Controlling motion artefact levels in MR images by suspending data acquisition during periods of head motion

Head movements are a major source of MRI artefacts. Prospective motion correction techniques significantly improve data quality, but strong motion artefacts may remain in the data. We introduce a framework to suspend data acquisition during periods of head motion over a predefined threshold.


| I NT ROD UCTI ON
Head movement leads to severe artefacts in MR images that hamper the diagnostic assessment of brain MRI scans 1 and have a strong impact on morphological, 2 diffusion-based, 3 and functional 4 measures of the brain. Quite naturally, the effects of head movement are biased toward noncompliant populations, which limits the use of MRI for the study of brain pathology, an essential cornerstone of clinical neuroscience.
Motion correction can be implemented using prospective or retrospective methods and using passive or active markers of head position. 5 Prospective motion correction (PMC) techniques act in real time during image acquisition by adjusting the MRI scanner to account for patient movements. 6 With PMC methods, the acquired MRI data is already corrected for motion effects, avoiding the need for dedicated processing steps. 7,8 In this study, we use the PMC method introduced by Zaitsev et al., 6 based on external optical measures of brain motion using the tracking system introduced by Maclaren et al. 9 (KinetiCor, HI, Honolulu). This technique may be implemented with minimal changes to pulse sequences but requires specially dedicated software libraries to integrate the motion information into the software environment of the MRI scanner. A functional MRI study using this system has reported improvements in SNR by 30% to 40% and an increase in the numbers of significantly activated voxels by 70% to 330%. 4 Improvements in the precision of relaxometry measures in the range of 11% to 24% have also been reported using this PMC technique. 10 The lag time between tracking of the marker position and acquisition update (32 ms 6 ) can induce erroneous motion correction for fast motion, imposing an upper limit on head velocities that can be accurately corrected. 6,7 To overcome this limitation, we introduce a framework that allows the interruption of data acquisition during periods of head motion detrimental to image quality. Data acquisition is resumed automatically by the MRI scanner following motion periods. A similar strategy has been introduced recently to allow reacquisition of data for excessive head motions 6,[11][12][13][14] or motion artefacts. 15,16 Here, we base our strategy on a predictor of the impact of head motion on image quality, which accounts for the amplitude of the signal acquired at the time of the motion as well as head speed. This predictor allows suspension of data acquisition with optimal efficiency by continuing the sampling process during periods of strong motion when the amplitude of the acquired signal is low. Importantly, the presented framework allows the user to set the maximal amount of image degradation deemed acceptable in the MRI data. This determines the threshold for triggering of the suspension of data acquisition and governs the tradeoff between the quality of the MR images and the extension of the scan duration due to the suspension. This value can be adjusted on a studyspecific basis depending on the compliance of the population of interest and the desired level of image quality.

| Predictor of motion impact
To obtain a predictor of image degradation from motion trajectories, we follow the framework introduced by Todd et al. 4 They highlighted the importance of the amplitude of the signal acquired at the time of the motion in developing aggregate measures of motion impact over the whole duration of the acquisition. The encoding-weighted integrated motion metric (Mew) is defined as: where S i represents the head speed at the time of the acquisition of the ith k-space line and W i is the weight of the ith kspace line to Mew-the aggregate measure of motion impact over the whole 3D acquisition. For the FLASH acquisition used here, the total number of acquired lines N is the product of the number of phase-encode steps in the phase and partition (secondary phase-encoded) directions.
In line with Ref. 4, we take W i as the norm of the ith acquired line, calculated by taking the Fourier transform of the magnitude image and summed over all readout points. Figure 1A shows an example set of weights that show high-/ low-frequency variations due to the inner(partition)/outer (phase) phase-encoded directions, sampled using a Cartesian trajectory. S i is calculated as in Ref. 4: where v i and x i are the translation and rotation velocities of the center of the FOV of the encoding box (in mm/s and deg/s), computed from the optical camera data as the difference in position and orientation between the (i-1)th and ith timeframes divided by the corresponding time interval. Figure 1C shows the time course of Mew i calculated from the weights and head speeds of Figures 1A and 1B. Because the weights vary by several orders of magnitude throughout the acquisition, a given head speed may lead to values of Mew i , which may vary by the same amount depending on the time of occurrence of head motion. This forbids the use of a unique threshold value of Mew for the whole acquisition to detect head motion and trigger the suspension of the acquisition. Instead, this threshold value should be defined by considering deviations of Mew i (DMew i ) from its values when no significant head motion is present: Here, Mew compliant,i is the value of Mew i in compliant populations. Significant motion can then be identified in real time as yielding values of DMew i above 0, regardless of the occurrence of the motion in k-space (see Figure 1D).
Our definition of Mew compliant,i should account for all possible values of S i due to noise in the estimation of the marker position/orientation and head motion of physiological origin (e.g., heartbeat and respiration). The former represents an intrinsic limit of the PMC system (typically a few microns 9 ), and the latter can be accurately corrected by the  Magnetic Resonance in Medicine PMC system and does not need to be addressed here. Therefore, Mew compliant,i was computed from: 1. A set of reference weights obtained from PDw and T 1 w images acquired in 10 compliant participants (same acquisition sequence and image resolution as the current study) to account for the variability in weights across a population. These 10 participants (5 females and 5 males, age mean/SD: 37.5/13.8) were independent from those involved in the motion experiments.
2. Samples of S obtained from 3 motion time courses of 5 compliant participants. A histogram of these samples of S is shown in Supporting Information Figure S1.
Multiple sample values of Mew compliant,i were obtained by combination of the reference weights with all possible permutations of the S samples. The reference Mew compliant,i values were obtained by taking the mean across the Mew compliant,i values and adding 3 SDs. During image acquisition, the instantaneous value of Mew i was calculated from the product of the instantaneous head speed with the corresponding weight, calculated by taking the mean across the reference weights. These mean reference weights, calculated from an independent dataset as described above, were used for all study participants and motion experiments. Periods of motion are identified in real time as those yielding values of DMew i above 0 (see Figure  1D). From these values, an aggregate measure of motion degradation over the entire acquisition can be calculated using: The aggregate measure of motion degradation DMew is the sum of all the values of DMew i above 0 encountered during the acquisition.

