| ObjectiveTo evaluate the value of US radiomics,CEUS radiomics and the combination of them in the diagnosis of benign and malignant breast tumors.Materials and Methods166 patients(170 lesins,including 49 benign lesions and 121 malignant lesins)with complete surgical pathological results were selected,which treated at the Cancer Hospital and Institute,Peking Union Medical College and Chinese Academy of Medical Science from March 2019 to January 2020.The patients were examined for US and CEUS before the surgery.Two ultrasound doctors manually sketched the ROI of obtained images,and then the radiomics features were extracted from the ROI and analyzed.To diagnose the benign and malignant breast tumors,multivariate logistic regression models were constructed in US,CEUS and US combined with CEUS separately.The features used for those models were selected by the mRMR algorithm.Finally,the performance of the diagnostic model was evaluated by five-fold cross-validation method.ResultsRadiomics features of US and CEUS extracted from breast tumors were first selected by the method of mutual information to do single feature analysis.The results showed that the texture features of the filtered images has the strongest correlation in the benign and malignant identification.Then,the US,CEUS and US combined with CEUS radiomics multivariate logistic regression models constructed by mRMR algorithm all have good diagnostic performance in the differential diagnosis of benign and malignant breast tumors,and both better than aforementioned single radiomics feature.Specially,the diagnostic performance of US model is better than CEUS model,while the US combined with CEUS model has the best diagnostic performance.The average AUC of the training set is 0.953.In the evaluation on the test set,the accuuracy is 0.852,sensitivity is 0.869,specificity is 0.808,positive predictive value is 0.918,and negative predictive value is 0.715.Conclusions1.US and CEUS radiomics have certain value in the diagnosis of benign and malignant breast tumors,which could help ultrasound doctors to find the hidden texture features of images,and then assist in diagnosis;2.US combined with CEUS radiomics model can further improve the diagnostic accuracy of US radiomics model in identifying benign and malignant breast tumors,reflecting the effective auxiliary diagnosis value of CEUS radiomics,which have certain clinical application value.ObjectiveTo investigate the predictive value of US,CEUS and US combined with CEUS radiomics in molecular subtypes of breast cancer.Materials and Methods119 patients(120 breast cancer lesins)with complete immunohistochemical results were selected,which operated at the Cancer Hospital and Institute,Peking Union Medical College and Chinese Academy of Medical Science from March 2019 to January 2020.According to the results of immunohistochemistry,they were divided into luminal A and non-luminal A,luminal B and non-luminal B,her-2 overexpression and non-her-2 overexpression,TNBC and non-TNBC,HR positive and negative,her-2 positive and negative groups.The patients were examined for US and CEUS before the surgery.Two ultrasound doctors manually sketched the ROI of obtained images,and then the radiomics features were extracted from the ROI and analyzed.To evaluate the predictive value of the aforementioned 6 groups of breast cancer,multivariate logistic regression models were constructed in US,CEUS and US combined with CEUS separately.The features used for those models were selected by the mRMR algorithm.Finally,the performance of the diagnostic model was evaluated by five-fold cross-validation method.Results1.In single feature analysis,the log1.5_glrlm_RunLengthNonUniformity value in CEUS radiomics features is significantly different between luminal A and non-luminal A group(p<0.05);The glszm_HighGrayLevelZoneEmphasis value in US radiomics features is significantly different between TNBC and non-TNBC group(p<0.05);And the log2.0_glrlm_RunLengthNonUniformity value is significantly different between the HR positive and negative group(p<0.05).2.The accuracy of US and CEUS radiomics multivariate logistic regression model in predicting breast cancer molecular subtypes are all above 0.7.Among them,the CEUS model has improved the accuracy of the US model in predicting luminal A,luminal B,HR positive and her-2 positive breast cancer.However,the accuracy of the CEUS model in predicting her-2 overexpression breast cancer and TNBC is similar to the US model.In addition,the US model has a high accuracy in predicting her-2 overexpression breast cancer and TNBC(0.841,0.867),which are all above 0.8.Conclusions1.The texture feature based on US and CEUS radiomics have the potential to become imaging biomarkers that predict breast cancer molecular subtypes independently.2.US radiomics multivariate logistic regression model has strong ability in predicting her-2 overexpression breast cancer and TNBC.3.In the prediction of luminal A,luminal B,HR positive and her-2 positive breast cancer,the multivariate logistic regression model of US combined with CEUS radiomics has the best performance,means that CEUS radiomics could improve US radiomicis with certain auxiliary predictive value,and provides a new way for the evaluation of breast cancer molecular subtypes before surgery. |
