A Maximum-likelihood Approach To PET Emission/attenuation Image Registration

Abstract

An algorithm for retrospective registration of PET emission and attenuation images is introduced based upon the assumption that no activity should be present outside the body. Since the attenuation image defines anatomy, it is segmented to identify a foreground template of the subject. This template is mapped to the emission image by a geometric transformation, and the emission image pixels that do not correspond to the foreground template are considered the background pixels for that transformation. An optimal geometric transformation mapping the foreground template to the emission image is sought by maximizing the likelihood of the background emission image pixels given a background intensity distribution; this background intensity distribution is estimated from the background pixels for the initial position of the template. The current implementation of the algorithm is confined to 2-D rigid motion, although these constraints are not fundamental to the method. In a preliminary phantom study where the phantom was moved between the acquisition of emission and attenuation data, the registration algorithm predicted the rotation to within five degrees (worst case) and usually to within three degrees, and it predicted the translation to within one and a half pixels; it predicted the rotation to within one degree and the translation within one pixel when performed on images reconstructed with a matching attenuation image. Possible improvements to the algorithm are suggested, based on the observation that the predicted motion is biased by the attenuation correction artifact in the images.