Parallelization of 3-D PET BpjF reconstruction on a DSP cluster

Abstract

This work presents the parallelization of Three-Dimensional PET Backprojection Filtering reconstruction with reference to an implementation on a multiprocessor architecture organized in cluster mode. The authors took as reference a 24 rings, 784 detectors per ring, scanner, and a corresponding volume of interest of 128*128*47. The multiprocessor architecture is based on six commercial digital signal processors connected to an interface processor and global memory on a shared bus and together via their port links. In the Three-Dimensional backprojection stage the volume of interest data are stored within processors local memories, and are accessible from other processors. LORs data are dynamically scheduled onto the processors providing a good load balance and a low communication overhead. The Three-Dimensional filtering has been considered in the frequency domain using Fast Fourier Transform techniques on an enlarged array of samples in order to take into account border artefacts drawbacks. The Backprojection Filtering reconstruction time for a 3-D acquisition on this apparatus takes, according to the authors' simulations, approximately five minutes.