Deep Learning for Breast Region and Pectoral Muscle Segmentation in Digital Mammography

Abstract

The accurate segmentation of a mammogram into different anatomical regions, such as breast or pectoral muscle, is a critical step in automated breast image analysis. This paper evaluates the performance of u-net deep learning architecture on segmenting breast area and pectoral muscle from digital mammograms and digital breast tomosynthesis. To minimise the image variations due to vendor and modality specifications, Volpara$^\text {TM}$ algorithm was used to normalise the raw image to a unity representation that is independent of imaging conditions. Four factors and their interactions were investigated for their effects on the performance of u-net segmentation: image normalisation; zero and extrapolated padding techniques for image size standarisation; different contrast between breast and background; and image resolution. By training u-net on 2,000 normalised images, we obtained median dice-similarity coefficients of 0.8879 and 0.9919, respectively for pectoral and breast segmentations from 825 testing images. The model training speed was boosted by using down sampled images without compromising segmentation accuracy. Using normalised breast images by Volpara$^\text {TM}$ algorithm, u-net was able to perform robust segmentation of breast area and pectoral muscle.