TU-H-CAMPUS-IeP2-04: Efficient Binary Tree Description of High Resolution Breast Models for in Silico Imaging




We introduce a new geometric model for Monte Carlo (MC) simulation of breast tomosynthesis and mammography images with very high resolution digital breast phantoms voxelized at resolutions down to 50 microns or less. This fine sampling of the anatomy is required to reproduce the spiculated lesions and microcalcifications visible in clinical mammograms with pixel sizes under 100 microns. Phantoms voxelized at this resolution require such a large amount of computer memory that it might not be possible to execute the simulation in regular computers or in Graphics Processing Units (GPU), limiting the realism that can be achieved in virtual clinical trials.


The standard voxel-based geometry model used in MC codes is inefficient in terms of memory use, because uniform regions are described with a large number of identical voxels. We have developed a new geometric model based on a binary tree partition of space. The binary tree recursively divides the object in two equal halves, alternating the x, y, and z dimensions, until a uniform volume or a final voxel are found. A top-down search of the tree structure is used to determine the exact composition of a particular voxel during the ray-tracing process.


The binary tree geometry was implemented in PENELOPE and its GPU-accelerated version MC-GPU. An example tomosynthesis scan was simulated using a computational breast phantom voxelized at 50 micron. This phantom was too large to fit in the GPU memory directly, but its binary tree version required 10 times less memory. The time spend searching the binary tree reduced the simulation speed in half approximately.


We have introduced a new geometry model for in silico imaging of very high resolution phantoms. The binary tree model increases the computation time but, more importantly, enables the simulation of more detailed phantoms than previously possible.