A Deep Learning Mammography-based Model For Architectural Distortion Benign And Malignant Differentiation

ObjectiveArchitectural distortion is a feature of a variety of benign and malignant breast lesions.Due to the lack of typical characteristics,architectural distortion has been ignored by doctors for a long time.Although mammography is currently being as an important screening tool for breast lesions,there are still some limitations such as lower displaying rate resulted by gland overlap effect,so the positive rate of architectural distortions diagnosis is not high.Hence,improving the ability to diagnosis this important feature with mammography will be a guidance for the choice of clinical treatment plans.For this purpose,this paper initially evaluated the value of deep learning model based on mammography in the classification of benign and malignant of architectural distortion.MethodsCases of breast architectural distortion confirmed by pathology from The Second Affiliated Hospital of Guangzhou University of Traditional Chi nese Medicine from January 2010 to January 2020 were collected.Clinical d ata including age,menstrual history and pathological types were recorded.Meanwhile,the details in mammography report such as BI-RADS assessment ca tegory,breast composition category and the constituent ratio of calcification were also recorded.Two mammography experts manually delineated the architectural distortion in the same case independently,and calculated the intersection over Union(IOU)to evaluate the coincidence degree of the two regions of interest(ROIs).If IOU≤0.5,the ROIs were necessary to renegotiate.Finally,the ROI were put into the model.A classification model was constructed based on the ResNeXt-50 network and the benign and malignant of pathology were used as the label to train the model.All cases were randomly divided into training set,validation set,and test set at a ratio of 6:2:2.The performance of the model with indicators such as Area Under Curve(AUC),accuracy,sensitivity,specificity,and F1 score.ResultsA total of 349 cases of architectural distortion of mammography were included in this study with a median age of 48 years.There were 190 cases in the malignant group,the median age of the patients was 49(43-56)years,and the number of menopausal was 102.There were 1 case of type a(almost entire fat),38 cases of type b(scattered fibroglandular tissue),146 cases of type c(heterogeneous fibroglandular tissue)and 5 cases of type d(extremely fibroglandular tissue)of the breast composition categories in malignant group.There were 6 cases of category Ⅱ,1 case of category Ⅲ,3 cases of category Ⅳa,3 cases of category Ⅳb,67 cases of category Ⅳc and 110 cases of category V of the BI-RADS assessment in malignant group.115 malignant architectural distortion were combined with the signs of calcification.There were 159 cases in the benign group,the median age of the patients was 49(43-56)years and the number of menopausal was 62.There were 2 case of type a,19 cases of type b,126 cases of type c and 12 cases of type d of the breast composition categories in benign group.There were 10 cases of category Ⅱ,8 case of category Ⅲ,15 cases of category Ⅳa,2 cases of category Ⅳb,115 cases of category Ⅳc and 9 cases of category V of the BI-RADS assessment in benign group.91 benign architectural distortion were combined with the signs of calcification.there were no significant differences in midian age and the constituent ratio of calcification between the two groups(P>0.05),while there were significant differences in constituent ratio of menopausal、breast composition category and BI-RADS assessment category(P<0.05).In the test set,the deep learning model had an area under the receiver operating characteristic curve(AUC)of 0.61,an accuracy of 0.64,a F1 score of 0.60,a specificity of 0.81,a sensitivity of 0.49 to classify benign and malignant of architectural distortions.ConclusionsThe deep learning model based on mammography had a certain value in the diagnosis of architectural distortion of benign and malignant.The model had high specificity and could reduce the misdiagnosis of benign architectural distortion.It is expected to become an important auxiliary diagnostic tool for doctors.