Abstract
Breast arterial calcifications (BACs) are amongst the different types of benign calcifications observed on mammograms. BACs have been found to correlate with cardiovascular risk factors, cardiovascular mortality, and coronary artery disease (CAD). Considering that women are recommended to undergo routine screening mammography for the early detection of breast cancer, identifying BACs on mammograms could help identify women at risk of cardiovascular diseases (CVD) without the additional cost or radiation of other tests, such as coronary artery computed tomography (CT). In this paper, we present a difference of Gaussian generative adversarial network (DoG-GAN) model for segmenting BACs in mammograms. It combines the multi-scale difference of Gaussian (DoG) pyramid with the U-net as a generator of the GAN. This approach allows the model to explore image details at each scale and exploit the edge information to effectively segment BACs. We evaluated the performance of our model on a set of synthetic 2D images from digital breast tomosynthesis exams (DBT) collected and prepared for this task. The experimental results show that our model outperforms the state-of-the-art methods.
•Segmenting BACs in mammograms by using a deep learning model named DoG-GAN.•The difference of Gaussian (DoG) pyramid in the generator improves the GAN result.•DoG-GAN can extract the important and strong features from the mammogram.•Creating a new BACs ground-truth dataset annotated by expert radiologists.•DoG-GAN model results outperform the vanilla GAN, pix2pix, and DoGUnet.