Impact of motion correction on [¹⁸F]-MK6240 tau PET imaging

Abstract

Objective:

PET imaging of tau deposition using [¹⁸F]-MK6240 often involves long acquisitions in older subjects, many of whom exhibit dementia symptoms. The resulting unavoidable head motion can greatly degrade image quality. Motion increases the variability of PET quantitation for longitudinal studies across subjects, resulting in larger sample sizes in clinical trials of Alzheimer’s disease (AD) treatment.

Approach:

After using an ultra-short frame-by-frame motion detection method based on the list-mode data, we applied an event-by-event list-mode reconstruction to generate the motion-corrected images from 139 scans acquired in 65 subjects. This approach was initially validated in two phantoms experiments against optical tracking data. We developed a motion metric based on the average voxel displacement in the brain to quantify the level of motion in each scan and consequently evaluate the effect of motion correction on images from studies with substantial motion. We estimated the rate of tau accumulation in longitudinal studies (51 subjects) by calculating the difference in the ratio of standard uptake values in key brain regions for AD. We compared the regions’ standard deviations across subjects from motion and non-motion corrected images.

Main Results:

Individually, 14% of the scans exhibited notable motion quantified by the proposed motion metric, affecting 48% of the longitudinal datasets with three time points and 25% of all subjects. Motion correction decreased the blurring in images from scans with notable motion and improved the accuracy in quantitative measures. Motion correction reduced the standard deviation of the rate of tau accumulation by −49%, −24%, −18%, and −16% in the entorhinal, inferior temporal, precuneus, and amygdala regions, respectively.

Significance:

The list-mode-based motion correction method is capable of correcting both fast and slow motion during brain PET scans. It leads to improved brain PET quantitation, which is crucial for imaging AD.

Introduction:

Alzheimer’s disease (AD) is a neurodegenerative disease characterized by the presence of two histopathological lesions: plaques formed by a deposit of amyloid protein and tangles composed of tau protein (Hyman et al., 2012). The recent introduction of Positron Emission Tomography (PET) radiotracers has revolutionized the field of AD research (Villemagne et al. 2015) by enabling the observation of tau propagation in vivo with longitudinal PET scans. [¹⁸F]MK-6240 is a novel tracer developed to achieve high binding affinity and selectivity to tau protein (Hostetler et al., 2016). The rate of tau accumulation as time advances is the measurement of interest in AD research as it correlates with the decline in cognitive function (Jack et al., 2010). Longitudinal studies are, therefore, an integral part of clinical trials in the drug development process. A successful drug candidate would significantly reduce the rate of tau accumulation (Takeda, 2019) in the treated group compared to the placebo group. To evaluate [¹⁸F]MK-6240’s properties and compare them in subjects with and without treatment as required by clinical trials, full dynamic imaging is performed since it contains the maximum amount of quantitative physiological information (Guehl et al., 2019). The long acquisition duration, the age of subjects imaged, and their dementia diagnosis are all factors contributing to the unavoidable head motion that occurs during the acquisition leading to higher variability in the measured rate of tau accumulation across subjects and, therefore, a larger required sample size and more expensive clinical trials. Head motion can be a major source of such variability, although other sources include the uncertainties of the PET measurement itself (Salinas et al. 2020), as well as physiology effects such as the heterogeneity of diagnoses in subjects with high tau uptake (Nelson et al. 2012), the complex spatiotemporal pattern of tau propagation (Theofilas et al., 2018), and off-target binding (Betthauser et al., 2019) coupled with the low abundance of tau in the brain. The blurring of the image due to motion is an important confounding factor in PET quantification (Carles et al., 2018) especially in tau PET imaging: the high focal nature of tau uptake (Braak & Braak, 1991) in the early stages of AD (Sanchez, 2021) and the presence of off-target binding make it harder to pinpoint the original signal especially when motion is present. As a result, head motion correction is crucial to reduce the variability of PET-derived quantitative measures.

