Quantitative validation of data-driven motion correction for brain PET using phantom with motion generator system Authors: Yuto Kamitaka, Muneyuki Sakata, Keiichi Oda, Akie Katsuki, Hirofumi Kawakami, Kei Wagatsuma, Masato Kobayashi, Kenji Ishii Published: 2025 ABSTRACT AbstractBackgroundHead motion during brain positron emission tomography (PET) degrades image quality and quantitative accuracy. Therefore, a data-driven motion correction (MC) method utilizing ultrafast list-mode reconstruction technology has been proposed and shown to considerably improve image quality. However, reproducing accurate actual motions and motion-free images using clinical data alone remains challenging. This study aimed to quantitatively evaluate data-driven MC using a brain phantom for known tracer distributions and a custom-made motion generator system for variable known motions.MethodsHoffman 3D brain phantom was filled with 20 and 3 MBq of [18F]fluoro-2-deoxy-D-glucose (FDG) to simulate high- and low-radioactivity conditions corresponding to brain FDG PET and amyloid PET acquisitions, respectively. Two separate phantom measurements were performed accordingly. Motion simulation was conducted using a custom-designed motion generator, incorporating 15° and 30° rotations about the z-axis, 3° and 6° rotations about the x-axis, and 5 mm and 10 mm translations along the z-axis in the PET image coordinates. The data-driven MC was applied with frame durations of 1, 2, 5, 10, and 20 s for motion estimation. The estimated motions were compared with the motions measured using an external optical tracker system. %contrast and gray matter coefficient of variation (CV%) were calculated from the motion-corrected PET images.ResultsThe motion generator system successfully reproduced the designed motions. Motion estimation remained stable under high-radioactivity condition but showed reduced stability under low-radioactivity condition, particularly with shorter frame durations. Under both conditions, longer frame durations led to underestimation of continuous motion. The data-driven MC improved %contrast and gray matter CV% across all conditions, with shorter frame durations providing better correction for quick or continuous motions. However, shorter frame durations increased statistical noise, especially under low-radioactivity condition.ConclusionThe data-driven MC effectively improved the quality of motion-affected PET images under both high- and low-radioactivity conditions, indicating its broad applicability. However, correction accuracy deteriorated under the lower-radioactivity condition.Keywords: Head motion, Data-driven motion correction, Image quality, [18F]FDG, Amyloid PET, Dementia, Alzheimer’s disease Introduction Brain positron emission tomography/computed tomography (PET/CT) is a minimally invasive imaging modality that facilitates both visual and quantitative evaluation of brain metabolism, amyloid-β deposition, and tau aggregation as core biomarkers for dementia diagnosis [1–3]. The recent approval and clinical implementation of Alzheimer’s disease (AD)-modifying drugs, such as lecanemab and donanemab, across various countries have increased the demand for brain PET scans. This modality employs radiopharmaceuticals, including [18F]fluoro-2-deoxy-D-glucose ([18F]FDG), [18F]florbetapir, [18F]flutemetamol, and [18F]florbetaben, which are essential for the early and differential diagnosis of dementia [1]. According to the guidelines of the Japanese Society of Nuclear Medicine (JSNM) and the Society of Nuclear Medicine and Molecular Imaging (SNMMI)/European Association of Nuclear Medicine (EANM), these FDG and amyloid scans require an acquisition time of 10–30 min [4–6]. Findings from the Japanese Alzheimer’s Disease Neuroimaging Initiative (J-ADNI) study indicated that approximately 80% of participants, including cognitively normal individuals, exhibited head motion during a 30-min FDG scan [7]. Furthermore, mild head motion was observed, even during shorter 10-min acquisitions [8]. Such head motion significantly degrades image quality and compromises the quantitative accuracy of PET data [7, 9–11]. Head motion causes image blurring and introduces errors in attenuation and scatter corrections because of the misalignment between CT and PET images [12, 13]. Owing to significant improvements in the spatial resolution of PET/CT through advancements in hardware and software [14–17], motion artifacts have become increasingly pronounced. The administration of AD-modifying drugs requires evidence of positive amyloid biomarkers (amyloid PET or cerebrospinal fluid test) [18, 19]. Although amyloid PET is typically assessed visually as positive or negative in clinical practice, quantitative measures, such as the standard uptake value ratio (SUVR) [20] or centiloid scale [21], have become more prevalent in clinical research and drug development for AD. For instance, quantitative amyloid PET is increasingly used in both longitudinal and cross-sectional evaluations to monitor treatment response [22]. Therefore, minimizing the influence of head motion, which degrades the PET image quality and quantitative accuracy, remains a critical challenge for achieving accurate brain PET imaging. Several methods of motion correction (MC) have been proposed. A typical hardware-based approach involves tracking markers attached to the head using external devices, such as Polaris® [12, 13, 23, 24]. Accurate reflection of motion direction and magnitude measured by external devices in PET coordinates requires precise calibration to align the external device coordinates with those of the PET/CT system [23]. This approach increases the burden on both subjects and staff [25]. Another approach involves frame-based methods, where imaging data are divided into multi-minute frames, each reconstructed separately and subsequently aligned [7, 26, 27]. However, the images used for alignment contain motion-related errors owing to inaccurate attenuation and scatter correction [12, 13]. A data-driven MC method utilizing ultrafast list-mode reconstruction technology has been developed [11, 28]. This method divides the list-mode data into short temporal frames ranging from less than a second to several seconds, and the motions are estimated using images of the short temporal frames [11, 29]. Subsequently, the estimated motions are corrected by adjusting the list-mode data, followed by image reconstruction using the same parameters as those in ordinal clinical conditions. This approach eliminates motion-related errors in attenuation and scatter corrections. Although data-driven MC has been increasingly reported [11, 27, 30], most evaluations have been limited to clinical images. Evaluations using clinical images are practical; however, they rely on relative improvements in qualitative and quantitative image attributes, as accurately reproducing absolute motions and motion-free images from clinical data alone is challenging. A study using artificial simulated motions applied to clinically measured list-mode data evaluated registration accuracy under short frame durations [29]; however, the MC performance with real measured motions should be quantitatively verified. Therefore, this study aims to quantitatively validate data-driven MC using phantom measurements with predefined motions and known tracer distributions. Our custom-made motion generator system realized more variations in the designed motions compared to the study by Odagiri et al. [31], and the phantom’s radioactivity concentrations were configured to reflect the clinical brain FDG and amyloid PET imaging conditions. Materials and methods Motion generator system Figure 1 shows the motion generator system, comprising three stepper motors controlled by a microcontroller and nonmetallic components fabricated using a 3D printing system. Part of this system was designed using computer-aided design (CAD) software FreeCAD ver. 0.19 (https://www.freecad.org) and printed using Guider 2 (FLASHFORGE, Hangzhou, Zhejiang, China). For the simulation of head motion, the system could rotate about the x- and z-axes and translation along the z-axis in the PET image coordinates. These are most commonly experienced head movements in clinical situations. The motion ranges were defined as 9° rotation about the x-axis, 360° rotation about the z-axis, and 30 mm translation along the z-axis. It should be noted that while the rotational axis about the z-axis ((a) in Fig. 1) coincides with the z-axis in the PET image coordinates, the rotational axis about the x-axis ((b) in Fig. 1) was parallel, but does not coincide with the x-axis in the PET image coordinates. Motion generator system for simulation head motion. The system enables rotations about the z- (a) and x-axes (b) of the PET image coordinates and translation along the z-axis (c). The