|
Abstract:
|
In statistics, one way to model spatial correlation is to use the Ising model or its generalization, the Potts model. In the study of medical imaging, a variation of the Potts model, the repulsion Potts model (RPM) has been used to capture spatial relationship among voxels. To generate samples from the RPM, single site updating, for example Gibbs sampling, is easy but may mix slowly. We propose a new multi-site updating algorithm. The new algorithm is tested on both the simple RPM without an external field, and also the setting of MRI tissue classification using Bayesian hidden Markov normal mixture model, where an external field is involved. Results show that, compared with the Gibbs sampler, the newly proposed algorithm improves mixing of the Markov chains.
|
