Efficient Method for Generating System Response Functions using Monte Carlo Simulations

We have developed and tested a new method for generating imaging system matrices using a GPU-based Monte Carlo (MC) simulation package. Traditional methods of generating a system response matrix involve measurement with small sources translated throughout the field of view (FOV), or simulations of the detector responses to the photons emitted at each location. GPU-based MC simulations have the advantage that they can simultaneously simulate millions of photons at a time. In terms of system-response generation, GPU-based simulation suffers because of the inefficiencies of the data transfers and the difficulties of determining when a simulation has progressed enough to move on to the next locations. In addition, when a system matrix with a different voxel resolution is needed, this typically requires an entirely new simulation as the previous results are not transferable.We have developed a new, efficient method for using GPU MC techniques to generate system response matrices where photons are uniformly sampled from within the entire FOV. During simulation, each photon is tagged with its point of origin, and list-mode data are generated and stored instead of binned images. These data files can be quickly processed at any voxelization as long as enough photons have been simulated. Additional runs only add to the fidelity of the system response and thus, there is no wasted simulation time. The cost of this method is large data files, but we have found that these are very manageable even with desktop computers. We demonstrate this technique using for a multiple-pinhole cardiac SPECT system being developed at Mass General Hospital. Example reconstructions are shown at various resolutions all derived from a set of list-mode calibration data. We will present how this method automatically accounts for sub-voxelization, aperture sub-sampling, and penetration through thin regions of the collimator. We will also discuss the photon number requirements.