Acceleration of SimSET photon history generation

Abstract

SimSET (a Simulation System for Emission Tomography) is widely, used for studying PET and SPECT. As emission tomography simulation has become a more mature field, the scope of the research being performed, and thus the complexity of the simulations required, has grown immensely. Researchers are increasingly interested in clinically realistic simulations, and in some cases need hundreds or thousands of realizations. To meet these needs, we are investigating methods for accelerating SimSE'T. SimSET has always incorporated importance sampling (IS). Early studies showed the use of IS led to efficiencies 10-100 times greater than those achieved using analog (conventional) simulation. However, as the simulation became increasingly realistic the assumptions underlying the IS algorithms were violated. The efficiency improvement fell as low as a factor of two for some simulations. We are addressing this loss of efficiency by updating SimSET's algorithms, code optimization, and by modifying the software to run on multiple processors. We hope, with the new IS, to be able to simulate a 3D PET FDG brain scan (300 million detected events) in 3 hours on a 2 GHz processor. This would be a factor of 20 speedup over the currently, distributed software. To date we achieved a factor of 1.5-3 speedup by changing three algorithms and doing some code Optimization. We have several more algorithm improvements and another round of code optimization planned. We have made significant progress on parallel processing. Prototype code based on the last distributed version of SimSET achieved a speedup very close to the number of processors used. The new software also allows for multiple realizations of the same simulation to be automatically, generated on multiple processors.