marker for optical tracker was attached to the top center of the phantom cylinder Data acquisitions Two phantom measurements were performed on different days, using a Hoffman 3D brain phantom (Data Spectrum Corporation, Hillsborough, NC, USA) and a motion generator system. All PET data for the phantom were acquired using a Discovery MI PET/CT system (GE HealthCare, Milwaukee, WI, USA), a PET scanner comprising a lutetium-yttrium oxyorthosilicate (LYSO) silicon photomultiplier (SiPM) PET instrument with a 64-slice CT scanner [17]. [18F]FDG was used for radioactivity measurements, and the activities at the start of the acquisition were 20 and 3 MBq, which were defined as the high- and low-radioactivity conditions, respectively. The activities were determined by considering the estimated activities according to standard protocols used in clinical studies. The brain radioactivity level 30 min after the injection of 185 MBq [18F]FDG was estimated to be 20 MBq [4, 5, 32, 33]. Similarly, 3 MBq at the start of acquisition corresponded to the estimated lowest brain radioactivity level among the protocols with amyloid PET tracers and represented the estimated brain radioactivity levels 50 min after the injection of 555 MBq of [11C]PiB and 90 min after the injection of 185 MBq of [18F]flutemetamol [4, 5]. The motions simulated in this study were 15° and 30° rotation about the z-axis (head rotation) (phases (2) and (3)), 3° and 6° rotation about the x-axis (chin down) (phases (4) and (5)), 5 mm and 10 mm parallel translation along the z-axis (phases (6) and (7)) and 30° continuous rotation about the z-axis from side to side at 1°/s and 2°/s speed (phases (8) and (9)). Table 1 lists the order of the simulated motion and scan duration for each phase. For the rotations about the z- and x-axes and the translation along the z-axis (phases (2)–(7)), the phase was first started in the original posture and subsequently moved to the designed posture for 60 s. The acquisition continued in this posture, and the phantom was returned to its original posture at the end of the phase. Continuous rotation about the z-axis enabled movement from the start to the end of the motion phase. The scan duration per phase was adjusted, considering radioactive decay, to achieve the same number of counts in each phase and eliminate differences in image quality owing to differences in the number of counts. Because the phantom may not move exactly as programmed, the real positions of the phantom were simultaneously observed with Polaris Vicra (Northern Digital Incorporated, Waterloo, ON, Canada), an optical tracker used to evaluate motion estimation. The markers for the Polaris were attached to the top center of the phantom cylinder. Table 1Order of the programmed motion and scan durationPhaseProgrammed motionsScan duration(1)Reference (without motion)5′00”(2)15° rotation about the z-axis (Head rotation)5′10”(3)30° rotation about the z-axis (Head rotation)5′20”(4)3° rotation about the x-axis (Chin down)5′31”(5)6° rotation about the x-axis (Chin down)5′42”(6)5 mm translation along the z-axis5′56”(7)10 mm translation along the z-axis6′10”(8)30° continuous rotation about the z-axis (1°/sec)6′25”(9)30° continuous rotation about the z-axis (2°/sec)6′41” Order of the programmed motion and scan duration Motion correction and image reconstruction The PET data were motion corrected, and the images were reconstructed on a workstation running the PET Duetto reconstruction toolbox using lmDuetto ver. 2.17 (GE HealthCare) for MATLAB 2020b (The MathWorks Inc., Natick, MA, USA), based on the implementation described in prior research [11, 29]. The frame durations for the motion estimations were 1, 2, 5, 10, and 20 s. The images for motion estimation were reconstructed from the list-mode data using the maximum likelihood expectation maximization method with four iterations, without applying attenuation or scatter correction. For the motion estimations, the reconstructed images were aligned to the image of the first 120 s using rigid registration implemented as the method of “Realign” in SPM12 (the Wellcome Centre for Human Neuroimaging, University College London, UK). The motions measured by Polaris and those estimated from the alignment of the images were compared. The data for list-mode events, attenuation correction factor, scanner sensitivity factor, and estimated scatter contribution were corrected