Registration of segmented attenuation and emission data in PET

Estimates of patient motion between acquisition of attenuation and emission data in PET can be used to align the emission and attenuation data and sets of time-binned emission data. A two-region maximum likelihood (ML) algorithm (Costa et al., 1993) was developed to estimate rigid motion parameters between segmented attenuation and emission images. The authors previously reported that this algorithm produces registration parameter estimates which are biased toward the position of the object in the attenuation data used to weight the emission data before reconstruction. The performance of the two-region likelihood algorithm is evaluated using images of a homogeneous elliptical phantom reconstructed without attenuation correction. This results in unbiased estimates of the registration parameters. However, the two-region ML registration algorithm returns biased estimates of the registration parameters for a more clinically relevant torso phantom. For comparison, an alternative algorithm similar to that described by Watanabe et al. (1992) is also applied to the torso phantom. This algorithm estimates the registration parameters more accurately, but its precision is hindered by the existence of local maxima in the merit function maximized to obtain the registration parameters