PT  - JOURNAL ARTICLE
AU  - Brendan Vastenhouw
AU  - Changguo Ji
AU  - Frederik Beekman
TI  - Fast multi-pinhole SPECT image reconstruction with multi-core CPUs
DP  - 2010 May 01
TA  - Journal of Nuclear Medicine
PG  - 1340--1340
VI  - 51
IP  - supplement 2
4099  - http://jnm.snmjournals.org/content/51/supplement_2/1340.short
4100  - http://jnm.snmjournals.org/content/51/supplement_2/1340.full
SO  - J Nucl Med2010 May 01; 51
AB  - 1340 Objectives U-SPECT-II is a stationary multi-pinhole small animal scanner of which the pinholes can focus on a tumor or a specific organ, but it is also able to image large volumes using Scanning Focus Mode (SFM). With SFM the animal is translated through the collimator with 75 focusing pinholes. Since a large number of non-overlapping projections are involved in SFM and small voxels are required for high-resolution images, iterative image reconstruction becomes very computational intensive. Here we present methods to efficiently perform SFM reconstruction using a parallel implementation which takes advantage of the large number of processing cores of modern workstations. Methods We implemented an adapted version of OSEM that uses all projections simultaneously to reconstruct the entire volume. The same transition matrix is used for all positions while the voxel volume is shifted analog to the translation of the animal in the collimator. A special subset choice is not based on projection images, but on division of the detector pixels (POSEM, Branderhorst et al 2009). To enable the use of multi-core CPUs, the matrix is divided into different parts and assigned to different threads, which allows the calculation of the forward and backward projection in parallel. Each thread calculates part of the forward and backward projection and the partial results are summed afterwards. Performance is validated using a total body mouse bone scan. Results Traditional MLEM reconstruction for multi-pinhole SFM requires 5128 minutes for a total body mouse scan (0.375mm voxels). Using the implementation with the shifted volume combined with POSEM, the reconstruction time is reduced to 65 minutes. The parallel implementation was tested on a workstation with four quad-core CPUs, and resulted in an additional speedup of about 10×. Conclusions High resolution POSEM reconstruction of a total body mouse scan can now be performed in 6 minutes using 16 cores and can be further reduced by using more cores