Statistical Inference of Parameters of a 3D Solid Breast Texture Model from Clinical Data

In this paper, we investigate the statistical inference of the parameters of a 3D mathematical breast texture model, previously introduced for virtual clinical trials (VCT) in x-ray breast imaging. To improve the anatomical variability of the 3D texture model, a two-step inference method, referred to as inference from reconstruction, is proposed, allowing the estimation of medium scale texture model parameters from clinical breast computerized tomography (bCT) reconstructed volumes. A multiple birth, death and shift algorithm was employed to first reconstruct a set of random ellipsoids from each ground truth bCT volume. Result of the reconstruction step indicates that the medium scale intra-glandular adipose compartments can be modeled as systems of random ellipsoids exhibiting clustering interations. Then, during the inference step, the Matern cluster process was fitted to the reconstructed ellipsoid centers using the minimum contrast method based on the pair correlation functions. Distributions of the ellipsoid shapes were finally estimated from their empirical distributions. Finally, 12 sets of new medium scale model parameters were obtained from 12 bCT volumes. Visual evaluation of simulated 2D and 3D breast images using the new texture model parameters shows fairly high visual realism and improved morphological variability compared with images simulated from previous prototype using empirical parameters. Recently published studies demonstrate the realism of 2D and 3D breast images simulated from the new texture model, as well as successful applications of the new texture model in clinically relevant VCT tasks.