Head motion correction in PET imaging has been well-researched in the past. Image registration for motion detection and correction is often used in dynamic studies where multiple frames are acquired (Costes et al., 2009, Woo et al. 2011). In [¹⁸F]-MK6240 studies, the frames are usually reconstructed into 5-min bins and then registered to each other to correct for motion (Guehl et al., 2019). Additional processing steps needed to correct the motion-induced mismatch between the PET data and the attenuation map are often ignored. In addition, this approach cannot detect or correct intra-frame motion. Real-time, or near real-time, motion tracking is required to detect all movements. This can be achieved using an optical device tracking external markers attached to the head to follow motion during the acquisition (Bloomfield et al., 2003). This method requires additional devices, which can pose multiple challenges and not be well tolerated by subjects, especially by those with cognitive impairment. Data-driven motion correction techniques were developed to detect motion from the list-mode data by tracking the center of mass from time of flight (ToF) information. Initially, the axis where motion was most important was selected as a reference for partitioning the list-mode data into moving and static frames (Markiewicz et al., 2016; Lu et al., 2020). The motion index was also proposed to combine all 3D motion information (Sun et al., 2019). The limitation of these techniques is the need for a compromise between the number of frames to reconstruct and the computational time required. In addition, the coordinates of the center of mass cannot be accurately determined with dynamic activity distribution. Recently, a new ultra-fast list-mode reconstruction framework (Spangler-Bickell et al., 2021, Spangler-Bickell et al., 2022) was proposed that combines a data-driven image-based motion estimation technique and a list-mode event-by-event motion-corrected reconstruction. The computation speed achieved in the motion detection step allows for reconstructing very short frames with enough signal to perform image registration. Once the rigid transformation parameters are determined, the motion matrices are added to the reconstruction framework for a second list-mode reconstruction of all events detected in all the short frames yielding the motion-corrected image. There are many advantages to using this technique for motion correction: the speed of the short frames reconstruction makes it useful in clinical settings as it allows for near real-time data-driven motion tracking; there is no need for an external tracking system so it can be applied to data retrospectively (for example, after discovering that the images are blurry due to motion) as long as the list-mode data is available; it is highly flexible since partitioning of short frames is based on the desired duration or the number of events detected; it can handle a changing activity distribution which is helpful in dynamic studies.

In this work, we take advantage of the large number of list-mode datasets available to us to apply the ultra-fast list-mode reconstruction method for motion detection and correction in [¹⁸F]-MK6240 studies. We also evaluate its performance in two phantom experiments. Visual inspection, as well as quantitative results, are reported. We investigate the effects of motion correction on the measurements of the rate of tau accumulation inferred from longitudinal studies and its potential benefits in the design of clinical trials.

Methods:

Data acquisition

Two phantom experiments were conducted to evaluate the motion estimation and correction technique. First, a Hoffman phantom (Data Spectrum, USA), filled with 111 MBq ¹⁸F, was scanned on a Discovery MI (GE Healthcare, USA) PET/CT scanner for 15 min in list-mode while undergoing continuous rotation and translation movements introduced by a QUASAR system (Modus QA, Canada). A similar acquisition was repeated on a Mini Hot Spot phantom (Data Spectrum, USA) consisting of hollow channels of diameter 1.2, 1.6, 2.4, 3.2, 4.0, and 4.8 mm filled with 74 MBq ¹⁸F. For both phantoms, static acquisitions without introducing any motion were performed to serve as the ground truth for the activity distribution. For moving acquisitions, each phantom was placed on a 20° wooden slope and pushed by the Quasar’s piston in a back-and-forth fashion with a stroke of 40 mm for the Hoffman phantom and 20 mm for the Mini Hot Spots. The induced movement consisted of periodic and continuous rotations and translations thanks to the motorized system. We believe that this motion pattern, while not representing clinical reality, is complex enough to encompass severe scenarios of subjects’ motion. Motion was tracked continuously using a Polaris Vega Camera (NDI, Canada). It is a device approved for medical environments that tracks a target equipped with spherical reflective markers, with a 0.12 mm volumetric accuracy. The motion capture frame rate was selected to be 60 Hz. Temporal synchronization and spatial calibration were performed. The Polaris’ tracking data were used to generate rigid transformation parameters (rotations and translations) in the image space to relate every tracked pose of the phantoms back to the initial pose before motion was introduced. The conversion between the Polaris coordinates and the PET image coordinates was modeled by a 3D rigid transformation matrix. Its parameters were optimized from paired measurements of the positions of markers (manually spotted in the reconstructed CT volume from the PET/CT scanner and directly measured by the Polaris camera). To account for differences introduced by displacing the camera in subsequent scans, another transformation matrix is derived solely in the camera space to link the target markers placed on the phantoms to reference markers fixed to the gantry. As such, the composition of the two matrices transforms the target markers coordinates measured by the Polaris into the PET image space. The camera-tracked poses served as ground truth for the transformation parameters obtained by the list-mode based motion estimation technique.

