University of Michigan - Department of EECS
Positron emission tomography (PET) imaging provides valuable diagnostic information about physiology. Because PET measurements are very noisy, statistical methods for image reconstruction can produce better image estimates than conventional Fourier methods. Since radioactive decay is a Poisson point process, nearly all statistical image reconstruction methods have been based on a simple Poisson model for the measurement distribution. However, the conventional Poisson model is incorrect for most PET scans, due to the effects of real-time correction for accidental coincidence events. The exact statistical model for such measurements is intractable, so we have developed an approximate distribution using saddle-point methods. The proposed distribution is remarkably accurate yet tractable, i.e., suitable for implementing within a penalized-likelihood (aka Bayesian) image reconstruction method. Results demonstrate that the proposed approach leads to image estimates with significantly reduced variance relative to the conventional Poisson model, yet with negligible increase in computation time. In the talk I will also provide some background about PET, and highlight some recent "success stories" of statistical image reconstruction methods.