A retrospective image-based head MC method that corrected for transmission-emission as well as emission-emission misalignments was proposed and investigated in this study. The MC method had to be easy to use and not add tremendous burden to any human brain PET study, so as to make it practical for common use. The shape of a human head changed very little with movement, so rigid body transformations were used to model the relative positions during a dynamic PET scan (11
). Matching of two images was performed by finding the translations and rotations that optimize some matching function of the images (11
). This type of MC method is of value to those who want to extract the fine details of tracer behavior in a PET image, but are restricted in doing so due to the degradation present in the image caused by patient movement during the scan time. With the proposed MC method presented in this article, valuable data once hidden in a brain PET image can come to light, making otherwise questionable studies useful.
Various strategies have been employed to address the problems of patient head movement in PET (1
). An attractive method of late has been the acquisition of PET data in list-mode while simultaneously tracking the patient’s head movement with an optical motion-tracking system (1
). The optical motion-tracking system emits infrared light and detects the translational and rotational information of the head during image acquisition from the light that is reflected back from markers positioned on a patient’s head. Each detected event in the list-mode data is then corrected for by the movement information provided by the motion-tracking system and the image is reconstructed thereafter (1
). A big advantage of this method is its tracer independence (i.e., there is no reliance on the PET data for determining the head movement). However, there are technical issues with this method that still need to be addressed (1
). Also, optical-tracking systems are of no help in trying to retrospectively correct for head movement in PET images previously acquired. For these cases, image-based methods would offer a more practical solution.
One of the limitations of image-based methods (besides its susceptibility to the quality of the PET data) is that they do not account for motion within a frame. However, this problem can be minimized by shortening the frame duration in the PET protocol. It must be cautioned, though, that excessive shortening of the frame duration might also increase the noise due to lower counting statistics. Furthermore, in the MC method presented in this paper, there is no assumption or restriction set on having no movement between the transmission scan and the first emission scan, as is required by some other methods (18
). This restriction was removed by the alignment of the transmission image to the non-AC emission frames using the mutual information maximization criterion. Although mutual information is a powerful algorithm for the registration of multi-modality images (and which worked well in our study), the optimization of other cost functions may be more appropriate for other tracers where there is too little “mutual information” between the transmission image and the reference emission frame or between the emission frames themselves.
Recently, Costes et al. (20
) investigated a co-registration based frame realignment method for dynamic PET images using simulated [11
C]-raclopride PET data. Upon optimizing the choice of target volume and similarity criterion, a correction strategy was designed that took into account transmission-emission misalignment as well as realigning the individual time frames. Their optimal method consisted of using non-AC emission images and the cross-correlation criterion. Moreover, Mourik et al. (21
) evaluated four different “off-line” frame-by-frame motion correction methods. Their optimal method (based on simulated [11
C]-flumazenil and (R
C]-PK11195 studies) consisted of aligning non-AC emission frames to a summed image of the early non-AC frames (0–3 min). They then used a common attenuation map to reconstruct a series of aligned AC emission images. The study assumed there was no patient motion during the early emission frames, nor between the transmission scan and the start of the emission scan. Although both of these studies mirrored our study in certain respects, there were some important differences: (1) our method was derived from real patient data, not simulated data, (2) our method specifically accounts for the attenuation due to the head holder, (3) normalized mutual information was the matching criterion used in our study, (4) our method does not necessarily assume no mismatch between the transmission scan and the early frames of the emission scan, and (5) we worked with [18
F] instead of [11
With regard to the results reported in this paper, we saw that in Supplemental Figure 1
, the head movements of ControlSubject1
were relatively negligible while the head movements of ADSubject2
could not be ignored, especially along the longitudinal z-axis. The DVR images of ControlSubject1
before and after MC were very similar () and showed no apparent signs of image degradation that might have been introduced by the MC procedure. This should be the case if the MC procedure worked as it should since the subject had relatively little head movement to begin with. However, this does not mean that all control subjects will have negligible head movement. For example, there were control subjects in this study whose head movements were substantial and could not be ignored (e.g., control subject 3 in ). The MC procedure should thus be applied to all subjects in a brain PET study, regardless of whether head movements were apparent during the scan.
Previous PET studies have shown significantly higher [18
F]-FDDNP binding in the frontal, parietal, and temporal regions of the brain in patients with AD than in older control subjects without cognitive impairment (6
). After MC, the image quality of the [18
F]-FDDNP DVR images in subjects that moved was improved and [18
F]-FDDNP binding in the aforementioned regions was more clearly defined. Quantitatively, this can be seen in the increased separation of the mean DVR values between controls and AD in frontal, medial temporal, lateral temporal, and global ( and Supplemental Figure 3
). This increased separation explained in part why the discriminant analysis performed better after MC. Another reason for the improved performance was the considerable decrease (ranging from 42% to 91%) in P
-values after MC for frontal, parietal, posterior cingulate, medial temporal, lateral temporal, and global. The two reasons above allowed for the calculation of a more refined discriminant function that was used to correctly classify individuals from the sampled population. The resulting discriminant function can thus be of significant help to diagnose new AD cases based on [18
F]-FDDNP DVR images.
Additionally, before MC, we saw that medial temporal and subcortical white matter did not show significant differences between controls and AD. Since the medial temporal lobe is the brain region earliest affected in AD (23
), these results can be interpreted based on the possibility that elderly control subjects may already have significant pathology present even in the absence of neuropsychiatric symptoms. However, and even though it is possible that control subjects may have elevated medial temporal lobe signal with [18
F]-FDDNP PET, we saw that after MC, medial temporal did show a significant difference (P
= 0.0077) between controls and AD while subcortical white matter remained to be non-significant between the two groups. Medial temporal region is thus sensitive to head movement. The fact that DVR values in subcortical white matter did not differ significantly between the two groups also offers the possibility of using subcortical white matter as a reference region for [18
F]-FDDNP Logan analysis.
Moreover, the variability in regional DVR values decreased after MC. This decrease was most apparent in subcortical white matter in the AD group. MC should be considered if subcortical white matter is used as a reference region for Logan analysis, as its DVR is highly variable without MC in AD. The reduction in within-group CV would impact the design considerations of an experiment as well as the costs involved, since the sample size needed for detecting a given percentage change between means is a function of the CV (16
). In addition, regional TACs were distorted in subjects with large head movement when compared to the kinetic data after MC. This is a problem since the Logan plots of distorted TACs would yield less reliable DVR values. The proposed MC method would thus be a significant contributor to the precision of the data.