Parallel approach to iterative tomographic reconstruction for high resolution PET imaging

Abstract

We present a parallel approach to iterative reconstruction algorithms for high resolution 2-D and 3-D PET imaging on a multiprocessor machine where processors are connected by fast Ethernet connections. An efficient parallel implementation of the ML-EM and OS-EM approaches is formulated. It makes use of pre-calculated transition matrices, utilizing the combination of four techniques for storage reduction: (1) a sparse matrix approach, storing only non-negligible values, (2) the elimination of all symmetries, (3) the elimination of matrix elements outside the aperture of the tomograph and (4) the partition of the transition matrix amongst the different processors. This approach can be used in practice for very large 2-D images, as it reduces the storage size of transition matrices per processor by a factor of 8200 for images of 64x64 up to a factor of 98000 for images of 1024x1024. A performance analysis of the parallel algorithm is presented. The parallel ML-EM approach with a pre-computed transition matrix can reconstruct 128x128 pixels images in 0.25 secliteration, as opposed to several minutes when the matrix is calculated on the fly by a single processor. Some parametrisations of the parallel OS-EM approach offer total reconstruction times of 2 seconds for 128x128 pixels images.

List-mode image reconstruction for real-time PET imaging

“…Then, only the dataset corresponding to a single projection angle could be stored and the rest may be restored online easily. This complete rotational symmetry is lacking with conventional square pixels, but even then there are a few symmetries that can be exploited [49].…”

Parallel MLEM on Multicore Architectures

“…During the 1990s, various approaches to parallelizing MLEM have been proposed [1,2,6]. They exploit the symmetries in the scanner geometry and partition the FOV and thus the image vector in a compatible way.…”