A Preliminary Study On The Value Of Radiomic Based On Mri In Predicting Microvascular Invasion Grading In Hepatocellular Carcinoma

Purpose: To evaluate the diagnostic efficacy of 2-classification and 3-classification prediction tasks based on clinical imaging indexes and imaging omics features extracted from enhanced MRI in identifying different MVI grades of HCC,to explore the feasibility of imaging omics in diagnosing VETC,to compare the diagnostic performance of models established by different machine algorithms,and to screen the best machine learning model for preoperative prediction of VETC.Methods: 340 HCC patients with preoperative enhanced MRI examination(M0: 177;M1:121;M2:42;VETC+:21;VETC-:143)were included.The process of radiomics analysis included tumor segmentation,feature extraction,data preprocessing,dimension reduction,modeling and model evaluation.Different machine learning algorithms were used to establish clinical＿radiological,single sequence imaging omics,multi sequence imaging omics and clinical＿radiological＿radiomics models.In order to explore the differential ability of radiomics between the three MVI grades,2-and 3-class tasks were established respectively.The 2-class classification tasks included M0 vs M1(Task 1),M1 vs M2(Task 2)and M0 vs M2(Task 3).For the 3-class classification task,340 lesions were randomly divided into training set(236 cases)and test set(104 cases).For the VETC prediction model,164 lesions were randomly divided into training set(114 cases)and test set(50 cases).The discrimination ability of radiomics features was determined through the training set and tested through an independent test cohort.The predictive performance of the task was evaluated by subject operating characteristic(ROC)curve,calibration curve analysis and decision curve analysis(DCA),and the AUC was compared by Delong test.Results: In the 2-class task: the pathological grade and maximum diameter of tumor are the independent predictors of Task 1;AST,AFP and tumor edge are the independent predictors of Task 2;AST,AFP,maximum diameter and tumor edge are the independent predictors of Task 3.Combined with the clinical independent predictors,the clinical＿radiological＿radiomics(C＿R＿RO model)performed good diagnostic ability.The maximum areas(AUCs)under the ROC curve of Task 1,Task 2 and Task 3 were 0.766(95% CI 0.714-0.813)?0.830(95% CI0.763-0.884),0.941(95% CI 0.901-0.968)respectively.In addition,AFP,pathological grade,maximum tumor diameter and tumor edge were independent predictors of different MVI grades in the 3-class task.The C＿R＿RO model of 3-class showed good discrimination ability in distinguishing M0,M1 and M2,with the AUC 0.802(95% CI 0.747-0.857)?0.706(95% CI 0.638-0.775)?0.894(95% CI 0.836-0.952).Moreover,the random forest model in VETC prediction model represented the highest diagnostic efficiency,with the AUC 1.000(95% CI 1.000-1.000),0.859(95%CI 0.794-0.924)in training set and test set respectively.Conclusion: AFP,pathological grade,maximum tumor diameter and tumor margin were independent predictors of different MVI grades.The combined model combining clinical＿radiological independent predictors and MRI related radiomic features is expected to identify different MVI grades of HCC,and the random forest model based on MRI imaging omics is an effective tool for preoperative prediction of VETC.