Correction of Head Movement on PET Studies: Comparison of Methods

MATERIALS AND METHODS

Movement Correction Methods

FBF Realignments Without MT.

The nonattenuation dynamic image was denoised using a level-2, order-64 Battle–Lemarie wavelet (22,23). Frames were realigned to a single 300-s emission frame acquired 25 min after injection using a mutual information algorithm (24), and the transformation parameters were then applied to the corresponding attenuation-corrected dynamic image.

MT System.

An optical tracking system (Polaris) used near-infrared sources to track the position of 4 reflective beads that were securely attached to a neoprene hat worn by the volunteer. Movement between the hat and the scalp was minimized by choosing the smallest hat size that the volunteer could tolerate. Head position was recorded 5 times per second. The Polaris device returns spatial information in the form of rotations in quaternion space, followed by translations in 3 dimensions, and a root-mean-square error (RMSE) on the measurement. The Polaris data were then converted to a scanner coordinate system as described by Hutton et al. (11), with the simplification that in our case the Polaris system was in a fixed position. To assess the effect of motion on ROI radioactivity, a composite index of motion was calculated for each Polaris time point. The RMSE was calculated by measuring the distance of 3 points in the positions of the sampled ROI (cerebellum, right and left striatum) from their position during the transmission scan. In addition, a mean RMSE for the whole scan was calculated as an index of total movement during emission scanning.

Methods Using MT.

Three methods of image realignment using Polaris data were compared (Fig. 1): The first, referred to as MTfbf, is an image-based realignment method. For every frame, it repositioned the transmission image to the average position during that frame, performed the normal precorrection and reconstruction (using 3DRP) method, and finally realigned the reconstructed image to the transmission scan. In the second method, hereafter referred to as MT correction, has been described in detail by Bloomfield et al. (13), with enhancements to improve quantification (14). Lines of response (LORs), individually recorded in list mode, were realigned to the head position during the transmission scan when the list-mode data were rebinned into sinograms taking normalization factors into account. These sinograms were then postprocessed by dividing each sinogram element by the percentage of the frame duration during which the corresponding LOR was measured (i.e., fully inside the detector ring). This percentage was approximated as follows: For every time frame, 10 evenly spaced time points were taken, for which a sinogram was computed by counting the number of repositioned LORs (one for each detector pair) that contribute to every sinogram element. The average of those 10 sinograms was then divided by the equivalent sinogram obtained without motion. Note that this computation is central processing unit–intensive (±30 min per time frame on an Opteron 280 [AMD]) compared with the rebinning (±60 min for the whole scan) due to the number of detector pairs, which is about 200 million for the EXACT 3D. Attenuation correction and scatter correction were then applied to the emission data. The corrected data were then made consistent by replacing all sinogram elements that were measured for <30% of the frame duration by the forward projection of an initial estimate (14) and finally reconstructed using the 3DRP algorithm. The third method, referred to as LMCfbf (15) realigned all detected LORs within a frame to an average position during the frame, again taking normalization into account, repositioning the attenuation image to this average position, performing the same type of corrections and reconstruction as in the MT case and followed by moving the reconstructed image for the frame to the transmission scan position. For the MTfbf and LMCfbf methods, attenuation correction factors were derived from forward projecting an attenuation image after repositioning. The motivation for these 3 different methods is given in the Discussion.

• Download figure
• Open in new tab
• Download powerpoint

FIGURE 1. 

Different stages involved in realigning emission data with Polaris-based methods. MTfbf is a 1-stage process, where reconstructed frames (F1 and F2) are moved to position of the transmission scan (TR), using the average frame position derived from Polaris data. MT is a 1-stage process in which lines of response (LORs) are moved to the transmission scan position based on Polaris data. LMCfbf is a 2-stage process: First, LORs within a frame are realigned to average position in the frame using Polaris data; then reconstructed frames are moved to transmission scan position using the average frame position derived from Polaris data. For all methods, only 1 Polaris coordinate is shown for simplicity. Full details of the different methods are given in the text.