according to the estimated motions [11]. The images used for the analysis were fully reconstructed using three-dimensional ordered-subset expectation maximization (3D-OS-EM) with time-of-flight (TOF) (4 iterations, 16 subsets) and a 4 mm full width at half maximum (FWHM) Gaussian post-filter, from the data with and without motion corrections. The parameters of the reconstructed images were 128 × 128 matrix, 71 slices, 25.6 cm FOV, and 2.0 × 2.0 × 2.79 mm/voxel. Image analysis Image evaluation was based on quantitative image quality indices and visual evaluations of the reconstructed images. The recovery rate of gray-to-white matter contrast, relative to the true contrast of the Hoffman phantom, was used as a quantitative accuracy metric [4, 34]. To calculate the contrast, the JSNM region of interest (ROI) templates (Fig. 2) was used [4, 5], and the defined ROIs were applied after the co-registration of the digital phantom to the reconstructed images of the reference phase. The %contrast was calculated as follows: where GMp and WMp are the ROI mean activities of gray and white matter measured on the phantom images, respectively, and GMd and WMd are the ROI activities of the gray and white matter in the digital phantom, respectively. The gray matter coefficient of variation (CV%) was used as an additional quantitative index to assess image smoothness and the level of statistical noise. The gray matter CV% was calculated as follows: where SDGMp and GMp are the standard deviation and mean of the activities of the gray matter ROI in the phantom images, respectively. Region of interest (JSNM ROI templates; red for gray matter, green for white matter) defined in the digital Hoffman phantom to calculate %contrast and gray matter CV% [4, 5] Results Figures 3 and 4 show the motions measured by Polaris and those estimated by data-driven MC under high- and low-radioactivity conditions, respectively. In the plots for Polaris and each frame duration condition used for motion estimation, the top panels show rotational motion, while the bottom panels show translational motion. Motions measured by Polaris and estimated by data-driven motion correction (MC) at each frame duration under high-radioactivity condition. Each plot shows rotational motions about each axis (Rx, Ry, Rz) and translational motions along each axis (Tx, Ty, Tz) in the PET image coordinates relative to the original position of the marker or the origin of PET image Motions measured by Polaris and estimated by data-driven motion correction (MC) at each frame duration under low-radioactivity condition. This figure follows the same layout as Fig. 3. Each plot shows rotational motions about each axis (Rx, Ry, Rz) and translational motions along each axis (Tx, Ty, Tz) Table 2 shows a comparison of the head motion measured by Polaris and estimated by data-driven MC under high- and low-radioactivity conditions. For the rotations about the z- and x-axes and translations along the z-axis (phases (2)–(7)), the data were presented as the mean ± standard deviation (SD) of the measured or estimated angles or distances from the original posture during the stable posture following the designed motions. Stable posture intervals were determined as the time between the end of the designed motions and the beginning of the return motions to the original posture, with added margins. The duration of these intervals was in the range of 185–240 s, depending on the scan duration of each phase. For continuous z-axis rotations (phases (8) and (9)), data were presented as the mean ± SD of the absolute measured or estimated peak angles. The motions of the two acquisitions measured by Polaris were nearly identical, confirming that there was no considerable difference between the motions in the high- and low-radioactivity conditions. The means of the estimated motions matched those measured by Polaris system, confirming proper calibration of coordinates and phantom settings with the motion generator. From the stabilities of the plots in the figures and SDs in the table, the estimated motions during the stable postures were more unstable in the low-radioactivity condition than in the high-radioactivity condition, and shorter frame durations for the motion estimation worsened the stabilities in both the high- and low-radioactivity conditions. For 30° continuous rotations about the z-axis, faster motions and longer frame durations used for