A phantom study of the quantitative behavior of Bayesian PET reconstruction methods

The authors examine the behavior of a Bayesian protocol for reconstruction of PET images in terms of the quantitative accuracy of the resulting Region-of-Interest (ROI) statistics. The authors´ protocol uses a rapidly converging algorithm to compute a MAP estimate of the 3D PET image and includes novel methods for accurate attenuation and scatter modeling. The image is modeled using a Gibbs prior whose hyperparameters are estimated simultaneously with the PET image so that no critical user specified parameters are required. The authors computed ROI statistics over reconstructions of multiple frames of a computer generated brain phantom and an experimental chest phantom scanned using a Siemens/CTI ECAT931 scanner. Results indicate significant improvements in both bias and variance when compared to a standard filtered backprojection (FBP) based protocol.