RESULTS

An example of the output from Polaris and corresponding time–activity curves of the raw PET data and after different methods of head movement correction is shown in Figure 2. The pattern of movement is typical, with occasional large movements superimposed on a slow drift. The raw data time–activity curve shows greater evidence of head movement compared with the time–activity curve after movement correction. The RMSE for each individual scan is shown alongside the respective time–activity curve in Figure 3.

• Download figure
• Open in new tab
• Download powerpoint

FIGURE 2. 

(A) Plots of Polaris translation files (X, left–right; Y, anterior/posterior; Z, dorsal/ventral): x-axis time = min; y-axis from an arbitrary starting point = mm. a and b indicate large movements. (B) Time–activity curve and PET summated images of (from top to bottom): raw PET data, FBF realignment, MTfbf realignment, MT realignment, and LMCfbf realignment. CB = cerebellum; x-axis time = min; y-axis activity = kBq/mL.

• Download figure
• Open in new tab
• Download powerpoint

FIGURE 3. 

Time–activity curves (left) and RMSEs of motion (right) during scanning for each volunteer. (Left) ▵, DS; □, VS; ♦, cerebellum; y-axis = kBq/mL; x-axis = min. (Right) y-axis = mm; x-axis = min.

Noise Levels

Noise levels in the 2 sampling periods are shown in Figure 4. As expected, noise was greater in the second sampling period (F_1,7 = 43.05, P < 0.01), and noise varied across regions (F_2,14 = 9.54, P < 0.01). Importantly, realignment methods reduced noise overall (F_4,28 = 6.28, P < 0.01). Although there were no significant interactions between these variables, it was important to understand the differences between methods; therefore, post hoc t tests were performed and significant differences are shown in Figure 4.

• Download figure
• Open in new tab
• Download powerpoint

FIGURE 4. 

Noise levels (coefficient of variance, COV) in different regions after 5 different analyses: raw data, FBF realignment, and 3 Polaris methods (MTfbf, MT, LMCfbf). ANOVA showed significant effects of region, sample time, and method of analysis, with no interactions (error bars indicate SD; brackets indicate P < 0.05).

BPs

Measured BPs in the first and second sampling periods are shown for the 5 methods in Table 2 and Figure 5. ANOVA showed significant effects of region (F_1,7 = 60.28, P < 0.01) and method of analysis (F_1,7 = 7.19, P < 0.01) but not sample time. There were interactions between method of analysis and time (F_4,28 = 4.02, P = 0.01) and a trend interaction between method and region (F_4,28 = 2.25, P = 0.09). A post hoc t test showed no significant difference between the first and second sampling period in either ROI after any one method of analysis. Other comparisons are shown in (Fig. 5). There was no correlation between total RMSE motion and change in BP (data not shown).

• Download figure
• Open in new tab
• Download powerpoint

FIGURE 5. 

Mean BPs for VS and DS ROIs (brackets indicate paired t test, P < 0.05).

View this table:

• View inline
• View popup

TABLE 2

BPs in VS and DS After 5 Methods of Analysis

SPM Analyses

Glass brain depictions of results are shown in Figure 6. The raw data showed significant reductions in BP from period 1 to period 2 during the scan that were restricted to the anterior portions of the striatum (significant cluster peak at 26, 14, −2 (x, y, z), t = 5.97, p_corr = 0.04) and increases in anterior and posterior portions of the striatum (16, 6, −4, t = 6.58, p_corr < 0.01; −32, −16, 8, t = 5.74, p_corr = <0.01). No other contrasts had regions with significant differences between scanning periods when corrected for multiple comparisons.

• Download figure
• Open in new tab
• Download powerpoint

FIGURE 6. 

SPM comparisons of BP between first (A) and second (B) sampling periods. All voxels were significant at P < 0.01, uncorrected. Data are sagittal (top rows) and axial (bottom rows) glass brain views. Arrowheads indicate peak voxels in significant clusters.

