Parametric imaging with Bayesian priors: a validation study with (11)C-Altropane PET.

It has been suggested that Bayesian estimation methods may be used to improve the signal-to-noise ratio of parametric images. However, there is little experience with the method and some of the underlying assumptions and performance properties of Bayesian estimation remain to be investigated. We used a sample population of 54 subjects, studied previously with (11)C-Altropane, to empirically evaluate the assumptions, performance and some practical issues in forming parametric images. By using normality tests, we showed that the underpinning normality assumptions of data and parametric distribution apply to more than 80% of voxels. The standard deviation of the binding potential can be reduced 30-50% using Bayesian estimation, without introducing substantial bias. The sample size required to form the a priori information was found to be modest; as little as ten subjects may be sufficient and the choice of specific subjects has little effect on Bayesian estimation. A realistic simulation study showed that detection of localized differences in parametric images, e.g. by statistical parametric mapping (SPM), could be made more reliable and/or conducted with smaller sample size using Bayesian estimation. In conclusion, Bayesian estimation can improve the SNR of parametric images and better detect localized changes in cohorts of subjects.