Development of GPU based image reconstruction method for clinical SPECT

The aim of this work is to develop, validate and preliminarily evaluate two graphic processing units (GPU) implementations of the OS-EM iterative image reconstruction algorithms for a clinical SPECT system. The Monte Carlo software SIMSET was used to simulate the data acquisition of a typical SPECT camera with a parallel-hole collimator for the Shepp-Logan phantom. CUDA was adopted for the GPU programming of the OS-EM algorithm with high complexity. For the first implementation, the system matrix used for reconstruction was pre-computed by Monte Carlo simulation. A major task in the GPU programing in this case was to efficiently access every element of the pre-stored system matrix and compute the projection and back-projection for each image update. In the second implementation, the system matrix was calculated on-the-fly. A rotation-based projector and back-projector was employed, which reduced the complexity for the GPU programming and can achieve higher efficiency theoretically than other methods of computing projector and back-projectors. We modeled the depth-dependent collimator detector response in the matrix computation. The results showed that, for both GPU implementations, there are seldom differences of images reconstructed by the GPU and CPU implementation of the OS-EM algorithm. In the first matrix computation strategy, an acceleration ratio of 31.9 was achieved by using the GPU implementation compared with the CPU implementation of the same algorithm. In the second strategy, the acceleration ratio was 33.9. The images reconstructed with the first strategy demonstrated higher quantitative accuracy due to the more accurate modeling of the system matrix. But the corresponding reconstructed time was more than the second strategy. We also applied the second GPU implementation of the OS-EM method to clinical SPECT patient cardiac data and achieved the reconstruction of a 64×64×64 image within 8 seconds. Compared with the FBP reconstruction image, the OS-EM reconstruction image had a better performance. In conclusion, GPU implementation of the OS-EM algorithm has a great potential in SPECT image reconstruction.