A total of 65 subjects (55 Cognitively Normal – CN, 7 with Mild Cognitive Impairment – MCI, and 3 with Alzheimer’s Disease – AD) aged 68 ± 11 (mean ± standard deviation) were included in this study. All were scanned on the GE Discovery MI PET/CT scanner for 20 min, 90 min after administration of ~185 MBq [¹⁸F]-MK6240. Fifty-one subjects (45 CN, 4MCI, 2 AD) underwent a second scan 7±4 months later, and 23 subjects (19 CN, 3 MCI, 1 AD) were scanned a third time 8±4 months later. Additionally, all subjects underwent an MRI scan using a standard Magnetization-Prepared RApid Gradient-Echo (MPRAGE) sequence (Mugler & Brookeman, 1990) on a Siemens 3T system, optimized for FreeSurfer (Fischl, 2012) to generate high-resolution anatomical regions of interest for volumetric analyses. All subjects provided informed consent, and the study was approved by the institutional review board at Massachusetts General Hospital.

Motion correction

The list-mode based motion correction framework is composed of two steps for each PET study. First, the list-mode data were divided into multiple short frames which were separately reconstructed into low-resolution (4 mm × 4 mm × 2.8 mm) images using ultra-fast list-mode reconstruction toolbox provided by GE Healthcare (Spangler-Bickell et al., 2021). Four iterations of a maximum-likelihood-expectation-maximization (MELM) algorithm with a fast time-of-flight (TOF) based ray-tracing projector were used in this reconstruction step. No point spread function modeling was performed. Attenuation and scatter corrections were also ignored in this first step. The rigid registration of the short frames achieved motion estimation. Second, the obtained transformation parameters were applied to the endpoints of the measured events to correct for motion. A list-mode reconstruction of the motion-corrected events was performed with all relevant PET corrections. This reconstruction was achieved with an ordered-subsets-expectation-maximization (OSEM) algorithm with 80 total updates where 10M events are processed in each update. The number of iterations and subsets were automatically determined and ranged between 5–7 iterations and 12–16 subsets. In this step, attenuation, random, and scatter corrections were incorporated.

In each of the two phantom studies, the list mode data were reconstructed into frames of ~0.3 seconds with 300k non-random events each. The continuous movements started 30 seconds after the beginning of the scan, so we used the mean of the first 40 short frames to form a motion-free reference image (spanning ~10 seconds) needed in rigid registration. Since no movement was introduced between the CT scan and the obtained reference frame, there is no need to correct for motion relative to the attenuation map for both phantom experiments. A second reconstruction step incorporating the transformation parameters was applied to the entire acquisition of 15 min to produce the motion-corrected (MC) image. The last step was repeated without incorporating the registration parameters to generate a non-motion-corrected (NMC) image used for comparison. The static acquisition was also processed for each phantom to generate a static image.

For each human study, the list mode data were reconstructed into frames of ~20 seconds with 1.5M non-random events each. The reference image was formed from the first 10 frames. All frames were registered to this reference through rigid transformation and then averaged to form a mean image, which was used as a fixed image for CT-to-PET registration to correct for motion relative to the attenuation map. A final list-mode reconstruction of all detected events was applied using the saved transformation parameters to generate the motion-corrected image. A non-motion-corrected image was also produced for comparison.

Frame-to-frame rigid registrations were performed with Matlab’s default image registration (MathWorks Website, 2022) using the least-square cost function with a gradient descent optimization algorithm. CT-to-PET registrations used a mutual-information-based cost function to handle the differences in intensity across modalities.

The processing was performed on 8 threads on one CPU. The motion estimation step lasted ~90 seconds for processing 20 minutes of PET data (~150M events including randoms). The rigid registration of the resulting ~60 short frames was done in ~30 seconds. The non-motion-corrected list-mode reconstruction was performed in ~45 minutes. When motion correction was added, the computation time varied based on the level of motion detected because the sensitivity image is produced for different positions grouped together according to a maximum allowed displacement. This step took between 54 min and 2.7 hours.