motion estimation resulted in underestimation of peak rotation angles. Because the rotational axis about the x-axis has offset from the x-axis in the PET image coordinates, the measured and estimated motion in the phase of the rotational motion about the x-axis (phases (4) and (5)) included the translational motions in addition to the designed rotational motions about the x-axis. Moreover, discrepancies in the translational motions along the y- and z-axes in the phases were observed because of the differences in the origins of the coordinates between the measurements by Polaris and the estimations by the MC. Table 2Comparison of motions measured by Polaris and estimated by data-driven motion correctionRotation about the z-axisRotation about the x-axisTranslation along the z-axisAbsolute peak anglein 30° continuous rotation about the z-axisProgrammed motions15°30°3°6°5 mm10 mm1°/s2°/sPhase(2)(3)(4)(5)(6)(7)(8)(9)High radioactivityPolaris14.41 ± 0.0229.35 ± 0.023.08 ± 0.066.28 ± 0.064.69 ± 0.059.77 ± 0.0529.32 ± 0.1229.13 ± 0.22MC1 s/frame14.31 ± 0.1629.09 ± 0.163.23 ± 0.136.37 ± 0.144.65 ± 0.089.78 ± 0.0829.20 ± 0.2429.00 ± 0.132 s/frame14.31 ± 0.1129.09 ± 0.123.24 ± 0.096.38 ± 0.104.66 ± 0.059.78 ± 0.0529.00 ± 0.1928.54 ± 0.195 s/frame14.30 ± 0.0829.10 ± 0.083.24 ± 0.066.39 ± 0.074.66 ± 0.039.79 ± 0.0328.33 ± 0.1526.97 ± 0.1910 s/frame14.30 ± 0.0529.10 ± 0.053.25 ± 0.046.39 ± 0.054.66 ± 0.029.79 ± 0.0226.90 ± 0.1224.24 ± 0.2220 s/frame14.30 ± 0.0329.10 ± 0.033.26 ± 0.026.40 ± 0.034.66 ± 0.019.79 ± 0.0122.45 ± 0.1517.38 ± 1.89Low radioactivityPolaris14.39 ± 0.0229.34 ± 0.023.06 ± 0.066.17 ± 0.064.93 ± 0.059.88 ± 0.0429.44 ± 0.0629.30 ± 0.24MC1 s/frame14.28 ± 0.4429.23 ± 0.423.16 ± 0.316.35 ± 0.404.87 ± 0.219.85 ± 0.2128.88 ± 0.3328.95 ± 0.412 s/frame14.27 ± 0.3129.22 ± 0.313.17 ± 0.216.37 ± 0.284.87 ± 0.169.85 ± 0.1428.84 ± 0.4028.25 ± 0.345 s/frame14.26 ± 0.1629.20 ± 0.223.19 ± 0.136.39 ± 0.154.87 ± 0.099.86 ± 0.0827.99 ± 0.1226.69 ± 0.2110 s/frame14.28 ± 0.1029.21 ± 0.133.19 ± 0.116.40 ± 0.104.87 ± 0.059.87 ± 0.0526.90 ± 0.0724.23 ± 0.1520 s/frame14.32 ± 0.0729.23 ± 0.123.21 ± 0.076.41 ± 0.064.87 ± 0.049.87 ± 0.0422.58 ± 0.1317.45 ± 1.45MC: motion correction Comparison of motions measured by Polaris and estimated by data-driven motion correction MC: motion correction Table 3 compares the %contrasts. The difference in the %contrast for the reference phase between high- and low-radioactivity conditions was limited. The %contrast without the MC was degraded owing to the motions, and considerably improved regardless of the frame duration for motion estimation with the MC. The recovered %contrasts almost reached or slightly exceeded those of the reference phase in both high- and low-radioactivity conditions, but tended to decline with increasing frame durations (10 or 20 s/frame) for motion estimation. This decline was most prominent in the continuous rotation phase about the z-axis. In the high-radioactivity condition, under single motion phases (phases (2)–(7)), the highest %contrast was achieved with a 1 s/frame duration, or the differences in %contrast between 1 s/frame and the best result from other frame durations (2 or 5 s/frame) were minimal (< 0.2%). In the low-radioactivity condition, the %contrast with 1 s/frame was not the best in all single motion phases (phases (2)–(7)), and the differences in %contrast between 1 s/frame and the best result from other frame durations (2 or 5 s/frame) were relatively larger, such as 0.9% in phase (4), compared to those observed under the high-radioactivity condition. Table 3%contrast without and with data-driven motion correction in high- and low-radioactivity conditionsReferenceRotation about the z-axisRotationabout the x-axisTranslation along the z-axisContinuous 30° rotation about the z-axisProgrammed motions–15°30°3°6°5 mm10 mm1°/s2°/sPhase(1)(2)(3)(4)(5)(6)(7)(8)(9)High radioactivityWithout MC71.5243.7230.1851.9336.8861.6145.3046.8747.93MC1 s/frame71.5172.1472.5471.5970.0871.6471.4171.8772.602 s/frame71.4972.1372.5571.6369.9671.7071.4771.8472.425 s/frame71.4972.2372.3471.5869.7371.7271.5271.6570.8110 s/frame71.5071.9271.5071.5869.6471.6771.4970.3567.2120 s/frame71.5271.6870.5370.8869.2371.5971.3665.5858.89Low radioactivityWithout MC69.8546.1631.1553.8938.8559.9447.8245.8846.78MC1 s/frame70.1167.5269.2169.9065.9968.4269.0068.3868.582 s/frame69.9667.9169.2470.2366.7768.4569.5968.3168.495 