Research On MR-based Radiomics In The Differentiation Of Hepatocellular Carcinoma And Hepatic Hemangiomas

Background:Hepatocellular Carcinoma(HCC)is the most common type of primary malignant hepatic tumor,which is the third leading cause of cancer-related death worldwide.Hepatic hemangioma(HHE)is the most frequently diagnosed benign hepatic tumor.The treatment for HHE is markedly different than HCC.The HCC patients are better performed by surgical resection,liver transplantation or interventional therapy,whereas the HHE patients generally do not need any intervention unless the symptoms are obvious.Therefore,the determination of the follow-up treatment plan depends on the accurate identification of the two kinds of tumors in clinical practice.Although HCC and HHE have different pathological structure,while the imaging findings of them are often similar,which will decrease the accuracy and consistency of of the visual judgement of medical images in the traditional diagnosis methods.Radiomics is a new research direction in the field of medical imaging,which exploits the high-throughput feature from medical images to combine with machine learning method for disease diagnosis,tumor grading,and prognosis,etc.Therefore,Radiomics can be used as a new tool to identify hepatocellular carcinoma and hemangioma.Purpose:To assess the value of radiomic features to differentiating Hepatocellular Carcinoma from Hepatic Hemangioma base on enhanced magnetic resonance images.Methods:A Gd-EOB-DTPA-MRI epatobiliary phase transverse image data set containing 134 patients with 186 tumors(105 HCC and 81 HHE)was retrospectively enrolled in this study from Southern Hospital of Southern Medical University.The lesions at the largest cross-sectional slice are manually extracted by two experienced radiologists and then are used to extracted radiomic features.The extracted result contain 373 features,including first-,second-,and higher-order statistics features.The intra-class correlation coefficient(ICC)was used to estimate the reproducibility.An independent t-tests or Mann-Whitney U test was used to assess the differences between two groups,and receiver operator characteristic curve(ROC)was used to assess classification ability for each feature.The feature selection was carried out by the Least absolute shrinkage and selection operator(LASSO)algorithm and the minimal Redundancy maximal Relevance(mRmR)method based on maximal information coefficient(MIC)respectively.The selected feature subset is used to train the support vector machine for building the diagnostic model.The data set was divides into training set(including 81 HCC lesions,HHE 60 lesions)and test set(including HCC,HHE and 24 lesions and 21 lesions)according to the acquisition time.The 5 repeated 10 fold cross validation experiments are used to evaluate the expected classification performance of the above methods.This paper uses the parameters of area under the ROC curve(AUC),sensitivity,specificity and accuracy to evaluate the classification performance.The tested ROC acquired using the final models produced by the two futures selection methods was assessed by Delong test.Results:ICC values of more than 90%features(343/373)are larger than 0.90,including more than 23%features(80/343)with significance difference between the group of HCC and the group of HHE.The ROC analysis results show that the feature LZHGE of GLSZM can obtain best performance of classification,and the AUC,sensitivity,specificity and accuracy are separaetly equal to 0.81(p<<0.01),0.70,0.80 and 0.74.In feature selection and classifier training,the models obtained by the two different algorithms showed good discrimination performance on both the training set and test set.The radiomics signature consisting of 11 selected features based on LASSO algorithm obtain an AUC value of 0.834(95%CI 0.807-0.861),sensitivity of 0.795(0.755-0.836),specificity of 0.667(0.631-0.701),accuracy of 0.741(0.716-0.764)in the traning set.And the AUC obtained by the signature in the test set are 0.851,with sensitivity of 0.791,specificity of 0.667 and accuracy of 0.733.The radiomics signature consisting of 15 selected features based on the MIC-mRmR algorithm achieved an AUC value of 0.915(95%CI 0.891-0.940),sensitivity of 0,888(0.855-0.922),specificity of 0.803(0.756-0.850)and accuracy of 0.852(0.824-0.879).The test AUC,sensitivity,specificity,accuracy obtained by the signature are separaetly 0.892,0.875,0.760 and 0.822.The results show that performance of the model based on the MIC-mRmR algorithm is better than LASSO,but the test ROC difference between the two models is not statistically significant(P = 0.21).Conclusion:Using radiomic feature based on Gd-EOB-DTPA-MRI images can well identify hepatocellular carcinoma and cavernous hemangioma,which potentially provide a favorable method for clinical diagnosis.