Data analysis and performance evaluation

Profile lines were drawn on the phantom images to compare the static, MC, and NMC images. For the Mini Hot Spot phantom, the profile line crossed four 3.2 mm channels. For the Hoffman phantom, two rectangular regions of interest (ROIs) were drawn to calculate the ratio of the signal in a cold region to that in a hot region. The moving acquisition was divided into 15 noise realizations of 1 min each to calculate the bias and the standard deviation of the ratio, compared to the static image, in MC and NMC images.

All subjects’ MR images were processed through FreeSurfer (Becker et al., 2011) to delineate key brain regions. In this study, we focused on the cerebellum gray matter, commonly used as a reference region in tau PET imaging; the hot regions where tau uptake is expected to be high in prodromal AD, such as the entorhinal cortex, the inferior temporal gyrus, the precuneus, and the amygdala; and the cold regions where tau uptake is supposed to be low such as the lateral, the third, and the fourth ventricles. Activity concentration values were calculated for each MC and NMC image in each brain region after MR-to-PET registration. Standard Uptake Value Ratios (SUVR) were derived afterward using the cerebellum gray matter as a reference region.

We defined a metric used to quantify the level of motion in each scan in the following way. We first calculated the average voxel displacement for each short frame from the transformation parameters obtained in the motion estimation step. Only the voxels included in the subject’s head were considered. We then computed the displacement index, expressed in millimeters, by summing the average voxel displacement over all short frames in each scan. We carefully examined the images from cases where subjects exhibited notable motion for whom at least one scan had a displacement index higher than 4 mm.

The processed data included longitudinal studies with at least two time points. We estimated the rate of tau accumulation from the change in SUVR between consecutive scans, denoted by δ, and expressed it as a yearly relative change from baseline:

$$\delta =\frac{1}{\mathrm{\Delta }t}\times \frac{{\mathrm{S}\mathrm{U}\mathrm{V}\mathrm{R}}_2-{\mathrm{S}\mathrm{U}\mathrm{V}\mathrm{R}}_1}{{\mathrm{S}\mathrm{U}\mathrm{V}\mathrm{R}}_1}$$

where $\mathrm{\Delta }\mathrm{t}$ refers to the time elapsed between the two scans (in years), ${\mathrm{S}\mathrm{U}\mathrm{V}\mathrm{R}}_1$ is the tau measurement at baseline in one ROI, and ${\mathrm{S}\mathrm{U}\mathrm{V}\mathrm{R}}_2$ is the value obtained from the follow-up imaging.

For all subjects with two time points, we calculated the standard deviation of δ for MC and NMC images in the four hot regions: entorhinal, inferior temporal, precuneus, and amygdala.

Results:

The motion estimation method succeeded in detecting the introduced motion in both the Hoffman and Mini Hot Spot phantoms. The periodic movements introduced were correctly detected by both the optical and the PET-list-mode-based tracking. Figure 1 shows the obtained transformation parameters from both methods for the first 90 seconds of each moving scan. In general, we found a good correspondence in parameters between the two methods. We noted that there is a slight temporal shift (less than 0.5 sec) in the motion parameters for the Mini Hot Spot phantom, likely due to an error in synchronization between the Polaris and the scanner.

Figure 1:

Transformation parameters with list-mode and optical tracking methods for phantoms’ experiments (first 90 seconds of the moving acquisitions). There was good correspondence in the transformation parameters derived from both methods, which correctly detected the periodic movements introduced by the Quasar system.

As shown in Figure 2, the MC images were similar to the static images, whereas the NMC images were blurred by the introduced motion for both phantoms. On the left panel, the rectangular ROIs drawn on the foreground and the background of the Hoffman phantom’s images are shown. They are used to calculate the bias and the standard deviation of the ratio across 15 noise realizations of 1-min acquisitions. The biases added by MC and NMC methods were −1.0% and 32%, respectively. The relative standard deviations for MC and NMC were 8.9% and 7.5%, respectively. The profile plots on the zoomed structures also show the similarity between MC and static images and the discrepancies between NMC and static images in both phantoms.

