EM-ML PET reconstruction on multiple processors with reduced communications
✍ Scribed by Søren P. Olesen; Jens Gregor; Michael G. Thomason; Gary T. Smith
- Publisher
- John Wiley and Sons
- Year
- 1996
- Tongue
- English
- Weight
- 811 KB
- Volume
- 7
- Category
- Article
- ISSN
- 0899-9457
No coin nor oath required. For personal study only.
✦ Synopsis
Positron emission tomography (PET) reconstruction by the EM algorithm is an iterative computation of Poisson emission rates to maximize a likelihood function. The method is time consuming and, for real scanner data, requires large numerical arrays. To speed up the computation on multiple processors which have their own local memory and communicate by passing messages on a network, a parallel method has been implemented in which processors compute several iterations before exchanging their latest data with other processors. This method is convenient for iterative reconstruction using a relatively small number of interconnected, standard processors such as workstations on a local-area network. Computational aspects of the method are explained and illustrated with two-dimensional reconstructions from a simulation and from sinograms produced by a PET scanner. Five hundred twelve (512) iterations are computed on a local-area network of workstations and, for reference, on a distributed-memory multiprocessor computer. The method is capable of producing high-quality reconstructions with significant speed-up. 0 1996 John Wiley & Sons, Inc.
1. Introduction
In emission-tomography reconstruction by spatial statistical methods (cf., [ 1,2]), algorithms for maximum likelihood estimation of parameters tend to be iterations with long computation times involving large arrays. One way to address these problems of time and memory is to implement the iteration on multiple processors in parallel, for example, using a massively parallel computer [3,4] or taking advantage of specific interconnection topologies [5,6].
Parallel iteration can also be distributed over a network of conventional computer workstations. Distributed iteration generally reduces the memory required at any one processor by allowing each processor to store and use only portions of large, sparse matrices. Distributed iteration on multiple processors may also reduce the average time to compute one iteration; however, communication among processors is time consuming on typical networks, so an objective here is to reduce communications substantially while still producing satisfactory reconstructions.
This article describes a method for distributed-memory, itera-* Currently with the Dept. of