Millisecond SPECT image reconstruction acceleration using OSEM algorithm based on CPU / GPU hybrid parallel hardware platform

Currently SPECT image reconstruction algorithm mainly includes analytical method represented by FBP and iterative method represented by OSEM. The OSEM reconstruction algorithm can obtain better reconstructed image quality, at cost of reconstruction computation time overhead. In order to improve the computational efficiency of OSEM algorithm and to further promote the clinical application of OSEM algorithm, this work focus on the research of parallel computing applications in OSEM algorithm.In this work, we first analyze and exploit the general principles and methods of parallel computing applications on OSEM algorithm, and then implement OSEM algorithm on multi-core CPU, GPU, CPU / GPU hybrid parallel hardware platform, including some further targeted parallel acceleration algorithm combined with each hardware platform. Finally, we performance testing of optimized parallel algorithms on each hardware platform respectively, with comparison and analysis. Test is implemented on CPU, GPU, CPU / GPU platform respectively of various condition for optimization. A slight distinction emerged between reconstruction image implement on CPU and GPU. Performance of acceleration per iteration using OSEM contains: 1)3.987s on CPU with 17.54 time speed-up; 2)1.100s on GPU with 63.54 time speed -up 3)0.610s on CPU/GPU hybrid platform with 114.58 time speed -up. Test results with multiple sets of experiment show that CPU / GPU parallel hybrid algorithm get the best performance with reconstruction time of 0.61s per iteration, leading to a significant implement of OSEM algorithm for clinical application.