Figure 2:

Comparison of phantoms’ images from static, non-motion-corrected, and motion-corrected reconstructions. The motion caused blurring in the NMC images. The zoomed-in images and the profile plots show a good agreement between the static acquisition and the MC images for both phantoms.

Figure 3 shows the motion estimation step in one subject with a high displacement index. Sixty low-resolution images (spanning ~20 seconds of data each) from the 90–110 min post-injection list-mode data were reconstructed and registered to a common reference frame. A representative image from each minute is shown on the left panel. The images shown in Figure 3 were obtained without applying a post-reconstruction filter. However, the registration process used images smoothed by a Gaussian filter with a FWHM of 6 mm. A large translation on the z-axis is visible in frames 105, 108 and 109 minutes. The rigid registration parameters are also shown on the right panel. A z-translation of more than 8 cm is detected in the last two minutes.

Figure 3:

List-mode motion estimation in one subject with notable motion. Low-resolution images from the ultra-fast list-mode reconstruction are shown for the 90–110 min window. Notable head motion is visible in the last two images confirmed by the transformation parameters reporting an 80-mm translation along the z-axis starting at 108 min post-injection.

Figure 4 shows the histogram of the displacement index for all 139 scans included in this study. A displacement index threshold of 4 mm was chosen based on the histogram to identify scans with substantial motion. Although only 14% of scans exhibited notable motion, its effect on longitudinal studies is magnified when it involves at least one scan. In our study, based on the defined index threshold, motion affected 25% of all subjects, with 48% of those with three time points. The displacement index was significantly higher in MCI/AD subjects than in CN subjects (p<0.05 for a two-sample one-sided t-test). Motion affected subjects regardless of their age: the mean age of subjects with high displacement index is statistically equivalent to that of those with lower index (p<0.05 for a Two One-Sided t-Tests (TOST) equivalence test, assuming equivalence bounds of ± one standard deviation).

Figure 4:

The histogram of the displacement index for all scans is shown along with the threshold used to identify subjects with notable motion.

Figure 5 shows the final reconstructed images (with and without motion correction) for the dataset presented in Figure 3. The left panel zooms on the lateral ventricle region (delineated) and shows the SUVR in that region for the MC and NMC images. The MC value is smaller, which is expected for a cold region. In the NMC image, the large z-translation introduced signal from the surrounding areas into the ROI, leading to higher SUVR.

Figure 5:

Comparison between MC and NMC images for one subject with notable motion. The blurring introduced in the NMC image led to a higher SUVR in the lateral ventricle region (delineated). The region borders are well-defined in the MC image, and the lower SUVR is expected for a cold region.

In Figure 6, the longitudinal SUVR for one subject (age 36, CN) with three scans is plotted on the left panel. High variations of SUVR are present in the NMC values (dashed lines). Considering the age of the subject and their cognitively normal status, it is not expected to see an increase of 69% between the first and second scans in the entorhinal region. The subsequent decrease in SUVR strengthens this idea since tau buildup theoretically only increases with time in the absence of treatment. The second scan had a high displacement index (8.5 mm) as demonstrated by the motion parameters reported on the right panel. The substantial motion is the reason behind the unusual values seen in the NMC images. The findings of the MC method are more consistent with a young subject with no cognitive complaints. In this case, MC did not lead to the expected positive slopes for longitudinal SUVR, but it lessened the negative change to an amplitude small enough to be within the test/retest margin of error (Salinas et al., 2020).

Figure 6:

Comparison between MC and NMC longitudinal SUVR (left panel). Notable motion, confirmed by the motion parameters shown on the right, occurred in the second scan leading to unexpected high variation in the measured tau burden from the NMC images. MC processing decreased the variation between time points.

In longitudinal studies with two time points, MC reduced the standard deviation of the yearly relative change (δ) across subjects by −49%, −24%, −18%, and −16% in the entorhinal, inferior temporal, precuneus, and amygdala regions, respectively. Reducing standard deviation translates to a smaller sample size needed to achieve the same statistical power in clinical trials. To inspect the large reduction in the standard deviation, we investigated the distribution of the yearly relative change from the NMC images. We isolated two outliers where motion in one scan led to extreme values for the rate of tau accumulation. An example was shown in Figure 6. We excluded those outliers and recalculated the standard deviation from the remaining subjects. All values for the standard deviations are reported in Figure 7.

