Prediction Of Early Recurrence After Hepatectomy And Microvascular Invasion In Hepatocellular Carcinoma By Gd-EOB-DTPA MRI Radiomics Model

Objective:This study aims to establish a prediction model for preoperative microvascular invasion and early recurrence after liver cancer resection based on retrospective analysis of Gd-EOB-DTPA MRI radiomic features.Materials and methods:Data of patients who underwent resection of hepatocellular carcinoma from January 2010 to March 2022 in Nanfang Hospital were retrospectively analyzed.They all underwent preoperative Gd-EOB-DTPA MRI scans and pathological diagnosis.The clinical data,pathological reports and imaging characteristics of these patients were collected.The imaging features were further analyzed by two experienced radiologists.The VOIs were delineated semi-automatically.Radiomic features were extracted using professional software and checked for consistency.?Early recurrence prediction model:According to the review results within 1 year after surgery,the cases were divided into recurrence group and non-recurrence group.The extracted feature parameters were subjected to significance analysis,single-factor and multi-factor logistic regression analysis to screen out the key features.The screened features were incorporated into a binary logistic regression equation to build a prediction model.The prediction performance of each model is compared through decision curve analysis and ROC analysis.H-L goodness of fit test is performed,and a calibration curve is drawn to evaluate the models.Assessing the model's ability to discriminate early relapse risk populations by RFS survival analysis.?MVI preoperative prediction model:According to postoperative pathological results,cases were divided into MVI positive group and MVI negative group.Training set and test set(86/38)were split.Data standardization,variance threshold selection,significance analysis,random forest sorting,correlation analysis,recursive feature elimination on feature parameters were performed to reduce feature dimensions.SVM prediction model and RF model were established respectively and,the prediction performance of each model was evaluated to find out the best model.Results:93 texture features from the late arterial,portal and hepatobiliary images with,a total of 279 feature were extracted in first part.1781 features were extracted respectively from the late arterial,portal,and hepatobiliary images,with a total of 5343 radiomic features in the second part.?Early recurrence prediction model:recurrence group/non-recurrence group(13/41).The final inclusion features were AFP,ALT and 1 texture feature in portal venous phase,.The AUC of the radiomics model was 0.921(95%CI,0.815-0.977)calculated by ROC analysis.When the threshold was 0.37771,the prediction accuracy of the radiomics model was 0.926,the sensitivity was 0.846,the specificity was 0.951.The precision,the recall rate and the Fl-value are 0.846.The prediction performance of this model,whose significance level of H-L goodness-of-fit test was 0.254(>0.05),is superior to the clinical-pathological feature model.?MVI preoperative prediction model:MVI positive group/MVI negative group was 43/81.The features finally included in the radiomics model included 3 hepatobiliary phase features.The AUC of the training set and test set were 0.905(95%CI,0.823?0.958)and 0.843(95%CI,0.689?0.940),respectively.The sensitivity and specificity of the radiomics model were 0.900 and 0.821 in the training set and 0.846 and 0.760 in the test set,respectively.The prediction performance of the radiomics feature model was better than that of the clinical feature model.Inclusion of clinical and semantic features in radiomics model,did not improve the model performance.Conclusion:Gd-EOB-DTPA MRI radiomics features can be used to predict early recurrence and microvascular invasion of hepatocellular carcinoma,with better prediction performance compared to clinical features and semantic features.