Strategies for accelerating forward and backprojection in list-mode OSEM PET reconstruction using GPUs

Image reconstruction for the ECAT HRRT PET scanner with MOLAR is computationally demanding and requires a computer cluster for reasonable run times. Parallel computing using GPUs and CUDA offers a means to accelerate MOLAR. However, forward and backprojection operations present unique challenges that must be overcome to achieve acceptable speedup. In this study we implement GPU-accelerated versions of MOLAR´s forward projection, backprojection and algorithm update modules and compare their performance to CPU-only versions. During this implementation we optimized the GPU thread configurations for each of these modules separately, along with a hybrid forward-backprojection module that is used for algorithm updates. We also numerically evaluated the reconstruction results to assess the impact of floating-point to integer conversions dictated by the GPU architecture. We found forward projection to be 41 times faster than the CPU-only code, while backprojection was 20 times faster. We found the optimal thread configurations always assigned 64 threads to a thread block, but with different distributions across the nested indexing loops within each module. These results show that MOLAR´s forward and backprojection modules can be adequately accelerated to make the MOLAR reconstruction package much more efficient.