GPU-accelerated motion compensated OSEM list-mode PET reconstruction using a time-averaged sensitivity matrix

Abstract

We developed graphic processing unit (GPU) code for motion compensated PET reconstruction with a list-mode ordered subsets expectation maximization algorithm. The motion was modeled by applying serial 3D rigid transformations, one per acquired pose, to the sensitivity matrix at time zero, and computing its motion-interval weighted average. The GPU implementation was about 50 times faster for motion modelling and about 4 times faster for reconstruction than an equivalent single-CPU implementation. While excellent acceleration of the motion modelling was achieved, the reconstruction performance of the present GPU implementation was only equivalent to a quad core processor. With the aim of achieving a significant improvement in performance, we optimized the CPU including increasing the use of cache memory instead of on-chip shared memory, memory coalescing and better adapting the CPU architecture to the problem and found the acceleration rate of 7.5, which is equivalent to a multicore processor with 8 cores. Our GPU implementation is as fast as 8 cored CPU and needs further improvement using either multi-CPUs or distributed/shared memory computing to get more acceleration.