Breast density mapping based upon system calibration, x-ray techniques, and FFDM images

Abstract

Clinical studies have correlated a high breast density to a women's risk of breast cancer. A breast density measurement that can quantitatively depict the volume distribution and percentage of dense tissues in breasts would be very useful for risk factor assessment of breast cancer, and might be more predictive of risks than the common but subjective and coarse 4-point BIRADS scale. This paper proposes to use a neural-network mapping to compute the breast density information based upon system calibration data, x-ray techniques, and Full Field Digital Mammography (FFDM) images. The mapping consists of four modules, namely, system calibration, generator of beam quality, generator of normalized absorption, and a multi-layer feed-forward neural network. As the core of breast density mapping, the network accepts x-ray target/filter combination, normalized x-ray absorption, pixel-wise breast thickness map, and x-ray beam quality during image acquisition as input elements, and exports a pixel-wise breast density distribution and a single breast density percentage for the imaged breast. Training and testing data sets for the design and verification of the network were formulated from calibrated x-ray beam quality, imaging data with a step wedge phantom under a variety x-ray imaging techniques, and nominal breast densities of tissue equivalent materials. The network was trained using a Levenberg-Marquardt algorithm based back-propagation learning method. Various thickness and glandular density phantom studies were performed with clinical x-ray techniques. Preliminary results showed that the neural network mapping is promising in accurately computing glandular density distribution and breast density percentage.