrast-enhanced MRI images.The correlation between CHESS-DIS score and the transjugular HVPG value was studied.A positive correlation between CHESS-DIS score and HVPG was found with the correlation coefficient of 0.36(P<0.0001)and 0.55(P<0.01)in the training and validation cohorts,respectively.The intraclass correlation coefficients for assessing the inter-and intra-observer agreement were 0.846 and 0.841,respectively.AUCs for detection of CSPH were 0.8l and 0.91 in the training and validation cohorts,respectively.Conclusion:The non-contrast-enhanced MRI image-based CHESS-DIS could be used to detect CSPH in patients with cirrhosis.As CHESS-DIS score was significantly correlated with the invasive HVPG with a good reliability,it might be used to non-invasively assess portal hypertension in future.3.CT/MRI-based deep convolutional neural network models for non-invasive diagnosis of CSPHBesides including patients from the above mentioned CT and MRI cohorts,new participants were recruited to enlarge the sample size.The CT cohort comprised 406 participants with cirrhosis from four hospitals in China.Besides,a cohort of 271 participants without chronic liver diseases but had upper-abdominal CT scans were recruited form one hospital to balance the CSPH ratio.The MR cohort comprised 164 participants with cirrhosis from nine hospitals with eight in China and one in Turkey.And a cohort of 271 participants without chronic liver diseases but had upper-abdominal MRI scans were also recruited.In each cohort,participants were randomly divided into training,validation,and testing sets at a 3:1:1 ratio,in order to develop,fine-tune,and assess the proposed models,respectively.The CT-based deep convolutional neural network(DCNN)model identified CSPH with AUCs of 0.998,0.912 and 0.933 in the training,validation and test sets.The MR-based DCNN model analysis identified CSPH with respective AUCs of 1.000,0.924,and 0.940 in three sets.Conclusion:This study established the first DCNN model in the field of portal hypertension,which could facilitate the diagnosis of CSPH.