A new parallel EM algorithm with the optimal data replication on a hypercube multiprocessor for 3D PET image reconstruction

The authors present a new parallel expectation maximization (EM) algorithm with the optimal data replication on a hypercube multiprocessor. Although data replication has been frequently used to reduce data sharing overhead, none of the parallel EM algorithms have attempted to optimize the data replication. To maximize efficiency of the proposed parallel EM algorithm, data replication have been optimized through strict mathematical formulation and analysis in this study. To verify the theoretical analyses, the authors have implemented the proposed parallel EM algorithm on an Intel iPSC/860 and the experimental results are shown to be consistent with the theoretical results