Positron emission tomography(PET) is a noisy functional medical imaging model. In this paper a fully Bayesian PET reconstruction method is presented for combining a segmented anatomical membrane a priori. The segmented anatomical membrane a priori is based on the fact that the radiopharmaceutical activity is similar throughout each region and the anatomical information can be obtained from other imaging modalities such as CT or MRI. The prior distributions are formed as some kind of Markov random field. Due the non convex and the hyper parameters in the prior

it is difficult to use point estimator such as maximum a posteriori(MAP). So we used Dynamic Markov chain Monte Carlo posterior simulation method to get a minimum mean square error(MMSE) estimator which update the hyper parameters as well as density data. The variances and credit area of the reconstruction results can be easy gotten by MMSE. We compared the reconstruction result of ML

MAP and MMSE

and find that the segmented anatomical membrane a priori exhibit improved the noise and resolution properties and Dynamic Markov chain Monte Carlo is mostly suitable for fully Bayesian reconstruction.