Research On The Prediction Model Of Benign And Malignant Breast Tumors Based On Ultrasound Radiomics

Objective Based on conventional two-dimensional ultrasound images,the prediction model of benign and malignant breast tumors was established by using radiomics method to explore the value of ultrasound radiomics in the diagnosis of benign and malignant breast tumors.Methods A total of 424 lesions were collected from 414 women with pathologically confirmed breast tumors from June 2020 to June 2021.The cross-sectional images of the largest diameter section of the lesion were recorded and stored in DICOM format.Patients were randomly divided into a training set（n=254）and a validation set（n=170）in a 3:2 ratio.Sonographer used ITK-SNAP software to segment the region of interest（ROI）of breast tumor.Ultrasound radiomics features were extracted by Pyradiomics software and screened by random forest（RF）.Independent predictors were selected by univariate and multivariate analysis of two-dimensional ultrasound characteristics of breast tumors.For the ultrasound radiomics features and two-dimensional ultrasound features,Logistic Regression（LR）,Support Vector Machine（SVM）,Gradient Boosting Decision Tree（GBDT）and Random Forest（RF）machine learning algorithms were used to establish three different prediction models in the training set respectively,which including ultrasound radiomics features model,two-dimensional ultrasound feature model and ultrasound radiomics features+two-dimensional ultrasound feature comprehensive model.The Receiver Operating Characteristic Curves（ROC）of each model were drawn,and the area under ROC curve（AUC）,accuracy,sensitivity and specificity of each model were calculated.Finally,the same methods were used to test the predictive performance of the radiomics model in an independent validation set.Results 1.A total of 424 breast masses from 414 women were included in this study,of which 200 were pathologically diagnosed as benign and 224 as malignant;2.Multivariate Logistic regression analysis showed that ultrasound features,including morphology,margin and micro-calcification,were independent predictors of breast cancer（P<0.05）;3.In this study,a total of 107 radiomics features were extracted,and 10 effective features were selected to construct the model after random forest screening;4.For the ultrasound radiomics model,the AUC of LR,SVM,GBDT and RF models in the training set were 0.917、0.939、0.996、1.000,respectively.In the validation set,they were 0.886、0.898、0.870、0.864,respectively.For the two-dimensional ultrasound feature model,the AUC of LR,SVM,GBDT and RF models in the training set were0.862、0.864、0.875、0.875,respectively.The validation sets were 0.875、0.882、0.874、0.874,respectively.For the comprehensive model,the AUC of LR,SVM,GBDT and RF models in the training set were 0.958、0.957、0.979 and1.000,respectively.In the validation set,they were 0.949,0.948,0.942 and 0.901,respectively.Conclusions In this study,based on the ultrasound images of breast tumor patients,four different machine learning algorithms were used to construct the ultrasound radiomics model,two-dimensional ultrasound model and the comprehensive model of the two,respectively.All the models showed good performance in differentiating benign and malignant lesions,and the model based on LR and SVM algorithm was more stable,and SVM was slightly better than LR.In addition,compared with the ultrasound radiomics model and the two-dimensional ultrasound model,the comprehensive model has better diagnostic performance and is more effective in distinguishing benign and malignant breast tumors.