A Fast Planar PET Image Reconstruction Method Based On GPU

With the widespread application of planar positron emission tomography(PET)imaging system in clinical application and scientific research,researchers have put forward new requirements for the quality of image reconstruction.In order to obtain high quality reconstruction images,the depth of interaction(DOI)and other physical processes become the first issue to be solved.Based on Monte Carlo(MC)simulation,the system response matrix(SRM)fully considers the physical process of photon detection,and has great performance on eliminating the DOI effect.However,the SRM based on MC simulation is usually accompanied by high storage requirement,which will lead to an amount of time consumption when reading the huge SRM during the reconstruction process.Therefore,how to compress the SRM and accelerate the reconstruction process has great value to research.In this thesis,we first introduced the iterative reconstruction algorithm,the PET error influencing factors and the system response matrix.Then we briefly described how to compress the SRM using the symmetry of the planar PET system and perform the fast reconstruction in the graphic processing unit(GPU)based on the compressed SRM.Finally,the validity of the algorithm was verified by experiments,and its acceleration performance was evaluated.The main work of the thesis is as follows:(1)The SRM is compressed and stored using the symmetry of the planar PET geometry.Based on the geometric structure of planar PET,the description of planar PET SRM based on line of response(LOR)symmetry and voxel symmetry was studied respectively.By using LOR symmetry or voxel symmetry,only a subset of the system response matrix was required to characterize the overall SRM,and the storage scale of SRM was greatly reduced(approximately 400-600 times),which also made the time consumption of reading the huge SRM reduced significantly.(2)Accelerate the PET reconstruction based on the symmetry of the SRM and the use of compute unified device architecture(CUDA).Using the LOR-based SRM subset,the LOR-based accelerating reconstruction algorithm was realized by using the symmetry of LOR.Using the voxel-based SRM subset,the voxel-based accelerating reconstruction algorithm was realized by using the symmetry of voxels.Considering the characteristics of the two algorithms,the two algorithms were combined together.The LOR-based algorithm was in charge of the forward projection,and the voxel-based algorithm was responsible for the back-projection operation.Thus,the LOR and voxel union accelerated reconstruction algorithm could be realized.(3)Verify the validity of the algorithms and evaluate the acceleration effect by experiments.The experimental results show that the reconstructed results are the same before and after the algorithm acceleration.The computational speed of the algorithm based on LOR and voxel union is 3 times faster than that of LOR-based accelerating reconstruction algorithm,12 times faster than that of voxel-based acceleration reconstruction algorithm and 113 times faster than that of image reconstruction algorithm before acceleration.