Automated Segmentation of Breast Arterial Calcifications from Digital Mammography

Abstract

Breast arterial calcifications (BACs) are formed when calcium is deposited in the walls of arteries in the breast. The accurate segmentation of BACs is a critical step for risk assessment of cardiovascular disease from a mammogram. This paper evaluates the performance of three deep learning architectures, YOLO, Unet and DeepLabv3+, on detecting BACs in digital mammography. In comparison, a simple Hessian-based multiscale filter is developed to enhance BACs pattern, then a self-adaptive thresholding algorithm is applied to obtain the binary mask of BACs. As BACs are relatively small in size, we developed a new metric to better evaluate the small object segmentation. In this study, 135 digital mammographic images containing labelled BACs were obtained, in which 80% for training deep learning networks and 20% for validation. The results show that our Hessian-based filtering method achieves a highest accuracy on validation data, and DeepLabv3+ falls behind with little effectiveness. We conclude simple filtering technique is effective in BACs extraction, and DeepLabv3+ is an expensive alternative in terms of its computational cost and configuration complexity.