Breast percent density (PD%), as measured mammographically, is one of the strongest known risk factors for breast
cancer. While the majority of studies to date have focused on PD% assessment from digitized film mammograms, digital
mammography (DM) is becoming increasingly common, and allows for direct PD% assessment at the time of imaging.
This work investigates the accuracy of a generalized linear model-based (GLM) estimation of PD% from raw and postprocessed
digital mammograms, utilizing image acquisition physics, patient characteristics and gray-level intensity
features of the specific image. The model is trained in a leave-one-woman-out fashion on a series of 81 cases for which
bilateral, mediolateral-oblique DM images were available in both raw and post-processed format. Baseline continuous
and categorical density estimates were provided by a trained breast-imaging radiologist. Regression analysis is
performed and Pearson's correlation, r, and Cohen's kappa, κ, are computed. The GLM PD% estimation model
performed well on both processed (r=0.89, p<0.001) and raw (r=0.75, p<0.001) images. Model agreement with
radiologist assigned density categories was also high for processed (κ=0.79, p<0.001) and raw (κ=0.76, p<0.001) images.
Model-based prediction of breast PD% could allow for a reproducible estimation of breast density, providing a rapid risk
assessment tool for clinical practice.