Introduction
Positron emission tomography (PET) is a medical imaging technique that allows for measurements of tissue function by following the time course of a tracer labelled with a positron emitter. Most dynamic brain scans require an acquisition time of 60–90 min and, for accurate results, the subject should remain in exactly the same position. In practice, however, subject motion is not uncommon, especially not for specific patient groups such as, for example, patients suffering from Alzheimer’s or Parkinson’s disease.
Full utilisation of improvements in intrinsic spatial resolution of new PET scanners is increasingly hampered by patient motion [
1]. Patient motion during a PET scan may reduce effective spatial resolution [
2]. More importantly, patient motion may alter measured time-activity curves (TAC), especially for small regions of interests (ROI), thereby directly affecting the outcome of tracer kinetic analysis.
The simplest method to reduce patient motion during scanning is the use of head restraints. To date, a number of head restraints are available for reducing motion (e.g. [
2,
3]). As these head restraints do not eliminate all movements, even more restrictive head restraints exist that fix the skull completely [
3]. These restrictive head restraints, however, are very uncomfortable and therefore they are not used frequently. In addition, many patients (e.g. traumatic brain injury, obsessive-compulsive disorder) do not tolerate rigid head fixation.
An alternative is to register motion during scanning using an optical online motion tracking system. Most recent optical motion tracking systems [
4‐
6] enable online correction for motion that occurs within frames (
in-frame patient motion). Online motion tracking systems have two main advantages. Firstly, when using a motion tracking system, there is no mismatch between emission and transmission scans, as emission data are realigned to the position of the head during the transmission scan. This is very important, because a mismatch between emission and transmission scans leads to erroneous attenuation correction. Secondly, it is possible to correct for in-frame motion, as realignment may take place several times per second. However, motion tracking systems also have some disadvantages. Firstly, older data sets, acquired prior to installation of a motion tracking system, cannot be corrected for patient motion. Secondly, most optical online motion tracking systems require PET data to be acquired in list mode, which is not possible on older PET scanners. Thirdly, online (continuous) motion correction during reconstruction is not trivial and some difficulties with normalisation and attenuation correction still need to be investigated further [
7]. Finally, the use of optical (online) tracking systems is not always possible when the view of the patient in the gantry is limited. This is, for example, the case when scanning patients with traumatic brain injury, where the view within the gantry is partly blocked by auxiliary equipment, such as that needed for administering anaesthetics. In those patients, however, motion is observed frequently.
Frame-by-frame motion correction methods correct image data post hoc. Existing frame-by-frame methods use correlation coefficient [
8], cross-correlation [
9,
10], mutual information [
9,
11], standard deviation of the ratio of two images [
10,
12], sum of absolute differences [
10,
12], mean square difference [
10], stochastic sign change [
10] or (scaled) least-square difference images [
13,
14]. Although, frame-by-frame motion correction methods do not have the same advantages and performance characteristics as online (optical) motion tracking systems [
11], they are very useful when no list-mode data are available, when older data have to be reanalysed or when optical tracking is not possible because of a limited view into the PET gantry.
The purpose of this study was to evaluate four different off-line frame-by-frame motion correction methods, previously introduced by Perruchot et al. [
9]. Two of these motion correction methods in theory also correct for mismatches between transmission and emission scans. Methods were evaluated extensively using both simulation studies and several clinical data sets, covering both tracers with low and high cerebral uptake.
Discussion
Simulations
The simulation studies showed that motion had a large impact on both regional TACs (max. 98% for the motion parameters selected) and parametric tracer kinetic analysis (max. 45%). Furthermore, these studies showed that the best motion correction method from a theoretical point of view, where emission and transmission scans were aligned (method C or D), did not provide the best results. In fact, for SIM
FMZ these methods gave the poorest results (see Fig.
7a, green and light blue symbols) and frequently motion correction failed completely. One reason for these poor results in the case of SIM
FMZ may be that the (cupped) μ and PET images (either AC or NAC) only have limited corresponding contours or information (see Fig.
2). Although for SIM
PK more corresponding contours or information were seen, motion correction still failed frequently, especially for the last few frames (data not shown).
The best results were obtained for method B, NAC-on-NAC, which is consistent with a previous study performed by Perruchot et al. [
9]. This method assumes that there is no motion during the first
x minutes. If this assumption holds, and the patient has not moved between transmission and emission scan, emission and transmission scans are aligned. Using the optimal motion correction strategy B (NAC-on-NAC), corrected V
T values maximally differed by 3.5% from the original (true) V
T values.
Clinical data
Large movements
Both [11C]flumazenil and (R)-[11C]PK11195 data sets were corrected very well using the settings found in the simulation study. Using frame-by-frame movies and superimposing the contour of the first x frames is a useful tool to determine whether motion is present.
Differences between original V
T images and V
T images after motion correction were large (up to 49%). In addition, large artefacts, present in the original (
R)-[
11C]PK11195 V
T images, disappeared when motion correction was applied. Artefacts were less visible in the original [
11C]flumazenil V
T images, which is caused by the particular distribution of activity together with the smaller amount of motion in the [
11C]flumazenil data set (Table
1). Although visually there was close agreement between V
T images before and after motion correction, both ratio image and ROI analysis showed that the effects of motion were considerable with differences at a regional level being as high as 26.9%. Therefore, it is recommended that all data be inspected for the presence of motion, e.g. by visual inspection of frame-by-frame movies. In addition, results of the present study indicate that motion correction could always be applied, as effects in case of minor movements are small.
