Conditional entropy maximization for PET

Maximum Likelihood (ML) estimation is extensively used for estimating emission densities from clumped and incomplete measurement data in Positron Emission Tomography (PET) modality. Reconstruction produced by ML-algorithm has been found noisy because it does not make use of available prior knowledge. Bayesian estimation provides such a platform for the inclusion of prior knowledge in the reconstruction procedure. But modeling a prior distribution is a cumbersome task and needs a lot of insight. In this work, we have proposed a conditional entropy maximization algorithm for PET modality, which is a generalization of maximum likelihood algorithm. We have imposed self-normalization condition for the determination of prior distribution. It is found that as prior distribution tends to uniform distribution, the conditional entropy maximization algorithm reduces to maximum likelihood algorithm. Simulated experimental results have shown that images reconstructed using maximum entropy algorithm are qualitatively better than those generated by ML-algorithm.