E4915. Breast Arterial Calcification on Mammograms: Correlations to Cardiovascular Risk and Deep Learning
  1. Aneka Khilnani; George Washington University
  2. Akanshka Mohan; George Washington University
Breast arterial calcification (BAC) has emerged as a potential women-specific risk marker for cardiovascular (CV) risk. Although BAC presents as a medial calcification of the arteries, notably different from the intimal atherosclerotic process, BAC is associated with an increased risk (1.23) in CV disease in postmenopausal women and has higher diagnostic accuracy than other traditional CV risk factors in asymptomatic middle-aged women, especially under 60 years of age. The widespread use of mammography promotes the leveraging of BAC detection for further CV risk stratification in women undergoing breast cancer screening. A robust quantification is indispensable for an application in prevention; however, BAC appears in different shapes and sizes, as elongated paths or short inconsistent paths, and can vary considerably. There have been many attempts to find an accurate method to quantify and detect BAC from mammograms.

Educational Goals / Teaching Points
We evaluate the current literature on the association of BAC- detection by deep learning quantitative scoring algorithms and CV risk.

Key Anatomic/Physiologic Issues and Imaging Findings/Techniques
Wang et al. was among the first to utilize deep learning to detect BAC. A 12-layer CNN architecture was considered pixel-wise, and a patch-based approach was used to detect BAC. To detect BAC, thresholding and morphology operations are applied, and the resultant BAC is identified as an overlapping region, resulting in a 62.61% accuracy. Several other studies have used U-Net with numerous disease applications. AlGhamdi et al. proposed a U-Net model for the automatic detection of BAC in mammograms, combining both short skip (stop the model from learning unwanted features) and long skip (recover information that was lost during encoding) connections, yielding an accuracy of 91.47%. Guo et al. utilized Simple Context U-Net (SCU-Net), which also focuses on segmentation of vessels, resulting in a 95% correlation between the predicted mask of SCU-Net in comparison to measurement of BAC on CT.

Despite all these approaches, the accurate detection of BAC from mammography is still an unresolved issue, such as the challenge of deep learning models in detecting small vessels. Thus, severity detection remains a challenge. The narrow and varied appearance of BAC creates challenges for deep learning. Moreover, there are no large, annotated datasets of BAC available, and the use of large images makes the processing significantly difficult. The training of deep learning models from scratch with such limitations makes the problem challenging to solve.