Noise and resolution of Bayesian reconstruction for multiple image configurations

Abstract

Images reconstructed by Bayesian and maximum-likelihood (ML) using a Gibbs prior with prior weight beta were compared to images produced by filtered backpropagation (FBP) from sinogram data simulated with different counts and image configurations, Bayesian images were generated by the OSL algorithm accelerated by an overrelaxation parameter and modified by a simple averaging procedure to dampen instabilities caused by acceleration. For relatively low beta , Bayesian images can yield an overall improvement of the images compared to ML. However, for larger beta , Bayesian images degrade from the standpoint of noise and quantitation. Compared to FBP, the ML images were superior in a mean-square error sense in regions of low activity level and for small structures. Bayesian reconstruction can recover resolution without sacrificing noise performance and is dependent on the image structure and the weight of the Bayesian prior.<>