Application Of MR Radiomics In Prediction Of Pathological Grading And MVI Of Hepatocellular Carcinoma

Part Ⅰ Feasibility study of MR radiomics in preoperative predicting of pathological grading of hepatocellular carcinomaObjective:To investigate the clinical feasibility of MR radiomics in preoperative prediction of the pathological grading of hepatocellular carcinoma.Materials and methods:A total of 447 patients with surgically pathologically confirmed HCC with complete imaging data were collected from February 2012 to December 2017,all examinations were completed on 3.0T MR(scanning machine discovery MR750,3.0T,GE Healthcare).The pathological grading of HCC was divided into high grade and low grade by international general grading method.According to the inclusion and exclusion criteria,241 HCC patients were included in this study,including 198 males and 43 females,with an average age of 54.75±10.57 years.All patients were divided into training group(n=181)and test group(n=60)according to the ratio of training group to test group=3:1.Firstly,ITK-SNAP software was used to manually segment the same lesion displayed by T1 weighted imaging(T1WI)and T2 weighted imaging(T2WI)in DICOM format in two groups.This work was completed by the same radiologist with more than 10 years of experience in abdominal diagnosis;Secondly,the tumor radiomics features of all patients based on T1WI and T2WI images were extracted using Matlab 2010A(Mathworks,Natick,USA)software;Then the least absolute shrinkage and selection operator(LASSO)logistic regression model,which is suitable for the regression of highdimensional data,was used to select the useful features and generate the radiomics prediction model.Three radiomics models were established:T1WI based radiomics model-1,T2WI based radiomics model-2,and T1WI&T2WI based radiomics model-3.In addition,a clinical model-G was established which was composed of age,sex,tumor size,Alpha fetoprotein(AFP)level,history of hepatitis B,liver cirrhosis,venous tumor thrombus,portal vein hypertension,tumor pseudo capsule.By combining the above three radiomics models with the clinical model-G,three combined models for predicting HCC pathological grade were obtained,namely:combined T1WI based radiomics and clinical model(combined model-G1),combined T2WI based radiomics and clinical model(combined model-G2)and combined T1WI&T2WI based radiomics and clinical model(combined model-G3)respectively.Evaluation the ability of above seven model preoperative predict HCC pathologic stage.Results:In both the training group and the test group,T1WI based radiomics model-1,T2WI based radiomics model-2,and the combined T1WI&T2WI based radiomics model-3 were able to distinguish high grade HCC from low grade HCC successfully(p<0.05).In the test group,the area under the receiver operating characteristic(ROC)curve(AUCs)for predicting HCC grade by the three radiomics models were 0.712,0.721 and 0.743,respectively.AUC for predicting HCC grade by the clinical model-G was 0.645.AUCs for predicting HCC grade by the combined model-G1,combined model-G2,combined model-G3 were 0.769,0.796,and 0.820,respectively.In the multivariate analysis for preoperative prediction of HCC pathological grade based on the combined model-G3,AFP level and radiomics signature were independent predictors of HCC pathological grade(p<0.05).Conclusion:The radiomics models based on T1WI or T2WI and T1WI&T2WI can predict the pathological grade of HCC in some degree.The combined model combining radiomics signature and clinical factors can further improve the ability of preoperative prediction of HCC pathological grade.Among them,radiomics signature and AFP level can be used as independent predictors of HCC pathological grading.Part Ⅱ Feasibility study of MR radiomics in preoperative predicting microvascular invasion of hepatocellular carcinomaObjective:To investigate the value of MR radiomics in predicting Microvascular invasion(MVI)of hepatocellular carcinoma(HCC).Materials and Methods:The collection of the patients was the same as in Part Ⅰ.According to the inclusion and exclusion criteria,241 HCC patients were included in this study(due to the difference in pathological diagnosis results,the composition of the 241 patients in the two parts was not exactly the same).There were 196 males and 45 females,with an average age of 54.97±10.81 years;According to whether there are cancer cell nests in the vascular lumen of HCC patients under the microscope,they were divided into MVI positive and MVI negative.A total of 241 patients were divided into training group(n=179)and test group(n=62)according to the ratio of 3:1.The T1WI and T2WI images of the patients were loaded into ITK-SNAP software,and the Three-dimensional(3D)of the Region of interest(ROI)were manually segmented.MATLAB 2010A software was used to extract tumor radiomics features based on T1WI and T2WI images of all patients,and three radiomics model for predicting MVI was established,i.e T1WI based radiomics model-Ⅰ,T2WI based radiomics model-Ⅱ,and T1WI&T2WI based radiomics model-Ⅲ.Another clinical model-M was established by age,sex,tumor size,AFP level,hepatitis B,liver cirrhosis,portal hypertension,tumor pseudo capsule,etc.Combining the above three radiomics models with clinical models,three combined models for predicting MVI were obtained,i.e combined T1WI based radiomics and clinical model(combined model-MⅠ),combined T2WI based radiomics and clinical model(combined model-MⅡ),combined T1WI&T2WI based radiomics and clinical model(combined model-MⅢ).The ability of the the above seven models to predict HCC MVI before operation was evaluated respectively.Results:In both the training group and the test group,MVI-positive HCC and MVI-negative HCC could be distinguished successfully by radiomics model-Ⅰ,radiomics model-Ⅲ and radiomics model-Ⅲ(p<0.05),with AUCs of 0.698,0.734 and 0.763,respectively.The AUC for predicting HCC MVI by the clinical model-M was 0.656,while the AUCs for predicting HCC MVI by the combined model-MⅠ,combined model-MⅡ and combined model-MⅢ were 0.700,0.774 and 0.847,respectively.In the multivariate analysis for preoperative prediction of HCC MVI based on the combined model-M3,radiomics model and tumor capsule were independent predictors of HCC MVI(p<0.05).Conclusion:The radiomics models based on T1WI or T2WI and T1 WI&T2WI can distinguish MVI-positive and MVI-negative HCC in some degree.The combined model combining radiomcs signature and clinical factors can further improve the ability of prediction of HCC MVI.Among them,radiomics signature and tumor pseudo capsule can be used as independent predictors of HCC MVI.