2014 Volume 32 Issue 2 Pages 109-115
3D list-mode PET image reconstruction using TOF information has advantages in size of acquisition data and accurate reflection of TOF information to image in comparison to 3D sinogram reconstruction. However, computation time of list-mode reconstruction increases in proportion to the number of events. In this work, we investigated acceleration of 3D list-mode image reconstruction using CUDA (Compute Unified Device Architecture), especially forward and back projections which were the dominant computational load. We sorted list-mode data by main direction (largest LOR's direction cosine) for efficient memory access, and used texture memory to store TOF kernel function for reducing calculation amount. The reconstructed images computed by both CPU (8-cores, MPI) and GPU are almost identical, while our approach runs 7.8 times faster than an equivalent CPU implementation.
