Room: Room 202
Purpose: The Monte Carlo x-ray imaging simulation code MC-GPU has been modified to replicate a commercial full-field digital mammography and digital breast tomosynthesis device. We describe the state-of-the-art computational models implemented to improve the realism of the simulations. We also present a sensitivity analysis of the simulation input parameters, which is an important step in the software validation and verification.
Methods: The three main components of MC-GPU â€“ source, detector, and geometry â€“ were modified to more realistically reproduce the performance of a Siemens Mammomat Inspiration system. The source model includes now an extended focal spot with optional tube motion. The detector model represents a Selenium direct-conversion panel with a numerical anti-scatter grid, depth-of-interaction effects, x-ray fluorescence tracking, random charge conversion efficiency (Swank noise), and electronic fluctuations. The acquisition geometry reproduces a tomosynthesis scan with a static detector. A binary tree sorting algorithm was implemented to minimize GPU memory requirement to store 50-micron-voxel breast phantoms. The performance of the simulated detector model was compared to experimental measurements. A sensitivity study measuring the change in detectability of simulated lesions for a range of values of input parameters was performed.
Results: The MTF and DQE of the actual and simulated detectors are compared. The change in detectability for a Hotelling observer trained to detect spiculated masses and microcalcification clusters in mammography and tomosynthesis images of 100 simulated patients for five different exposure, energy spectra, focal-spot sizes, and motion blur values are presented.
Conclusion: The improved x-ray imaging simulator closely reproduces the performance of the replicated clinical device. The sensitivity to changes in the input parameters was evaluated. With the addition of realistic anatomical models, the MC-GPU software can produce large-scale image datasets that can be used for in silico clinical trials or training deep learning classifiers, without the need to irradiate patients.