Figure 7:

The standard deviation of the yearly relative change across subjects for NMC, MC, including, and excluding the two outliers. MC decreases the standard deviation across regions, but the reduction is smaller when subjects with notable motion are excluded.

The standard deviations from the MC approach are still smaller than NMC, but the differences are decreased when subjects with very large motion are excluded.

Discussion

In this work, we applied a list-mode reconstruction method using ultra-short frames to estimate motion followed by a full event-by-event motion-corrected list-mode reconstruction in two phantom experiments as well as a large dataset of [¹⁸F]-MK6240 PET studies. In the phantom studies, the proposed approach was able to detect and correct for continuous motion by utilizing very short frames (~0.3 seconds) for the motion estimation. In human studies, visual inspection showed blurry NMC images compared to improved MC images for scans with notable motion (i.e., higher displacement index). There was no difference visually between MC and NMC images for scans with minimal or no motion. In this study, we found that notable motion (defined as a displacement index higher than 4 mm) occurred in 14% of scans and affected 24% of studies with two time points and 48% of studies with three time points. Therefore, motion correction is more important for longitudinal studies by keeping a subject’s data usable even if one scan exhibits notable motion. Similarly, the effect of motion correction depends on the targeted population/disease: the displacement index which we used to quantify motion amplitude was significantly higher for MCI/AD subjects when compared to CN subjects. As such, the need for motion correction might be greater for a study restricted to subjects with dementia.

In the motion estimation step, the reference image was obtained by summing several short frames starting from the beginning of the list-mode data to be reconstructed. This approach is needed to decrease the noise level in the reference frame and to ensure robust registration, but it ignores possible motion at the start of the data. This work only focused on the 90–110min post-injection window where the activity distribution change is more gradual. Future work including the early frames immediately following tracer injection is needed to investigate the registration robustness with faster activity distribution change. A possible alternative would be to register the short frames to their adjacent frames instead of fixing a unique reference frame for the entire window. A more in-depth study into the minimal duration of the short frames would be needed in that case, as the noise in the images might become too high and lead to registration failure.

We demonstrated how motion correction can reduce the SUVR change variance in key brain regions. However, the reduction in variance was driven mostly by two subjects with notable motion resulting in unusual SUVR changes derived from the NMC images. One might argue that excluding high motion subjects (as demonstrated by the images from the scanner) would be an easier solution. In that sense, motion correction is a useful tool to keep valuable data, especially in expensive studies. The change in variance also depends on the way we calculate the change in SUVR. The yearly relative change is expressed as a percentage of the baseline value which adds another source of variability across subjects. Our study’s population covered a wide range of ages and diagnoses, and this physiological variability influences the resulting standard deviation and as such the clinical trial’s required sample size. Other factors include the region of interest, the reference region used for SUVR calculation, and the inherent variability in PET measurements. Motion correction cannot address all these factors, however, as demonstrated in this work, motion correction can reduce the required sample size when notable motion is present.

In future work, we would like to take advantage of the full dynamic scan available to us to reconstruct all the dynamic frames with and without motion correction and compare the effect of motion correction on the ratio of distribution volume (DVR) and its longitudinal change. We will have to tackle the problem of rapid activity distribution change in the early frames and the optimal duration of the short frames.

Conclusion

Motion is a major confounding factor in [¹⁸F]-MK6240 tau PET imaging. In this work, we executed a list-mode-based motion correction method for two phantom studies where motion was artificially introduced and tracked with a Polaris camera. We validated the method’s performance against optical tracking data. We then applied the approach to 139 scans and studied the effect of motion on the estimation of the rate of tau accumulation from longitudinal studies. We found that 14% of the scans exhibited notable motion as demonstrated by the proposed motion metric, the displacement index. The motion-corrected images demonstrated decreased blurring in cases with high displacement index and led to better quantitative results when compared to non-motion-corrected images. Motion correction greatly reduced the variance in the estimated rate of tau accumulation as measured by the change of SUVR between two time points, potentially allowing for a smaller sample size in longitudinal clinical trials evaluating the effect of a candidate drug against AD.