Deep learning software for the automated detection and quantification of breast arterial calcifications in screening mammograms.
Cardiovascular disease (CVD) is the leading cause of death globally. One in three women in the United States die from CVD every year. Women's symptoms often differ from men's, making it particularly challenging to diagnose. Breast arterial calcification (BAC) are calcium deposits that accumulate in the arteries of the breast. Although BAC is currently not considered a clinically actionable finding, BAC has been reported as a risk marker for cardiovascular disease, wherein increased BAC correlates with the higher risk of cardiovascular disease in women. Heavy depositions of calcium lead to calcifications that appear as white spots on mammograms. Thus, quantifying BAC from screening mammograms is a non-invasive and cost-efficient approach to assess risk of CVD among women. Since millions of women undergo mammography annually, an automated approach to find the relationship between BAC and CVD would provide an opportunity to improve risk stratification without additional cost and radiation exposure.
To address this unmet need, researchers at Emory and Mayo developed a deep learning software called Multitask Learning based Context U-Met (MTL CU-Net). MTL CU-Net is designed for automatic detection and quantification of BAC in female at-risk patients for CVD. This software can process the large image size of mammograms into smaller high-resolution patches and perform image segmentation to distinguish BAC in the image using reduced training parameters. Additionally, MTL CU-Net quantifies BAC via a “calcium score”. Thus, MTL CU-Net is a robust automated method that detects and quantifies BAC and can be used during routine mammogram screening.
Inventors are currently working on using predicative scores from their BAC quantification software to compare to actual CVD outcomes.