A systolic implementation of the MLEM reconstruction algorithm for positron emission tomography images

摘要

This work is part of a project concerned with parallel reconstruction of 3D images from the data delivered by positron emission tomography. The intention of this investigation is the development of a highly efficient parallel-computer assisted PET diagnosis system. Due to the Poisson-nature of positron emission processes, likelihood models can be used for their description. The investigations undertaken are based on an expectation-maximum algorithm that iterates to a maximum likelihood of an estimated radiosity image. A parallel imple- mentation of this algorithm is adapted to the modular architecture of a Quadrics DM-SIMD parallel-computer that can be tailored for special applications. Two approaches are described. The “slice-wise” approach calculates l5 slices of a 3D data set in parallel, one slice per pro- cessor. The systolic approach provides more precise calculations based on direct matrix-vector multiplication while making more efficient use of the Quadrics architecture. Problems are discussed that are raised from the definition of a huge and sparse transition matrix during systolic computation.