Deep learning-based Breast Imaging Reporting and Data System classification and establishment of diagnostic model in breast cancer diagnosis with automated breast ultrasound

Background: Deep learning (DL) holds substantial promise for ultrasonography interpretation. However, there is limited research on the differential diagnosis of tumors in automated breast ultrasound (ABUS). This study constructs a breast cancer diagnosis model based on DL-adjusted Breast Imaging Reporting and Data System (D-BI-RADS) and determines its diagnostic performance. Methods: This retrospective study evaluated 786 patients with 832 breast nodules who underwent ABUS examination, follow-up or surgery from November 2018 to January 2022. The three-dimensional (3D) Vision Transformer (ViT) DL diagnostic models (DMs) were established and ABUS BI-RADS was adjusted to D-BI-RADS. With univariable and multivariable logistic regression, incorporating ABUS features and D-BI-RADS, three breast cancer DMs were constructed. Area under the receiver operating characteristic curve (AUC) was compared between DL, ABUS BI-RADS, and D-BI-RADS. In addition, the three DMs were verified in the validation set and test set. Results: The training, validation, and test data sets included 599 [mean age, 55±1.41 years; 351 (58.5%) malignant], 67 [mean age, 47.5±0.71 years; 40 (59.7%) malignant], and 166 [mean age, 61.5±4.95 years; 96 (57.8%) malignant] nodules. In the training data set, the D–BI-RADS showed improved performance in breast cancer diagnosis [AUC, 0.82; 95% confidence interval (CI): 0.79–0.85; P<0.01] compared with A-BIRADS. With BI-RADS category 4b (3.5) as the cut-off value for benign and malignant diagnosis, the D-BIRADS showed improved performance (AUC, 0.919; 95% CI: 0.896–0.941). In addition, Logistic, Random Forest, and Support Vector Machine (SVM) methods achieved ideal performance in the test set (AUC, 0.93; 95% CI: 0.89–0.97; AUC, 0.934; 95% CI: 0.897–0.972; AUC, 0.938; 95% CI: 0.901–0.974; respectively). Conclusions: The DL can be a useful adjunct guide to the adjustment of BI-RADS, and the established DMs achieved ideal diagnostic performance in breast cancer.