Research On The Application Value Of MR Imagingomics In The Diagnosis Of Liver Fibrosis In Rats And The Selection Of Optimized Models

PartⅠ Application value of radiomics based on MRI in evaluation of liver fibrosisObjective To explore the application value of MRI-based radiomics in diagnosis and staging of liver fibrosis.Materials and Methods A prospective radiomics study was conducted by subcutaneous injection of carbon tetrachloride(CCl4)to induce and establish a rat liver fibrosis model.Signa-HDx 3.0 Tesla MRI(GE Medical Systems,Milwaukee,WI)was used to scan anesthetized rats to obtain three-dimensional fast spoiled gradient echo(3D-FSPGR)T1WI images.3D-Slicer software(free and open source medical image analysis visualization software)and IBEX software(MD Anderson Cancer Center,US)were used to draw the regions of interest(ROI)and extract feature parameters.After data preprocessing and dimension reduction,a binary classification radiomics model was established.The radiomics models for stage F0-F4,F0 and F1-F2,F1-F2 and F3-F4 diagnosis were established respectively.Results A total of 1767 sets of feature texture parameters were obtained,and 1742 sets remained after data preprocessing.Seven models were established by LASSO + SVM for the study of the above groups.The area under curve(AUC)of the model established in stage F0-F4 diagnosis was 0.90,0.85,0.71,0.89 and 0.83,respectively.The AUC of normal liver(stage F0)and early liver fibrosis(stage F1-F2)was 0.88,and the AUC of early liver fibrosis(stage F1-F2)and advanced liver fibrosis(stage F3-F4)was also 0.88.The sensitivity,specificity and accuracy of the seven models were all higher than 80%.Conclusions Raiomics can be used in the non-invasive diagnosis and staging of liver fibrosis,and it was a new diagnostic method of liver fibrosis.PartⅡ The selection of optimal reduced dimension model in the study of radiomics of liver fibrosis Objective To explore how to obtain more optimized dimensionality reduction and modeling methods for the diagnosis and staging of liver fibrosis.Materials and Methods 97 rats were randomly divided into a control group(25 rats)and a liver fibrosis group(72 rats).CCl4 was used to induce rat liver fibrosis models.Signa-HDx3.0 Tesla MRI was used to scan the rats to obtain 3D-FSPGR T1 WI images.Liver tissue biopsy was performed immediately after scanning to obtain liver fibrosis stage.3D slicer software and IBEX software were used for tumor segmentation and feature parameter extraction.Py Charm(Jet Brains.ro,version 4.5.4)and R-Studio software were used for data pre-processing,dimension reduction and establishment of multi-classification models.R-Studio software was used to draw the ROC curve and obtain the AUC,calculate the Kappa value of different models and visualize the data,draw the decision curve analysis(DCA)for different models.The fuzzy comprehensive evaluation method was used to weight the three statistical methods according to the ratio of 1:1:1 to obtain the best model for comprehensive diagnosis of liver fibrosis.Results A total of 1767 sets of feature texture parameters were obtained.After preprocessing the data,part of the incomplete feature texture parameters were deleted and 1742 sets remained.By using two different dimension reduction methods and six modeling methods,12 types of multi-class radiomics models were established.LASSO + BPNet and PCA + SVM had higher diagnostic efficiency and decision-making value for stage F0(AUC=0.92 and0.91,Kappa values were 0.81 and 0.87,respectively).LASSO + BPNet was the best model for diagnosis of stage F1(AUC=0.90,Kappa value was 0.84),LASSO + KNN was the best model for diagnosis of stage F2(AUC=0.75,Kappa value was 0.67).LASSO + SVM and PCA + GBDT were both the best models for diagnosis of stage F3(AUC=0.89 and 0.91,Kappa value were 0.92 and 0.95,respectively).LASSO + DT was the best model for diagnosis of F4 stage(AUC=0.90,Kappa value was 0.88).The analysis result of fuzzy comprehensive evaluation method showed that the comprehensive evaluation score of LASSO + SVM was 16,it was the highest score among all methods.Conclusions The optimal diagnosis models were different for different stages.LASSO+ SVM had the best comprehensive diagnosis ability for stage F0-F4.