The Study Of Visual Features And Radiomics Features Of Breast MRI For BI-RADS Category 4 And 5 Lesion

Objective:The aim of this study was to explore the differential diagnosis and subcategory of MRI BI-RADS category 4 lesions by using MRI visual features and radiomics features,and to explore the grouping of MRI BI-RADS category 5 lesions Materials and Methods:The T2 and DCE-MRI images of 229 patients with BI-RADS category 4 lesions(124 malignant and 105 benign)were retrospectively included,as well as the relevant clinical information.Visual features were assessed based on BI-RADS terminology and a differential diagnosis model was constructed by logistic regression.The doctors manually outlined the maximum level of the lesion and extracted the information of2 D imaging.Patients were randomly divided into training set and validation set.LASSO algorithm was used for dimension-reduction screening of features,and diagnostic models based on T2 radiomics features,CE radiomics features,and T2 radiomics combined with CE radiomics features were constructed respectively.A clinical diagnostic model was construct based on visual features.A combined model integrating radiomics features and visual features was constructed.The support vector machine algorithm was used to train and validate the model in the training/ validation set by using a five-fold cross-validation method,and performed tested in the test set.The diagnostic power of the model was analyzed using a receiver operating characteristic curve,and the clinical value was assessed using a decision curve analysis.A total of 296 patients with BI-RADS 5 lesions were retrospectively enrolled,among which 289 cases of malignant lesions were divided into three groups as HR positive group,HER2 positive group and TN group based on clinical treatment.Their T2 and DCE images and related clinical information were collected.The doctor sketched the largest slice of the lesion on the MR image and extracted the 2D image group information.Univariate analysis was used to screen the radiomics features that had significant differences among the three groups of patients,and the radiomics classification models based on T2 and CE and T2 CE MR were constructed respectively;A clinical classification model based on visual features and a combined classification model based on radiomics features and visual features were constructed,and linear discriminant analysis algorithm was used to train and verify the model by five-fold cross-validation method.Results:The BI-RADS category 4 lesions showed significant differences in age and the TIC early and delayed phases of DCE-MRI.Patients with malignant lesions are older and malignant lesions showed more rapid enhancement in the early phase,more outflow type and platform type in the delayed phase,while patients with benign lesions are younger and benign lesions showed more slow enhancement in the early phase and more gradual enhancement type in the delayed phase.The AUC of logistic regression model based on visual features was 0.872,the sensitivity was 87.10%,and the specificity was 77.14%.For the BI-RADS category 4 lesions,7 T2 radiomics features and 13 CE radiomics features were screened out by LASSO dimensionality reduction to construct the radiomics diagnostic model.The AUC of T2 radiomics model,CE radiomics model and T2 CE radiomics model in validation set were 0.663,0.717 and0.751,respectively,and in test set were 0.675,0.708 and 0.762,respectively.The AUC of the clinical model in the training set and the test set were 0.804 and 0.788,respectively.The combined model has the highest diagnosis efficiency,and the AUC of the training set and the test set are 0.859 and 0.861,respectively.DCA is confirmed that the combined model had the highest clinical benefit.For the BI-RADS category 5 lesions,the diagnostic efficiency of the differential diagnosis model of HR positive group and HER2 positive group(AUC 0.791,sensitivity 62.71%,specificity 83.18%),and the differential diagnosis model of HR positive group and non-HR positive group(AUC 0.781,sensitivity 61.86%,specificity 83.63%)were relatively better in the models constructed by visual features.For the BI-RADS category 5 lesions,11 T2 and 9 DCE radiomics features were selected to construct the radiomics model.The overall average ACC of T2,DCE,T2 DCE,clinical and combined models in validation set were 47.37%,48.77%,45.96%,50.88% and 54.74% respectively.The relative accuracy of HR positive group was the highest,and that of triple negative breast cancer group was the lowest.Conclusion:The logistic regression model of visual features and the combined model can effectively distinguish the benign and malignant lesions and subcategorizing the MRI BI-RADS category 4 lesions,and the diagnostic efficiency of combined model was better than that of clinical models or radiomics models,which has certain clinical value.Although the visual features of BI-RADS 5 lesions are different in the HR positive group,HER2 positive group and TN group,the performance of multiple diagnostic models based on MR radiomics features and visual features in BI-RADS 5lesions grouping is not good,and further exploration is needed.