Acceleration of image reconstruction with a ray-driven method using a GPGPU

The purpose of our research is to develop a fast image reconstruction algorithm with a ray-driven method using a general-purpose computing on graphics processing units (GPGPU). The ray-driven method based on a projection bin uses sample points that are located on the center of the bin. In the implementation with a ray-driven method using a GPU, a collision of memory accesses sometimes reduces the performance of the calculation. To avoid the collision of memory accesses, two methods were used in our proposed algorithm: one was the calculation order (the order of access to memories, which corresponded to pixels in an image matrix), and the other was the grouping of sample points that were assigned to threads in the GPU. The performance of the proposed method was compared with an image reconstruction with a CPU and that with the GPU using an atomic function, which was prepared to avoid collision. The results of the simulations confirmed the feasibility of our proposed reconstruction algorithm in the applications of the filtered backprojection method, ML-EM method and OS-EM method.