With the ever-increasing dependence on cone-beam CT (CBCT) for clinical decision-making, faster and more accurate CBCT simulation tools are necessary to develop the next generation of CBCT techniques and algorithms. GECCO, a hybrid GPU-Monte Carlo (MC) software toolkit was built for the dual purpose of increasing both speed and accuracy in CBCT simulation. GECCO contains two applications: The GECCO application, a rapid converter of CT to CBCT, and the real-time image simulation (RTIS) application, a purely analytical approach that updates CBCT images in real-time according to acquisition technique. The GECCO application calculates primary projections analytically and scatter with GPU-MC. The analytical model is a faster, less memory-intensive update to the Fastcat simulation toolkit, combining mono-energetic projections into a polyenergetic projection. Scatter is calculated using a modified version of GGEMs GPU-MC, denoised, and interpolated to reduce the number of MC histories needed for a full simulation. Conversely, the RTIS application is solely analytical and maximizes speed, only simulating a single CT slice and storing partial results to enable dynamic updating of the simulation. Validation compared GECCO simulations of an XCAT phantom, a virtual Catphan phantom, and an elementally decomposed planning CT of a central nervous system (CNS) patient to GGEMs MC projections, TrueBeam CBCT of a Catphan, and a registered CNS patient CBCT, respectively. The GECCO application took 63 min to simulate 915 full-resolution projection CNS patient CBCT on an NVIDIA 4090 GPU. The virtual Catphan sensitometry module showed an RMSE of 5.7 HU compared to the TrueBeam on-board imager (OBI). Likewise, RTIS had an initialization time of 917 ms and an update time of 36 ms, three orders of magnitude faster than any other simulation application. Profiles through the CNS patient showed 14 and 16 HU RMSE as compared to OBI for GECCO and RTIS, respectively. We introduce the GECCO software toolkit, containing the GECCO application for producing high-quality machine learning datasets and the RTIS application for real-time selection of imaging techniques for radiation therapists.
O’Connell et al. (Fri,) studied this question.