Architectural distortion detection based on superior–inferior directional context and anatomic prior knowledge in digital breast tomosynthesis.

Background: In 2020, breast cancer becomes the most leading diagnosed cancer all over the world. The burden is increasing in the prevention and treatment of breast cancer. Accurately detecting breast lesions in screening images is important for early detection of cancer. Architectural distortion (AD) is one of the breast lesions that need to be detected. Purpose: To develop a deep‐learning‐based computer‐aided detection (CADe) model for AD in digital breast tomosynthesis (DBT). This model uses the superior–inferior directional context of DBT and anatomic prior knowledge to reduce false positive (FP). It can identify some negative samples that cannot be distinguished by deep learning features. Methods: The proposed CADe model consists of three steps. In the first step, a deep learning detection network detects two‐dimensional (2D) candidates of ADs in DBT slices with the inputs preprocessed by Gabor filters and convergence measure. In the second step, three‐dimensional (3D) candidates are obtained by stacking 2D candidates along superior–inferior direction. In the last step, FP reduction for 3D candidates is implemented based on superior–inferior directional context and anatomic prior knowledge of breast. DBT data from 99 cases with AD were used as the training set to train the CADe model, and data from 208 cases were used as an independent test set (including 108 cases with AD and 100 cases without AD as the control group). The free‐response receiver operating characteristic and mean true positive fraction (MTPF) in the range of 0.05–2.0 FPs per volume are used to evaluate the model. Results: Compared with the baseline model based on convergence measure, our proposed method demonstrates significant improvement (MTPF: 0.2826 ± 0.0321 vs. 0.6640 ± 0.0399). Results of an ablation study show that our proposed context‐ and anatomy‐based FP reduction methods improve the detection performance. The number of FPs per DBT volume reduces from 2.47 to 1.66 at 80% sensitivity after employing these two schemes. Conclusions: The deep learning model demonstrates practical value for AD detection. The results indicate that introducing superior–inferior directional context and anatomic prior knowledge into model can indeed reduce FPs and improve the performance of CADe model. [ABSTRACT FROM AUTHOR]