Contrast-enhanced CT And MRI Based Radiomics For Preoperatively Predicting Microvascular Invasion In Hepatocellular Carcinoma

Objective:To investigate the value of contrast-enhanced computer tomography(CECT)and magnetic resonance imaging(CE-MRI)texture analysis combined with machine learning random forest model for preoperatively predicting microvascular invasion(MVI)in patients with hepatocellular carcinoma(HCC).The diagnostic efficacy between CT and MRI enhanced radiomics models of each phase images were also compared.Methods:CECT images of 111 patients and CE-MRI images of 116 patients with HCC were retrospectively analyzed.In addition,clinical factors(CF)and laboratory indicators within 1 week before surgery of these patients were collected.The data were divided into MVI-positive group and MVI-negative group according to the histopathological findings after tumor resection.Clinical and laboratory indicators were analyzed using the Mann-Whitney U test,x~2 test or Fisher exact test and multivariate logistic regression analysis to obtain clinically independent risk factors for predicting MVI.After standardization and preprocessing of each phase of CECT and CE-MRI images.Then these features were dimension reduction by three steps to filter out the optimal features.All the data were randomly divided into training group and test group according to the ratio of 8:2.Random forest models were build based on the clinical independent risk factors and the feature parameters selected in each phase of CECT and CE-MRI images in the training group.The receiver operating characteristic curve(ROC)was used to analyze the effectiveness of these models.The diagnostic efficiency of the clinical and radiomics models were further verified using the test groups,and the accuracy rate,sensitivity and specificity were calculated.Eighty-five of the above cases had undergone both CECT and CE-MRI examination.The texture features of the CECT and CE-MRI images were extracted,screened and analyzed again using the same method for each stage respectively.Again,they were randomly divided into training groups and test groups according to the ratio of 8:2.And the multi-phase random forest models were established.ROC was used to analyze the diagnostic efficacy of the models in each phase image and calculated the accuracy rate,sensitivity and specificity of the multi-phase models.The Delong test was used to compare the diagnostic efficacy between the above models.Results:After screening the texture features in CECT images of 111 patients,there were corresponding feature parameters were obtained in the arterial,portal venous and delayed phases respectively with statistical differences in distinguishing MVI-positive group from MVI-negative group(P < 0.05).And similarly in CE-MRI images of 116 patients,the corresponding feature parameters with statistical difference in different phase images were also obtained.In addition,a multi-factor logistic regression analysis of CF in the 142 patients included revealed that the maximum tumor diameter was the only independent risk factor for predicting MVI(P < 0.001).Further analysis of CF in the two groups of patients with CECT and CE-MRI,respectively,yielded the same results.In CECT,MVI-positive with the maximum tumor diameters of 7.0 cm(4.5-11.0)and MVI-negative groups with 4.0 cm(3.0-6.0).In CE-MRI,the maximum tumor diameters of MVI-positive and MVI-negative was 7.0 cm(4.0-11.0)and 3.3 cm(3.0-5.5),respectively.There were statistical differences between the two data sets(P <0.001).The AUC of the diagnostic efficiency of CF prediction models about CT and MRI in the test group were 0.704 and 0.749,respectively.The AUC of arterial phase(AP),portal vein phase(PVP),delayed phase(DP)and three-phase combined radiomics models of CECT were 0.749,0.773,0.763 and 0.768 in the test group,respectively.It was found that the diagnostic efficiency of the PVP radiomics model in the test group was slightly higher than that of the other three radiomics models,and the diagnostic efficiency of the CF model was lower than that of the radiomics models in each phase image.Similarly,The AUC of AP,PVP,DP and three-phase combination radiomics models of CE-MRI were 0.789,0.763,0.761 and 0.777 in the test group,respectively.The diagnostic efficiency of the AP radiomics model in the test group was slightly higher than that of the other three models,while the diagnostic efficacy of the CF model was lower than that of the radiomics models in all phase images.In addition,the combined PVP radiomics model and CF model of CECT was found to have higher diagnostic efficacy than the single phase radiomics model.And the combined three-phase radiomics model and CF model of CE-MRI had higher diagnostic efficacy than the single radiomics model.It indicated that the combined radiomics-clinical model could improve the predictive ability of HCC MVI.In the above two radiomics studies,there were 85 patients underwent both CECT and CE-MRI examination.Then the texture features of each phase of CECT and CE-MRI images were extracted for analysis and multi-phase random forest radiomics models were established.It was found that the diagnostic efficiency of PVP,DP and three-phase combined radiomics models in the test group were higher than that of anyone of the CE-MRI radiomics model and had no significant statistical difference.The preliminary suggestion was that the predictive efficacy of the two sets of radiomics models for HCC MVI were higher and consistent.Conclusion:The radiomics models were built by texture features based on CECT and CE-MRI images and they could predict HCC MVI preoperatively.No statistically significant difference in the diagnostic efficacy was found between the CECT and CE-MRI radiomics model with each phase image.Moreover,both CECT and CE-MRI radiomics models had high MVI predicting efficacy,but showed no significant difference.Thus,the radiomics models might be an effective tool to guide personalized treatment in clinic.