s/frame69.5767.8169.2570.8466.2668.5769.5368.0467.4810 s/frame69.6567.7468.3070.6166.2868.7069.6366.8864.1420 s/frame69.7367.3867.1768.9565.8668.1869.0063.0059.50MC: motion correction %contrast without and with data-driven motion correction in high- and low-radioactivity conditions MC: motion correction Table 4 shows a comparison of the gray matter CV%. The gray matter CV% of the reference phase in the low-radioactivity condition was higher than that in the high-radioactivity condition. The head motions led to increased gray matter CV%. For all phases, regardless of the frame duration for motion estimation, gray matter CV%s improved with MC, as the values almost reached or slightly exceeded those of the reference phase. Among the types of motion, the gray matter CV% in the phase of 6° rotation about the x-axis tended to be higher under both high- and low-radioactivity conditions. Table 4Gray matter CV% without and with data-driven motion correction in high- and low-radioactivity conditionsReferenceRotationabout the z-axisRotationabout the x-axisTranslation along the z-axisContinuous 30° rotation about the z-axisProgrammed motions–15°30°3°6°5 mm10 mm1°/s2°/sPhase(1)(2)(3)(4)(5)(6)(7)(8)(9)High radioactivityWithout MC14.6724.7934.4230.9040.9217.2731.4416.6216.93MC1 s/frame14.7015.1915.2515.3815.4914.5614.3514.5315.052 s/frame14.6815.0815.2315.3115.5414.5514.3614.5915.035 s/frame14.6615.0915.1515.2715.3914.5514.2414.6415.0810 s/frame14.6815.0715.1315.3515.6314.4914.2614.6115.0220 s/frame14.6715.1214.9115.4315.7614.5414.1614.9415.56Low radioactivityWithout MC22.8629.4539.2533.5742.5123.3835.3825.6525.72MC1 s/frame23.0622.0122.6122.7324.7221.4821.7322.7423.392 s/frame23.0222.2622.6922.7524.5121.2221.6022.8023.545 s/frame23.2022.2222.3222.5524.2121.3321.5822.4823.5010 s/frame22.9922.2522.3522.5424.7221.3421.5722.7623.1020 s/frame22.9322.1922.1922.7124.0421.0021.6122.8824.92MC: motion correction Gray matter CV% without and with data-driven motion correction in high- and low-radioactivity conditions MC: motion correction Figure 5 shows the phantom images acquired without MC, with MC, and a reference image at frame durations for motion estimations of 1 and 20 s/frame. Image noise increased in the low-radioactivity condition compared to that in the high-radioactivity condition. Images without MC were degraded owing to motion, and their degradation was recovered with MC, resulting in images closer to the reference image. For the phase of 30° continuous rotations about the z-axis at 2°/s with 20 s/frame for motion estimation, the influence of the motions remained in the reconstructed images. Comparison of phantom images with and without data-driven motion correction (MC) at 1 (upper) and 20 s/frame (lower). This figure compares phantom images under four types of head motion (A 30° rotation about the z-axis, B 6° rotation about the x-axis, C 10 mm translation along the z-axis, and D 30° continuous rotation about the z-axis at 2°/s). The red lines are aligned with the bottom edge of the brain area in the reference images as guides to facilitate the comparisons Discussion A data-driven MC provides robustness against head motions, and its ability must be evaluated through the experiments with the known motions and the known distributions. In this study, we quantitatively evaluated the MC using a custom-made motion generator system and a Hoffman 3D brain phantom. The experimental settings were designed for variations in head motion, and the conditions reflected clinical brain FDG and amyloid PET imaging. Minor discrepancies between the designed motions, which were specified as rotation amounts of the stepper motors, and the motions measured by Polaris were observed. Minor rotations or translations other than the designed motions, such as the small translations (Tx and Ty in Figs. 3 and 4) in the phases of the rotations about the z-axis (phases (2), (3), (8), and (9)), were observed. The discrepancies and errors can be attributed to the backlash in the gear system of the motion mechanisms, bending of the nonmetal materials, and misalignment of the rotation axis during the rotation of the heavy (6.5 kg) phantoms using nonmetal material support parts and gear systems. Despite these discrepancies and errors, the system demonstrated high motion reproducibility, enabling valid comparisons across varying radioactivity concentrations. In addition, the discrepancies in the translations (Ty and Tz