Manifold Learning for Density Segmentation in High Risk Mammograms

There is a strong correlation between relative mammographic breast density and the risk of developing breast cancer. As such, accurately modelling the percentage of a mammogram that is dense is a pivotal step in density based risk classification. In this work, a novel method based on manifold learning is used to segment high-risk mammograms into density regions. As such, finer details are present in the segmentations and more accurate measures of breast density are produced. A set of high risk (BI-RADS IV) full field digital mammograms with density annotations obtained from radiologists are used to test the validity of the proposed approach. By exploiting the manifold structure of the input space, segmentations with average accuracy of 87% when compared with radiologists’ segmentations can be obtained. This is an increase of over 12% compared with segmentation in the high-dimensional space.