Radiomics Models Of CT Scanning To Differentiate The Pleomorphic Adenoma,Warthin Tumor And Basal Cell Adenoma

Objective This research aims to differentiate pleomorphic adenoma,Warthin tumor and basal cell adenoma by building radiomics models of machine learning,and discuss the values of different models.Materials and Methods1.Information acquisition In this study,we retrospectively collected 242 patients with benign parotid tumors including 124 patients of pleomorphic adenoma,88 of Warthin tumor and 30 of basal cell adenoma in the Second Hospital of Dalian Medical University.These patients received maxillofacial CT scanning without contrast,then their CT image and basic information were collected.2.Feature extraction,screening and model constructionThe features of radiomics were extracted by Deepwise Research Platform(keyan.deepwise.com)with signing the maximal section of image by avoiding the maxilla artifacts,and finally 1015 features were obtained from the CT image including first-order features,morphological features(shape features),texture features anxd image filtering features,and then the data of features were standardized by Z-score standardization.These patients were divided into three groups.In the group of pleomorphic adenoma and Warthin tumor,they were randomly divided into training set(n = 148)and validation set(n = 64);in the group of pleomorphic adenoma and basal cell adenoma,they were randomly divided into training set(n = 107)and validation set(n = 47);in the group of Warthin tumor and basal cell adenoma,they were randomly divided into training set(n = 81)and validation set(n = 37).After that,Minimum absolute contraction operator,including LASSO,F test and tree-based,was used to screen features to obtain the most excellent.As a result,45 predicting models were established based on the classifiers of multivariable logistic regression(LR),random forest,(RF),linear discriminant analysis(LDA),support vector machine,(SVM),and extreme gradient boosting,(XG Boost).Finally,15 effective models of three groups were optioned with various classifiers.Prediction efficiency of the three models was evaluated by using the area under the curve(AUC),accuracy(ACC)and F1 value of the subject operating characteristic curve(receiver operating characteristic,ROC).The AUC of the different groups was evaluated by De Long test to compare the differences among five models of each group.Result 1.Feature extraction and screeningFrom the original image and the filtered image,1015 radiomics features were extracted from maxillofacial CT images of the liver of 242 patients.The dimensional reduction and importance range of radiomics features were carried out by LASSO algorithm,F test and tree-based,and the optimal features subset were selected respectively.Finally,20 predominant features were selected in each group.2.Model construction and effectiveness evaluationRadiomics models were built by LR,RF,SVM,LDA and XG Boost from extracted features.In the group of pleomorphic adenoma and Warthin tumor,SVM with F test is the most effective predicting model;in the training set,the AUC,ACC and F1 were 0.92,0.93 and 0.92 whereas 0.86,0.77 and 0.80 in the validation.In the group of pleomorphic adenoma and basal cell adenoma,LDA with F test is the most effective predicting model;in the training set,the AUC,ACC and F1 were 0.91,0.92 and 0.95 whereas 0.93,0.88 and 0.92 in the validation.in the group of Warthin tumor and basal cell adenoma,XG Boost with F test is the most effective predicting model;in the training set,the AUC,ACC and F1 were 0.91,0.85 and 0.91 whereas 0.88,0.78 and 0.87 in the validation.De Long test showed that there were significant differences in the group of pleomorphic adenoma and basal cell adenoma between LDA/LR and RF,SVM,XG Boost(P<0.05)and no difference between LDA and LR(P>0.05);both LDA and LR were better models.There was no difference among LR,RF,SVM,LDA and XG Boost in the rest groups(P>0.05).Conclusion 1.Radiomic models can differentiate pleomorphic adenoma,Warthin tumor and basal cell adenoma effectively and efficiently.2.LDA classifier models is the most effective predicting model in the group of pleomorphic adenoma and basal cell adenoma.3.SVM and XG Boost models boost were better in differentiation of pleomorphic adenoma and Warthin tumor,Warthin tumor and basal cell adenoma.Moreover,LR,RF and LDA are also available.