PARALLEL COMPUTING WITH BLOCK-ITERATIVE IMAGE-RECONSTRUCTION ALGORITHMS

Fully discretized models for image reconstruction from projections give rise to huge and sparse nonlinear optimization problems. We study the use of parallel and vector supercomputers for the iterative reconstruction of medical images. A block-iterative version of the Multiplicative Algebraic Reconstruction Technique (MART) is implemented on a CRAY X-MP/48. The implementation exploits the block structure of the algorithm which allows us to take advantage both of the vector architecture of the computer and of the multiple processors for parallel computations. Results indicate that block-iterative algorithms are suitable for parallel and vector implementations and can reconstruct with relative efficiency highly discretized images which give rise to very large optimization problems.