The Role Of Magnetic Resonance Based Radiomics In The Classification Of Parotid Diseases

Objective:To construct and evaluate a diagnostic radiomics model based on conventional MRI and machine learning algorithms in parotid diseases classification.Methods:Clinical data and preoperative MRI images,including T1WI、T2WI and CET1 WI,were collected from parotid diseases patients who underwent surgical treatment at Xiangya Hospital of Central South University and Hunan Cancer Hospital between January 2010 and September 2020.A total of 298 patients were included and randomly divided into training and test sets in a 7:3 ratio.ITK-SNAP software was used for layer-by-layer outlining of ROI on three sequences of MRI images."Py Radiomics" module was used for radiomics feature extraction.Feature filtering was subsequently performed using the SKB and LASSO algorithms.A three-step,four-classification machine learning model was constructed using XGBoost,SVM and DT,aiming to classify parotid tumors into four subtypes: malignant tumor(MT),pleomorphic adenoma(PA),Warthin tumor(WT)and other benign tumor(OBT).ROC and AUC were used as performance metrics.The diagnostic confusion matrixes for these models in the test set were then calculated and finally compared with the radiologist’s diagnostic confusion matrix.Results:No significant differences in clinical characteristics were seen between the training and test sets.The three steps were filtered to obtain 6,12 and 8 optimal features,respectively.In the training set,the AUC values were 0.857,0.882 and 0.908 for XGBoost,0.809,0.819 and 0.832 for SVM and 0.73,0.795 and 0.889 for DT for the three steps,respectively.In the test set,the AUC were 0.826,0.817 and 0.789 for XGBoost,0.783,0.833 and 0.821 for SVM,and 0.605,0.631 and 0.627 for DT for the three steps,respectively.In the diagnostic confusion matrixes,the overall accuracy of the prediction models using XGBoost,SVM,and DT were 0.708,0.596,and 0.461,respectively,and the overall accuracy of the radiologists was 0.492.Conclusion:A three-step,four-classification model based on SVM and XGBoost algorithms combined with conventional MRI radiomics can be used for the initial classification of parotid tumors,among which the XGBoost model outperforms SVM model.The three-step,four-classification models constructed in this study outperforms radiologists in the four-classification task of parotid tumors,which has the potential to be an effective diagnostic aid in clinical practice.They will contribute to the initial screening and staging of parotid diseases.