Bayesian Tomographic 재구성에 있어서 Gibbs Smoothing Priors의 효과에 대한 비교연구

A Comparative Study of the Effects of Gibbs Smoothing Priors in Bayesian Tomographic Reconstruction

Bayesian reconstruction methods for emission computed tomography have been a topic of interest in recent years, partly because they allow for the introduction of prior information into the reconstruction problem. Early formulations incorporated priors that imposed simple spatial smoothness constraints on the underlying object using Gibbs priors in the form of four-nearest or eight-nearest neighbors. While these types of priors, known as "membrane" priors, are useful as stabilizers in otherwise unstable ML-EM reconstructions, more sophisticated prior models are needed to model underlying source distributions more accurately. In this work, we investigate whether the "thin plate" model has advantages over the simple Gibbs smoothing priors mentioned above. To test and compare quantitative performance of the reconstruction algorithms, we use Monte Carlo noise trials and calculate bias and variance images of reconstruction estimates. The conclusion is that the thin plate prior outperforms the membrane prior in terms of bias and variance.