Adaptively altering hyper-parameter for improved reconstruction in PET

It is well known that as the iterations of the maximum likelihood algorithm are continued, density estimates become more and more noisy. In situations where some prior knowledge about the object distribution is available, one may utilize such information in the reconstruction procedure for improving the reconstruction. Fixed prior based image reconstruction process produces over-smooth reconstruction due to the penalizing nature of the potential. As the reconstruction process builds up, image properties like smoothness, frequency content etc., change and hence fixed prior based image reconstruction process may not serve the purpose. A new prior is proposed which is capable of producing improved reconstruction over those obtained by existing fixed prior based Bayesian algorithms. These are termed as dynamic priors, which unlike other priors modify itself recursively according to the noise level in the reconstruction. It is found that inclusion of prior knowledge in the reconstruction algorithm results in local minimums. In the present approach, appropriate prior energy is supplied to the estimate to overcome local minimums by scheduling Gibbs hyperparameter and subsequently the effect of prior is removed recursively as the estimate approaches global minimum.