An improved statistical approach to the estimation of spatial bias and variability in reconstructed PET data

The spatial bias and covariance characteristics of reconstructed PET data are important as routine assessments of image quality in PET scanners. Previous empirical analysis of uniform phantom data across the 43 sites in ACRIN shows that both spatial bias and covariance can be decomposed as a product of a trans-axial radial function and axial function. On the basis of this decomposition, a penalised maximum likelihood type approach for analysis of bias and variance patterns is developed. This can be viewed as a variation on the Π-method of Brieman. The proposed approach is evaluated using a simulation model and demonstrated on PET data collected on an operationally used clinical PET scanner with the dynamic brain ACRIN imaging protocol.