Multiresolution Bayesian ART-Type Procedure for Image Reconstruction fromProjections

摘要

The problem of reconstructing an image from measured line integrals (Radoninverse) is usually attacked on a single resolution level. The Bayesian approach, whereby the estimate of the image is approximated through iterations of the algebraic reconstruction technique, suffers in general under the poor a priori information about the mean of the distribution the image is taken from. To remedy the situation the authors propose a multilevel procedure which operates on a hierarchy of resolution levels of the image. From the solution on each level a 'fairly good' starting image (mean) is derived for the next finer level. To compare the multilevel with the unilevel approach both procedures were simulated for phantom images of different contrast type. Besides a high contrast and a continuous test pattern a head phantom with high and low contrast as mathematically defined objects was also used. Simulations were done for measurement data with and without added noise. Two interesting features of the multilevel procedure arise through local refinement. First it opens a way to dose reduction and second it enables the authors to decompose the global reconstruction problem into a number of independent local reconstruction problems of the same kind but of much smaller size. These subproblems could be solved in parallel. (Copyright (c) GMD 1992.)