Application Of MRI-based Radiomics In Predicting The Nature Of Nodules Associated With Liver Cirrhosis

Objective:To investigate the application value of radiomics based on MRI sequences in predicting hepatocellular carcinoma（HCC）,Regenerative nodules（RN）and Dysplastic nodules（DN）.Materials and methods:1.Case collectionModeling dataset:206 patients who underwent liver resection and liver transplantation with pathological diagnosis of Regenerative nodules（RN）,Dysplastic nodules（DN）and hepatocellular carcinoma（HCC）in the Second Affiliated Hospital of Dalian Medical University from January 2018 to January 2022 were collected,and 282lesions were obtained,including 56 RN,72 DN and 154 HCC.Independent external validation dataset:60 patients with pathology confirmed to contain RN,DN and HCC were collected in the Second Affiliated Hospital of Dalian Medical University from February 2022 to December 2022 at different time periods,and 80 lesions were obtained,including 20 RN,20 DN and 40 HCC.2.MR imaging equipmentAll patients underwent MRI using Siemens Skyra 3.0T,Siemens Verio 3.0 T and GE HDxt 1.5T MR scanners with scan sequences including:T₁WI,T₂WI,DWI,and enhanced T₁WI（T₁WI＿A,T₁WI＿V and T₁WI＿D）.3.MRI semantic feature evaluation methodsMRI semantic features for all patients,specifically including:maximum lesion diameter,T₁WI signal intensity（low,equal and high signal）,T₂WI signal intensity（low,equal and mild to moderate high signal）,T₂WI signal of uniform intensity,enhancing"capsule",non-circular arterial phase hyperenhancement,non-peripheral"washout",restricted diffusion,blood products in mass,periomal edema,mosaic architecture,fat sparing in solid mass and iron sparing in solid mass.Two diagnostic abdominal system imaging physicians（with 3 and 4 years of diagnostic abdominal imaging experience,respectively）performed independent analyses using a double-blind method to reach a consensus opinion and make a diagnosis,comparing their diagnoses with the final pathological diagnosis.When the two physicians’assessments did not agree,an imaging physician with 11 years of experience in diagnostic MRI of the abdominal system was involved in the discussion to reach a final consensus result.4.Image-pathology controlUsing the imaging pathology control method,the MRI tomographic images were used as a reference for histological blocking of the bulk liver specimen,with the orientation and thickness of the tissue block cross-sectioned in line with the orientation and thickness of the MRI cross-sectional images.Image localization and labeling of the lesions were performed on the corresponding cross-sectional MRI images based on Couinaud segmentation,followed by HE staining and immunohistochemical examination of the lesions to determine the nature of the nodules.Lesions pathologically characterized as RN,DN and HCC were selected for inclusion in this study.5.Model building process5.1 Image segmentationThe MRI images of all patients are imported into the 3D Slicer open source software in DICOM format,and the regions of interest（ROI）of the lesions are outlined layer by layer,and the resulting 3D volume of interest（VOI）data are uploaded to the cloud platform for feature extraction.5.2 Extraction and screening of radiomics featuresA total of 2264 features were extracted from the outlined VOI files and patient DICOM images by uploading them to the radiomics platform（United Imaging Medical Technology Ltd.http://urp.united-imaging.com/#/）.The variance threshold,select k best and least absolute shrinkage and selection operator were used for feature screening to obtain the optimal features for building the prediction model.5.3 Model buildingFive classifiers,decision tree,logistic regression,random forest,stochastic gradient descent and support vector machine,were used to establish the radiomic feature model,semantic feature diagnosis model and joint model based on semantic feature and radiomic feature,respectively,for the best eigenvalues screened by the five-fold cross validation.5.4 Model evaluationThe Receiver Operator Characteristic（ROC）curve,sensitivity,specificity,accuracy and F1 values were used to evaluate the effectiveness of different classifiers.Decision Curve Analysis（DCA）was used to evaluate the clinical application value of the model.Independent external validation data sets were used for validation to determine the model with the best final performance.6.Statistical methodsSPSS 26.0 and R software were used for statistical analysis of the data.The semantic features of patients’MRI matched the frequency of categorical variables.χ2test or Fisher exact test were used for comparison between groups,and multiple logistic regression was used to identify independent predictors of RN,DN and HCC.P<0.05was statistical different.Results:1.Analysis of clinical dataThe imaging characteristics of 282 lesions in 206 patients in the modeling dataset were analysed by univariate analysis and multiple logistic regression,showing that T₂WI signal intensity（low,equal and mild to moderate high signal）,enhancing"capsule",iron sparing in solid mass,restricted diffusion,non-circular arterial phase hyperenhancement,and non-peripheral"washout"is an independent predictor for differentiating RN,DN and HCC.2.Feature screening resultsThe Variance Threshold,Select K Best and Lasso algorithms were used to filter 6,6,7,8 and 8 optimal features from T₁WI,T₂WI,T₁WI enhanced,DWI and fused all sequence images out of 2264 features.3.Efficacy of different models3.1 The best model diagnostic efficacy was found for the joint model SVM classifier in the internal test set,where the AUC values（95%CI）,sensitivity,specificity,accuracy and F1 values were 0.936（0.923-0.949）,0.763,0.909,0.808 and 0.745,respectively.The DCA curves indicated a higher net benefit for the LR,SGD and SVM models,while the DT and RF models had relatively low net benefits.3.2 The joint model SGD classifier model with the best efficacy and stability in the external validation set,where the AUC values（95%CI）,sensitivity,specificity,accuracy and F1 values were 0.925（0.881-0.969）,0.767,0.914,0.825 and 0.772respectively.The DCA curves indicated a higher net benefit for the five classifiers HCC and a lower net benefit for RN and DN.Conclusion:1.The radiomics-based approach with different machine learning models has important clinical value for predicting RN,DN and HCC.2.In the external validation,the combined model had the highest efficacy and stability in predicting the diagnosis of RN,DN and HCC.3.The combined model obtained good performance in predicting RN,DN and HCC in both the internal test set and external validation set,and the diagnostic performance was better than that of the single model,which has broad application prospects.