Differentiating Benign And Malignant Epithelial Salivary Gland Tumors And Tumor Subtypes Of Salivary Gland Based On Multi-parameter MRI Radiomics Model

Objective: To evaluate the utility of radiomics analysis for differentiating benign and malignant epithelial salivary gland tumors and tumor subtypes of salivary gland based on multi-parameter（T1WI,T2 WI,ADC）MRI radiomics model and construct a classification model,so as to provide relevant information for clinical diagnosis and treatment.Objective: The T1 WI,T2WI and ADC images of 296 patients with salivary epithelial gland tumors confirmed by pathology were analyzed retrospectively.238 patients with benign tumors（115 pleomorphic adenomas,46 base cell adenomas,77 warthin tumors）,and 58 patients with malignant tumors were included.Benign and malignant salivary gland epithelial tumors and salivary gland subtypes were randomly divided into training data set and testing data set in the ratio of 8:2.ITK-SNAP software was used to manually draw the region of interest（ROI）,and extract the radiomics features form ROI.2600 features in T1 WI,T2WI,ADC respectively,and 7800 features after combination were screened by analysis of variance（ANOVA）and least-absolute shrinkage and selection operator regression（LASSO）.The models are established by using logistic regression（LR）classifier,and the diagnosis efficiency of the models are evaluated.Results: 1.Diagnostic model for differentiating benign and malignant salivary gland epithelial tumors: T1 WI model,T2 WI model,ADC model and T1 WI + T2 WI + ADC model obtained 74,21,43 and154 features after feature selection.In the training data set,the AUCs of T1 WI,T2WI,ADC and T1 WI + T2 WI + ADC models established by LR algorithm are 0.909,0.879,0.940 and 0.950 respectively;The sensitivities were 0.89,1.00,1.00 and 1.00 respectively;The specificities were 0.84,0.84,0.97 and 0.89 respectively;The accuracies were 0.85,0.88,0.98 and 0.91 respectively.In the testing data set,the AUCs of T1 WI,T2WI,ADC and T1 WI + T2 WI + ADC models were0.828,0.803,0.818 and 0.899 respectively;The sensitivities were 0.73,0.70,0.55 and 0.64 respectively;The specificities were 0.83,0.82,0.98 and 1.00 respectively;The accuracies were 0.81,0.81,0.90 and 0.93 respectively.2.Diagnostic model for differentiating histological subtypes of epithelial salivary gland tumor: T1 WI model,T2 WI model,ADC model and T1 WI + T2 WI + ADC model obtained 53,29,56 and100 most valuable features respectively after feature selection.The AUCs of the training data set（AUC: 0.89-1.00）and testing data set（AUC: 0.73-1.00）of the four models are all higher,and the efficiency of T1 WI + T2 WI + ADC model is best.In the training data set,the AUCs of pleomorphic adenoma,adenolymphoma,warthin tumors and malignant epithelial salivary gland tumor were 1.00,1.00,1.00 and 1.00respectively;The sensitivities were 0.99,0.86,0.97 and 1.00respectively;The specificities were 0.96,0.99,1.00 and 0.99respectively;The accuracy was 0.97.In the testing data set,the AUCs were 0.97,0.90,1.00 and 0.97 respectively;The sensitivities were 0.78,0.78,0.93 and 1.00 respectively;The specificities were 0.97,0.90,0.98 and 0.81 respectively;The accuracy was 0.86.Conclusion: The T1 WI,T2WI,ADC and T1 WI + T2 WI + ADC models established by LR classifier have good diagnostic efficiency for differentiating benign and malignant epithelial salivary gland tumors and histological subtypes.