Multiparameter MRI-Based Radiomics For Predicting Survival In Hepatocellular Carcinoma Patients Treated With Transarterial Chemoembolization

Purpose:The aim of this study was to develop and validate a survival prognostic model based on multiparameter magnetic resonance imaging（MRI）radiomics features and clinical characteristics of hepatocellular carcinoma（HCC）patients treated with transarterial chemoembolization（TACE）.Methods:A retrospective collection of 415 HCC patients treated with TACE between July 2014 and January 2020 was randomly divided into training set（n=290）and validation set（n=125）in the ratio of 7:3.MR images of Axial T₁-weighted imaging（T₁WI）,T₂-weighted imaging fat suppression（T₂WI-FS）,dynamic enhanced scan arterial phase（DCE-AP）and portal phase（DCE-PVP）were collected and imported into the ITK-Snap（version 3.8）to manually delineate the tumor region of interest layer by layer,and Py Radiomics（version 3.0）was used to extract the radiomics features.For each MR sequence,Pearson correlation analysis was firstly used for initial feature screening to obtain feature sets with low redundancy,and then two feature selection algorithms,recursive feature elimination（RFE）and filter selection（Relief）,were used to screen the radiomic signature associated with overall survival（OS）,and the classification algorithms used support vector machine（SVM）,random forest（RF）,logistic regression（LR）,Least Absolute Shrinkage and Selection Operator（LASSO）and Gaussian Process（GP）to construct the radiomics signature,and 5 fold cross-validation was used select the radiomic signature with best prediction performance.Clinical characteristics associated with overall survival（OS）were screened by univariate and multifactor Cox logistic regression analysis.Finally,radiomic model,clinical model,and clinic-radiomic combined model based on clinical characteristics and radiomic features were constructed using Cox proportional risk regression models.The predictive performance of the three models was evaluated by C-index as well as 95%confidence interval（CI）,and a nomogram were plotted to individually predict the 1-,2-,and 3-year survival of patients,and the calibration degree was evaluated using calibration curves.The optimal threshold was set by X-tile software（version 3.6.1）to classify patients into low-risk and high-risk groups,and Kaplan-Meier survival curves were plotted.Results:Radiomic signatures from LR-based T₁WI,LASSO-based T₂WI-FS,GP-based DCE-AP,and GP-based DCE-PVP obtained the best predictive performance.Among the clinical variables,postoperative changes in alpha-fetoprotein and postoperative combination with other treatments were independent risk factors for OS.Finally,the C-index of the clinic-radiomic combined model for OS prediction was higher than that of the radiomics model and the clinical model（training set:0.846vs.0.823vs.0.776;validation set:0.794 vs.0.780 vs.0.764）.The nomograms could individually predict the 1-year,2-year and 3-year survival of patients,and the calibration curves showed good agreement between the predicted and actual survival probabilities.Kaplan-Meier survival curves showed significant survival differences between patients in high-and low-risk groups（p<0.05）,and all three models could effectively distinguish patients with good and poor prognosis.Conclusion:The clinic-radiomic combined model can effectively predict the overall survival of hepatocellular carcinoma patients treated with TACE,and stratify the prognostic risk of patients,which can help clinical decision making.