摘要 |
Disclosed is a PET image reconstruction method based on GPU multicore parallel processing. The method comprises: (1) conducting particle sampling on voxels; (2) obtaining particle weights in a multicore parallel manner; (3) resampling particles; (4) obtaining a true value of the concentration of particles and true values of the weights thereof; and (5) calculating a voxel value of each of the voxels. The present invention defines a data model of noise in PET as a Poisson distribution rather than a Gauss distribution using a particle filter algorithm, which better conforms to the real situation in PET scanning, thereby enabling the noise filtration and optimization in a reconstruction process to be more effective, and enabling the obtained reconstruction result to be closer to the real situation in PET than traditional reconstruction methods such as ML-EM, etc. At the same time, the present invention accelerates the reconstruction process using a GPU multicore parallel processing technology to convert the calculations which would have been performed in sequence into simultaneous parallel calculations, thereby shortening the calculation time greatly, and enabling the PET image reconstruction method to be suitable for the requirements in clinical applications. |