3D PET image reconstruction on a mesh connected multiprocessor

The authors describe a practical parallel system and highly efficient parallel algorithms for 3-D PET (positron emission tomography) image reconstruction. The proposed parallel system, a wrapped-around 2-D mesh system, is compact enough for the medical environment and scalable for various PET sizes. This system is capable of performing both analytic and iterative reconstructions efficiently. The authors use the GTTR and EM (expectation maximization) algorithms for the analytic and iterative approaches, respectively. For the parallel GTTR algorithm, a wrapped-around 2-D mesh communication pattern is proposed. For parallelization of the EM algorithm, a new task and data partitioning scheme, called partition-by-view, has been developed. The partition-by-view scheme is more efficient than schemes proposed previously since some communication can be overlapped with computation by using a new communication pattern, called the multiple ring. Both communication patterns can be embedded in a wrapped-around 2-D mesh. The parallel GTTR algorithm and the partition-by-view scheme have been implemented using image sizes of 64*64*64 and 40*40*16, respectively. High efficiency, i.e., greater than 90%, has been achieved for most of the cases tested.<>