Minor movements
Although the simulation studies already showed that, in the absence of movement, there were no significant changes in VT values, minor movements were further evaluated for three clinical data sets. Only for the [11C]flumazenil study were small changes in VT after motion correction observed. This suggests that there was some small movement, although it cannot be excluded that the change was due to the correction algorithm itself. The latter is, however, unlikely as changes for both (R)-[11C]PK11195 and [11C]PIB were very small (max. 1.5 ± 1.3%). These results confirmed those of the simulation studies in that corrections were negligible for minor movements.
Limitations
One limitation of the proposed frame-by-frame motion correction method is its vulnerability to the quality of the scan data and noise characteristics [
4]. Montgomery et al. [
11] claimed that problems may occur within the last frame of the PET scan because of poor statistics. This may especially be the case for older data sets acquired on lower sensitivity scanners. In the present study, effects of noise were reduced as much as possible by reconstructing data sets with OSEM, which provides images with less noise then filtered back projection. In addition, images were smoothed (only) during the motion correction process, which was not only helpful for the motion correction optimisation process, but also for reducing the level of noise. Problems with the last frame, as reported by Montgomery et al. [
11], were not observed in the present study. Although it is possible that the motion correction algorithm might fail due to low image quality, this does not seem to be likely given the positive results with the low uptake ligand (
R)-[
11C]PK11195 in the present study.
Another limitation of the frame-by-frame motion correction method is the underlying assumption that there is no significant change in activity distribution within and between frames [
4]. This assumption can only be true for the last frames. In fact, the activity distribution varies rapidly in the first frames. For the optimal motion correction method B, the summed reference image has therefore another activity distribution compared to the later frames. However, for the present study, this did not cause any problems during the motion correction process.
The optimal motion correction method derived from the present study assumes that there is no mismatch between the transmission scan and the first x frames of the emission scan. This is a reasonable assumption, because most patient motion appears at later time points of the scan (>10 min) (data not shown). It is, however, recommended that correct alignment of transmission and early (in this case 0–3 min post-injection) emission scans be verified. Nevertheless, if motion appears between transmission and emission scan or within the first 3 min of the emission scan, the mismatch between transmission and emission scans remains. Although motion correction methods C and D correct for emission and transmission mismatch, no satisfactory results were obtained in this study. The use of different cost functions might improve the performance of these methods and therefore have to be investigated. However, for some tracers, like [11C]flumazenil with high cortical uptake, methods C and D will probably always fail because there is too little commonality between the emission (NAC) and µ-images. Even if a mismatch between emission and transmission scans would exist, results of the present study show that a frame-by-frame correction method provides a major improvement in accuracy of pharmacokinetic analyses over non-motion-corrected data.
As mentioned before, the final method derived from the present study does not correct for in-frame motion and therefore motion could still be present within a frame. However, the present method may be suited to identify suspicious frames and exclude those frames in the following analysis. An easy way to do this is to make a frame-by-frame movie of a data set and identify in which frame the motion starts. Subsequently, the frames before and after the initial motion-affected frame should be visually inspected for any unexpected blurring. A highly blurred frame indicates that there is considerable patient motion during the frame and therefore that frame should be excluded in further analysis.
The AIR package was chosen because it is freely available, easily adjustable and fast. However, it has the disadvantage that only three different cost functions are available. In some applications other cost functions, such as those based on mutual information, may be more appropriate. In those cases, however, it will be necessary to again determine optimal settings for that specific registration algorithm.
Clinical applicability
The present study shows that it is possible to perform an accurate off-line motion correction for dynamic brain studies. In theory, this method should also be applicable to other organs, provided observed motions are rigid. The latter requirement, however, will limit use of the proposed method for non-brain studies. Even when the method would be applicable, optimal settings for the motion correction algorithm need to be re-evaluated for such an application.
For older data sets, no raw data may be available. The optimal method presented here does not require sinogram or list-mode data, but only reconstructed PET and µ-images. Therefore, this method can also be used for accurate motion correction of older data sets.
In clinical practice, subjects are fixed using a head holder. The movements that are most frequently observed are rotations around the x-axis and axial translations. Therefore, only these kinds of movements were included in the simulation study. Clearly, other types of motion (e.g. rotations around the z-axis) may also occur. It should be noted, however, that these types of motion will also be corrected for using the method presented, as all rotational and translational movements are obtained during the registration process.
Conclusion
If no optical tracking system is available or when older data sets need to be reanalysed, a frame-by-frame motion correction method, based on non-attenuation-corrected images, provides major improvements in accuracy of pharmacokinetic analyses over non-motion-corrected data.
Acknowledgements
This work was financially supported by the Netherlands Organisation for Scientific Research (NWO, VIDI Grant 016.066.309). The authors would like to thank Anthonin Reilhac for his useful comments, Bart N.M. van Berckel, Hedy Folkersma, Reina W. Kloet, Ursula M. Klumpers, Nelleke Tolboom and Alie Schuitemaker for providing the clinical data, and the radiochemistry and technologists staff of the Department of Nuclear Medicine & PET Research for production of isotopes and acquisition of data.