Optimizing communications for parallel ML-EM image reconstruction on large clusters of processors

The ML-EM reconstruction algorithm is an iterative computation of tomographic images by maximization of a likelihood function of the Poisson emission rates. The method being time consuming, a parallel version can easily be implemented to distribute the reconstruction task on multiple processors. Since general-purpose supercomputers are costly and their parallel programming are platform-dependent, distributed ML-EM processing on a network of workstations is an attractive alternative. However, communications of the computed partial projections among processors are time consuming on typical networks such that there is a relatively low limit on the number of processors which can participate in the computation and still reduce reconstruction time. Optimizing the communication process by using efficient hardware and message passing interface, and by transferring sums of partial projections instead of the partial projections themselves over the network permits the distribution of the reconstruction task over a larger number of processors while reducing the overall reconstruction time.