Bayesian PET image reconstruction with an anatomically adaptive nonlocal prior

The incorporation of registered anatomical image as a prior to guide PET image reconstruction has been reported in many previous studies. Based on the nonlocal means filter and the regional information of anatomical image, an anatomically adaptive nonlocal prior (AANLP) is proposed. The information in this prior model comes from weighted differences between pixel intensities with in a nonlocal neighborhood. The weights of each pixel depend on its similarity with respect to the other pixels. The regional information of anatomical image is used to estimate a smoothing parameter which controls the decay of similarity function. This prior is determined and applied adaptively to each anatomical region on the PET image for the iteration in the reconstruction process. A two-step reconstruction scheme using the AANLP is proposed to update the image and estimate the parameter. The simulation results show that the AANLP reconstruction can dramatically preserve the edges and yield overall higher lesion-to-background contrast.