in Figs. 3 and 4) in the phases of rotation about the x-axis (phases (4) and (5)) were owing to the difference in the origins of the measured and estimated motions. The origin of the motion measured by Polaris was at the markers of the original posture attached to the top center of the phantom, and the origin of the motion estimated by MC was at the origin of the PET image coordinates. Because the locations and orientations of the two origins were manually calibrated to have only a z-axis offset, the differences in their origins did not theoretically affect the measurements and estimations in the other phases (rotations about the z-axis and translations along the z-axis). However, the enhancements of the minor translations other than the designed motions, such as those attributed to the misalignment of the rotational axis, were observed in the measured translations by Polaris. Excluding these discrepancies, the motions measured by Polaris and those estimated by the MC showed general agreement. However, underestimations of peak rotation angles during 30° continuous rotations were observed with longer frame durations for motion estimation in both the high- and low-radioactivity conditions. The stability of the estimated motions decreased with shorter frames and lower activity conditions. In the evaluations based on the parameters calculated from the reconstructed images, the gray matter CV% of the reference phase was higher in the low-radioactivity condition than in the high-radioactivity condition. The %contrasts values in the reference phase were similar; however, higher CV% in both gray and white matter regions of the Hoffman phantom led to slightly lower %contrast under low-radioactivity conditions than in the high-radioactivity conditions. MC improved %contrast and gray matter CV% to levels comparable to those of the reference phase under both high- and low-radioactivity conditions. However, slight differences were found in the results depending on the frame duration and radioactivity concentration. When using MC, shorter frame durations for motion estimation allow for the correction of finer motions, which results in fewer counts per frame, potentially leading to unstable interframe motion estimations. Conversely, longer frame durations ensure sufficient counts per frame, but may include large intra-frame motions, resulting in the degradation of motion-corrected images. Prior research, which only considered interframe motions, suggested that the accuracies of the registrations depend on the true and scattered counts per frame, and 4 × 105 true and scattered counts per frame provide sufficient levels of accurate registration [29]. While the scanners used were different from those used in prior research, this count level was achieved with 1.8 s and 11.5 s in high- and low-radioactivity conditions, respectively, in the present study. The experimental results demonstrated that motion estimation at 1 s/frame would be sufficient in the high-radioactivity condition, and when considering the combinations of %contrast and gray matter CV%, 5 s/frame would be the best choice in the low-radioactivity condition. Considering the combined effects of inter- and intra-frame motion, these results are generally consistent with the findings of prior research [29]. Additionally, the differences in the results across frame durations were minimal, with motion estimation at 1–5 s/frame yielding similar and acceptable results under both high- and low-radioactivity conditions. Note that, in this study, motion estimation frames were defined based on frame durations to evaluate the effects of duration on motions at a fixed speed. Count-based frame definitions for motion estimation, such as 4 × 105 true and scattered counts per frame, would yield more optimized results in clinical situations, as proposed in prior research [29]. Count-based frame definitions are adaptable for radioactive decay during measurements, and these optimum frame definitions may vary depending on the physical performance of the PET/CT device. Spangler-Bickell et al. reported that applying motion correction to clinical images without motion did not affect the image quality [11]. Consistent with these findings, the phantom tests in this study demonstrated similar results, indicating that MC can be applied universally to all datasets, and not solely to those containing motion artifacts. Regarding computational