A Fusion Model Based On Multi-source Data To Discriminate Benign And Malignant Breast Diseases

Part one Establishment and preliminary validation of a fusion model for differentiating benign and malignant breast diseasesPurpose:To establish a fusion model with multi-source data,including clinical breast physical examination features,traditional semantic features and radiomic features of mammography,for the identification of benign and malignant breast diseases,and to test its efficacy in an independent internal and external validation dataset.Materials and Methods:In this study,patients with pre-treatment mammography images and surgical pathology results from May 18,2009 to March 14,2018 were retrospectively collected,and they were divided into benign breast disease group and malignant breast tumor group according to surgical pathology results.Collected at the same time the patient’s clinical information including breast examination features a total of 12 items,12 semantic features and 1576 radiomic features,L1 regularization is used for our feature selection.Logistic regression was applied to construct a discriminant model.In the same way,we established three models based on clinical characters,imaging features and mammogram-based radiomic features independently.10-fold cross-validation was used to verify the reliability and repeatability of discriminant model.Patients with pre-treatment mammography images and surgical pathology results in Taizhou People’s Hospital from September 19,2015 to December 23,2019 were collected as an independent external validation dataset(Dataset 1)to validate these models.The evaluation parameters included AUC(Area Under The Curve),sensitivity,specificity,accuracy,Positive Predictive Value(PPV)and Negative Predictive Value(NPV).Results:A total of 1226 breast patients in our hospital were enrolled,including 584 benign diseases,642 malignant diseases,taizhou people’s hospital of 97 patients,45 benign patients,52 malignant patients.A total of 76 features remained made up of 62 radiomic features 8 clinical features and 6 image features after the preprocessing.The results showed that the AUC of clinical model,semantic model,radiomics model and fusion model were 0.817(95%CI:0.794～0.838),0.821(95%CI:0.798～0.842),0.815(95%CI:0.792～0.836)and 0.900(95%CI:0.882～0.917)respectively,the accuracy were 0.745,0.745,0.746 and 0.824 respectively.In dataset 1,the AUC values of the clinical model,semantic model,radiomics model and fusion model were 0.823(95%CI:0.702～0.910),0.753(95%CI:0.655～0.835),0.833(95%CI:0.744～0.901)and 0.860(95%CI:0.745～0.937)respectively,the accuracy were 0.712,0.722,0.814 and 0.864 respectively,while without significant differences in diagnostic accuracy and AUC values in the internal and dataset 1.Conclusions:In this study,a fusion model was established based on multi-source data,which was based on surgical pathology results as the gold standard.The performance of the model showed good diagnostic accuracy and high AUC value in both internal and external validation dataset.Part two The comprehensive verification of the fusion modelPurpose:To evaluate the factors affecting the diagnostic performance of the fusion model,including mammography equipment and breast density.Materials and Methods:Retrospectively collected patients with different mammography equipment of suqian first hospital on December 1,2015 to September 16,2020 with breast mammography before treatment and surgical pathologic results(dataset 2)and first people’s hospital of Changzhou on June 7,2018 to October 26,2020 with breast mammography before treatment and surgical pathologic results(dataset 3),and respectively analyzes the diagnostic performance of the model.Meanwhile,according to the ACR classification of breast density,the patients of our hospital and dataset 1-3 were divided into a,b,c and d.we classified type a and type b as non-dense breast group,type c and type d as dense breast group.We validate the model in different density of mammary gland group,effectiveness evaluation parameters including AUC value,sensitivity,specificity,accuracy,PPV and NPV.Results:Dataset 2 enrolled 118 patients,including 56 benign patients and 62 malignant patients,Dataset 3 enrolled 188 patients,including 82 benign patients and 106 malignant patients.There are 190 non-dense breast patients and 1036 dense breast patients in our hospital,16 non-dense breast patients and 81 dense breast patients in Dataset 1,34 non-dense breast patients and 84 dense breast patients in Dataset 2,44 non-dense breast patients and 144 dense breast patients in Dataset 3.In dataset 2,the AUC of the clinical model,semantic model,radiomics model and fusion model were 0.837(95%CI:0.757～0.898),0.887(95%CI:0.816～0.938),0.872(95%CI:0.798～0.927)and 0.922(95%CI:0.858～0.963),the accuracy were 0.763,0.822,0.839 and 0.864 respectively.In dataset 3,the AUC of the clinical model,semantic model,radiomics model and fusion model were 0.848(95%CI:0.788～0.896),0.904(95%CI:0.852～0.942),0.715(95%CI:0.645～0.778)and 0.906(95%CI:0.855～0.944),the accuracy were 0.798,0.851,0.723 and 0.867 respectively.The AUC values of the clinical model,semantic model,radiomics model and fusion model in the non-dense breast group were 0.811(95%CI:0.748～0.864),0.849(95%CI:0.790～0.897),0.879(95%CI:0.824～0.921)and 0.919(95%CI:0.869～0.908)respectively,and the AUC values of the clinical model,semantic model,radiomics model and fusion model in the dense breast group were 0.801(95%CI:0.775～0.825),0.799(95%CI:0.774～0.823),0.793(95%CI:0.767～0.817)and 0.890(95%CI:0.869～0.908)respectively.Conclusions:The fusion model has a certain tolerance for mammography images from different manufacturers and performs well in different breast densities.