DISCUSSION

This article describes, to our knowledge, the first quantitative comparison on patient data between raw, FBF, and Polaris-guided methods of motion-correcting PET data. Using several different criteria and analyses, all methods of realignment were superior to the raw data with Polaris methods performing better than FBF on some indices. Using such methods, the test–retest reliability and accuracy of PET data may be dramatically improved with consequent implications for interpretation of existing results and future study design.

In the VS, all methods of motion correction improved test–retest reliability with reduced variability. The ICC in the VS was increased by all methods compared with the raw data, indicating that between-subject differences were maintained or improved by the realignment methods. The implications of these findings on the design of future studies are exemplified by consideration of their impact on power calculations: To detect a 5% difference in the VS in the raw data would require 68 volunteers (power of 0.8, α = 0.05). The corresponding figures after the other methods are FBF, 28; MTfbf, 13; MT, 17; and LMCfbf, 14. Although studies using interventions may have different test–retest characteristics, such improvements have profound implications when considering the difficulty recruiting certain patient groups for clinical studies and the ethical issue of exposing volunteers to ionizing radiation.

All 4 methods of correction for head movement are likely to have improved the spatial accuracy of BP estimates. The degree of change in BP produced by head movements may be estimated by simulation studies in which the ROI is deliberately moved relative to the scan. For example, in the striatal subregions, small movements can cause changes in measured BP at least as large as those due to dopamine release (Fig. 7). Use of motion correction will improve the accuracy of BP measurements and enhance confidence in results, particularly in small ROIs such as the VS and DS. It is interesting to note that the changes of measured radioactivity in these 2 regions are quite different after head motion, either induced artificially, as in Figure 7, or as occurs within a clinical scan (Fig. 2B). This may be accounted for by the different local environments of the 2 regions. For example, as the VS and DS ROIs move rostrally in the z-plane, either due to movement of the ROI rostrally or to the volunteer moving caudally, the activity in the DS is more profoundly reduced than the activity in the VS. This is because the ROIs are no longer sampling the VS and DS and are, instead, sampling other local brain regions. In this case, the VS ROI samples what is actually DS (a high-activity region). The DS ROI, in contrast, samples brain regions with low radioactivity and, hence, shows a sharp drop in activity. In practice, movements made by volunteers during scanning do not occur in only 1 plane at a time, making it very difficult to account for the different patterns of uptake observed in Figures 2B and 3, for example. It is equally difficult to account for the different effects the various methods of motion correction have on the time–activity curve in Figure 2B because subtle variations in the degree of realignment between methods have correspondingly subtle and unpredictable effects on the time–activity curve.

• Download figure
• Open in new tab
• Download powerpoint

FIGURE 7. 

Measured BP in VS and DS ROI as ROI is moved a voxel at a time in the z-axis. Large changes in BP result from small changes in ROI position, demonstrating the susceptibility of measurements to movement.

The use of head movement correction is likely to be beneficial for other radioligands with different patterns of uptake than ¹¹C-raclopride. ¹¹C-Raclopride provides a good exemplar of the method because its highly localized pattern of uptake in the striatum with very little uptake in neighboring regions makes BP measurements very vulnerable to effects of head movement. However, radioligands with cortical uptake may be similarly vulnerable because of the close proximity of scalp. Because the Polaris-based methods do not rely on the PET data to compensate for movements, they should be applicable to other radioligands, where they are likely to have similar benefits of improved accuracy and test–retest reliability.

The SPM analyses showed that in this set of raw data, BP tended to reduce in the anteroventral portion of the striatum and increase in the dorsal as well as ventral regions between scanning periods. After different methods of movement correction, there were no voxels that significantly reduced in value between sample periods, although all methods continued to show voxels, which increased between the 2 periods (not corrected for multiple comparisons). It is likely that the different methods of head restraint used in different PET centers generate different patterns of apparent change in BP within the striatum (or, indeed, in extrastriatal regions), depending on the type of head movement they allow (1,4) and the regional distribution of the radioligand. A further factor that is likely to influence the type and degree of head movement volunteers make during a scan is the presence or absence of pharmacologic or behavioral stimulation during the scan. For example, agents such as amphetamines promote motor activity and may be anticipated to increase head movement. To date, the influence of such factors on PET and SPECT outcome measures has not been systematically evaluated. Few studies in the literature report whether scans have been rejected because of excessive head movement—perhaps because of the difficulty defining a cutoff point beyond which data are no longer reliable.