time, the reconstruction of each short frame for motion estimation was proportional to the number of counts in the frame, but was much faster than the duration of each frame [28]. Thus, the frame reconstructions and motion estimations were completed within several minutes in all cases of this study. Therefore, the frame durations for motion estimation did not make much difference in the total computational time for motion estimation [29]. The most time-consuming process was the full reconstruction with MC using the estimated motions. While the main part (such as 3D-OS-EM) of the reconstruction was independent of the frame duration for motion estimation, shorter frames for motion estimation had disadvantages compared to longer ones in the process of calculating the scanner sensitivity corrections. However, the computational time could be reduced by the optimized implementation, such as grouping the short frames based on thresholds for the magnitude of estimated motions. The present study has several limitations. The current study simulated head motions as rotations about a single axis or translations along a single axis, whereas in actual clinical examinations, head motions occur as a combination of multi-axis rotations and translations, which were beyond the scope of this study. Furthermore, the maximum velocities of head motion simulated in this study were relatively slow (2°/s and 0.2 mm/s). In the clinical setting, sudden and large motions may occur at speeds that exceed these limits. Conclusions This study evaluated data-driven MC using a motion generator system and the Hoffman 3D brain phantom. The experimental results with variable settings indicated that shorter frame durations effectively captured rapid or continuous motions but frequently produced unstable estimations of postures under low-radioactivity conditions. Conversely, longer frame durations stabilized motion estimation but resulted in incomplete correction of rapid or continuous motions. Overall, the data-driven MC effectively improved the quality of motion-included images to a level comparable to motion-free images under both high- and low-radioactivity conditions. However, optimization based on the device’s physical performance is necessary. This supports the effectiveness and robustness of data-driven MC. Acknowledgements We thank Mr. Masanari Sakai, Mr. Kosuke Nishino, and Mr. Taichi Matsushita for technical support with the cyclotron operation and radiosynthesis and Mr. Matthew Spangler-Bickell for technical support with the data-driven motion correction algorithm. Abbreviations [18F]fluoro-2-deoxy-D-glucose Three-dimensional ordered-subset expectation maximization Alzheimer’s disease Computer-aided design Coefficient of variation European association of nuclear medicine Full width at half maximum Japanese Alzheimer’s disease neuroimaging initiative Japanese society of nuclear medicine Lutetium-yttrium oxyorthosilicate Motion correction Positron emission tomography/computed tomography Region of interest Standard deviation Silicon photomultiplier Society of nuclear medicine and molecular imaging Standard uptake value ratio Time-of-flight Author contributions YK, AK, HK and KI designed the study. YK, MS, KO, and KI collected data. YK and MS processed the data. YK, MS, AK, HK, KW and MK interpreted the data. YK and MS drafted the manuscript, and all the authors revised the manuscript. All the authors have read and approved the final version of this manuscript. Funding AK and HK received salary support from GE HealthCare Japan. This study was partially funded through an investigator-sponsored research grant from GE HealthCare. Data availability The datasets used and/or analysed during the current study are available from the corresponding author on reasonable request. Declarations Ethics approval and consent to participate This article does not contain any studies with human participants performed by any of the authors. Competing interests AK and HK are employees of GE HealthCare Japan. The PET Duetto reconstruction toolbox used in this study was provided by GE HealthCare. YK, MS, and KI received support from investigator-sponsored research by GE HealthCare. Footnotes Publisher's Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Data Availability Statement The datasets used and/or analysed during the current study are available from the corresponding author on reasonable request.