A large-scale dataset for mammography and a model for predicting BIRADS Score

Background and aims: Breast cancer is among the most important etiologies of cancer mortalityuniversally. A breast cancer diagnosis at earlier stages provides more efficient treatments andenhances survival probability. Convolutional neural networks (CNNs), a type of deep-learningmethod, can potentially aid radiologists in the early diagnosis of images. In the current study, weaimed to assess CNNs use in discriminating malignant and benign breast lesions on mammographyand predicting the Breast Imaging Reporting and Data System (BI-RADS) score.Method: A total of ۱۰۰۰۰ mammography images from ۲۵۰۰ patients and their radiological reportswere collected. The images were gathered from Saee and Mahdieh radiological centers,Tehran, Iran. The lesions were classified into five groups according to the last version of BIRADSscores. With the help of CNN models, a deep learning model was processed to detect BIRADSfrom mammography images, and different algorithms were used to obtain the highest accuracyand specificity in diagnosis. To evaluate the accuracy of CNN models, an external dataset and aninternal dataset were assessed by both the two readers and the CNN models. Two human readersalso interpreted these test data and scored the probability of malignancy for each case using BreastImaging Reporting and Data System. Finally, specificity, sensitivity, and area under the curve(AUC) were calculated to evaluate the diagnostic function of the CNN models in predicting theBRADS scores of the lesions on images.Results: Our results showed that the average AUC of CNN models was ۰.۷۹۰ (۰.۷۱۵-۰.۸۹۵). Thefirst reader, second reader, and the best CNN model showed a sensitivity of ۰.۸۵, ۰.۷۸, and ۰.۸۹,respectively, in the ۴-۵ BIRADS group. The specificities were as follows: ۰.۸۳, ۰.۸۵, and ۰.۸۲ inthe same group. On the internal dataset, AUC was ۸۵.۷ (external ۹۲.۳) for the CNN and ۸۸.۶ ±۱.۳ (external ۹۳.۶ ± ۲.۵) for the readers.Conclusion: The diagnostic performance of CNN models in classifying the lesions according tothe BIRADS score was comparable to radiologist readers.