Eigenbreasts for statistical breast phantoms

Abstract

To facilitate rigorous virtual clinical trials using model observers for breast imaging optimization and evaluation, we demonstrated a method of defining statistical models, based on 177 sets of breast CT patient data, in order to generate tens of thousands of unique digital breast phantoms. In order to separate anatomical texture from variation in breast shape, each training set of breast phantoms were deformed to a consistent atlas compressed geometry. Principal component analysis (PCA) was then performed on the shape-matched breast CT volumes to capture the variation of patient breast textures. PCA decomposes the training set of N breast CT volumes into an N-1-dimensional space of eigenvectors, which we call eigenbreasts. By summing weighted combinations of eigenbreasts, a large ensemble of different breast phantoms can be newly created. Different training sets can be used in eigenbreast analysis for designing basis models to target sub-populations defined by breast characteristics, such as size or density. In this work, we plan to generate ensembles of 30,000 new phantoms based on glandularity for an upcoming virtual trial of lesion detectability in digital breast tomosynthesis. Our method extends our series of digital and physical breast phantoms based on human subject anatomy, providing the capability to generate new, unique ensembles consisting of tens of thousands or more virtual subjects. This work represents an important step towards conducting future virtual trials for tasks-based assessment of breast imaging, where it is vital to have a large ensemble of realistic phantoms for statistical power as well as clinical relevance.