Changes in VS regional activity (percentage change in radioactivity per hour) showed a trend reduction by LMCfbf realignment compared with raw data and were significantly reduced compared with the FBF data by LMCfbf. The values in the DS are smaller than those previously reported (5). Mawlawi et al. (5) found that DS regions reduced in activity by −12.9% ± 15.1% and −8.5% ± 13.5% per hour in dorsal caudate and dorsal putamen, respectively—changes that differed significantly from zero (it is not clear if raw or motion-corrected data are reported). In the VS, changes did not differ from zero and were of a similar order (−3.2% ± 13.9%) as those reported here. The only other equilibrium data reported in the literature are more difficult to compare with our data because the striatum is not subdivided (25). However, the values of −1.8% ± 9% for the striatum and 0.6% ± 12.6% in the cerebellum are very similar to our data.

Noise was significantly higher in the second sampling period and in the smaller VS ROI. This was presumably due to radioactive decay causing poorer counting statistics in later frames and smaller ROI. In the second period noise was reduced significantly in all ROIs by LMCfbf, by MT in the cerebellum, and by MTfbf in the VS and DS. Polaris-based methods performed better than FBF in the VS (Fig. 4). To our knowledge, only one other group has published a similar analysis of ¹¹C-raclopride BI data (5). Their study used somewhat higher injected doses of radioactivity (470 ± 166 MBq); however, noise levels in the data presented here are comparable with the raw dataset and are better in the realigned data. Mawlawi et al. (5) also found that FBF realignment reduced noise levels.

Three different Polaris-guided methods are reported here. Two methods used list-mode repositioning (MT and LMCfbf), whereas the other method (MTfbf) used image repositioning. Neither the FBF (which does not use Polaris data) nor the MTfbf method can correct for movement within the frame. If no within-frame movement occurs, the MTfbf and LMCfbf images are identical. It is possible that both frame-based methods would have performed better if the frame duration had been shorter, but this has not been examined to date. If large movements occur with respect to the position during transmission scanning, the MT method suffers from 2 drawbacks compared with LMCfbf: First, the standard scatter estimation is no longer accurate, which will introduce a bias increasing with the amount of the movement. The reason for this, as explained by Thielemans (15), is that the scatter distribution after realignment of the LORs is not the same as the scatter distribution that would have been measured in the target position (target position is the motion-corrected position). Second, because of the requirement of 3DRP for consistent data, some of the measured data cannot be used (14), which will lead to an increase in noise for larger movements. Of these 2 effects, the first was illustrated by Thielemans (15), showing quantitative difficulties of MT on phantom data with fairly large movements. The effect may be responsible for the slightly reduced BP measured in the DS when MT is compared with other methods. The expected reduction in noise of LMCfbf over MT is reflected in the data of Figure 4.

Fulton et al. (26) have published a quantitative comparison between MT and MTfbf on phantom data. They found that the MT method provided somewhat better quantification than MTfbf, which was considerably better than using no motion correction. However, this article has a different implementation of MT compared with ours. In our method, each LOR is weighted after repositioning with a factor that takes its measurement time into account (and some LORs are filled in by forward projection), whereas Fulton et al. use a global factor (i.e., identical for every LOR). In our experience, such a global factor is not sufficient for accurate quantification for arbitrary movements. This might be partially explained by the fact that we have used the 3DRP reconstruction algorithm, whereas Fulton et al. used Fourier rebinning plus filtered backprojection.

CONCLUSION

The use of a MT realignment system such as Polaris confers significant advantages to ¹¹C-raclopride PET studies, potentially allowing single-scan, dual-condition studies to detect small changes in radioligand binding in the VS with considerably increased statistical power compared with existing methods. Of the 3 Polaris-guided methods, LMCfbf is theoretically the most reliable.