Fully 3-D List-Mode OSEM Accelerated by Graphics Processing Units

Advanced list-mode image reconstruction algorithms such as fully 3D list-mode ordered-subset expectation maximization (OSEM) are needed to exploit the potential performance of high-resolution PET systems with depth-of-interaction capabilities. However, such algorithms are computationally intensive. With the aim to accelerate list-mode 3D-OSEM, we investigated the use of graphics processing units (GPUs). Primarily designed to deliver high-definition graphics for video games in realtime, GPUs are now increasingly being used as cost-effective high-performance coprocessors for scientific computing. Parallelizable algorithms such as list-mode OSEM are drastically accelerated when the 24 programmable units of the GPU are used to carry the heaviest part of the computation: the line forward and backward projections.