| Real-time suspension of data acquisition during periods of motion
In the first part of this study, we show the relationship between the quality of the acquired data (SD of R Ã 2 [SDR Ã 2 ]: see marker of image degradation section below) and DMew. This relationship establishes DMew as a predictor of the impact of head motion on image quality. In the second part of this study, we use this predictor to monitor in real time during data acquisition the impact of head motion on data quality and pause the acquisition when the predicted degradation exceeds a predetermined threshold.
Although in principle the quality of an MRI image cannot be predicted from a single time point alone, we consider the maximum tolerable image degradation and distribute this value over the longest likely duration of the motion periods.
According to motion behaviors reported from epileptic patients, 17 a maximum of 49 jerks may be expected over the duration of the acquisition of 1 image volume in the present study (7 min 21 s). This corresponds to a maximum of 147 s of movement during our acquisition (3 s per jerk), equivalent to the acquisition of 147 s/TR 5 6000 k-space readout lines. Given the maximum aggregate value of DMew observed in our experiments ($12 mm/s) (see Figure 2), the largest (instantaneous) value of DMew i may be 12/6000 5 2e -3 mm/ s. Motion experiments were therefore conducted with suspension of data acquisition when DMew i was exceeding threshold values DMew th 5 1.75e -4 , 3.5e -4 , 7e -4 , 1.4e -3 , and 2.8e -3 mm/s, to cover all possible scenarios of patients motion behavior.
Note that suspending data acquisition based on the value of DMew i is equivalent to suspending the acquisition based on the head speed according to: Figure  1E). With DMew th 5 0, the suspension is most stringent: any head speed beyond those of the compliant population leads to suspension of data acquisition, regardless of the location in k-space. With DMew th > 0, the threshold value of the head speed, S th , depends on the amplitude of the signal acquired at the time of the motion (see Figure 1E): S th is high/low when the amplitude of the acquired signal is low/ high. Note that to avoid excessive additional scan times in case of continuous participant motion, the total duration of the suspension period was not allowed to exceed the nominal duration of the scans. Upon reaching this limit, the scanner was required to resume scanning without further suspension, leading to a doubling of the acquisition time. This situation was never encountered in our experiments.

| Marker of image degradation
Maps of the MRI parameter R Ã 2 were computed from the PDw data acquired in the motion and no-motion conditions (see data acquisition and motion experiments below). The R Ã 2 maps were obtained by voxel-wise linear fitting of the log of the signal amplitudes across echoes, as described in Refs. 18,19. Because head motion leads to deviations from the exponential decays, 20,21 maps of the SD of the residuals E (standard error of the R Ã 2 estimates) were calculated as measures of the goodness of fit of the monoexponential model. When low levels of motion artefact are present in the MRI data, R Ã 2 values are known correlates of iron and myelin concentration in brain tissue. 22-24 R Ã 2 maps are known to be particularly sensitive to motion. 20 Head motion increases the spatial variability of R Ã 2 maps, and region-specific coefficients of variation of R Ã 2 have been shown to characterize image degradation due to head motion. 10 Building on this idea, we used estimates of the SDR Ã 2 values in white matter as markers of motion artefact levels in the acquired data. The white matter masks were calculated by segmenting the magnetization CASTELLA ET AL.
transfer maps using SPM's unified segmentation 25 and including voxels with a minimal white matter probability of 0.95. The white matter masks were eroded by 2 voxels to reduce the contribution of regions of inhomogeneous B 0 field to the SDR Ã 2 estimates. This white matter mask was also used to calculate the mean E estimates in white matter.

| Data acquisition
Data were acquired on 5 participants on a 3T Siemens Prisma MRI scanner (Erlangen, Germany) using a 64channel and a 20-channel head-neck coil and a custom-made 3D FLASH acquisition. 18,26 The k-space trajectory of the FLASH sequence was Cartesian-centered (i.e., the center of k-space was acquired halfway through the acquisition). The image resolution was 1.5 mm isotropically, leading to a voxel size of 1.5 3 mm 3 . The matrix size was 160 3 150 3 120, and the acquisition time was 7 min 21 s per image volume. The TR was 24.5 ms. Eight echo images were acquired with a bipolar readout (readout bandwidth: 466 Hz/pixel) following each (nonselective) RF excitation, with TEs ranging from 2.34 ms to 18.72 ms in steps of 2.34 ms.
Prior to the motion experiments (see below), subjects were requested to remain still. During this period of no motion, data were acquired with proton density (PDw), T 1 (T 1 w), and magnetization transfer-weighted contrast (RF excitation flip angles: 68, 218, 68, respectively). 18 Magnetization transfer-weighting was achieved using an off-resonance Gaussian-shaped RF pulse applied prior to RF excitation, typical amplitude: $150 to 200 V, duration: 4 ms, nominal flip angle: 220 8, frequency offset from water resonance: 2k Hz. 18 B 1 mapping data 27,28 was used together with the FLASH data to compute quantitative magnetization transfer maps offline according to Ref. 29. Only PDw data was acquired for each condition of the motion experiments.
Data acquisition was repeated across multiple scanning sessions for each motion experiment, with and without suspension of data acquisition and with different head coils. The subjects were taken out of the MRI scanner between sessions, and scanner calibration (e.g., shimming) was only conducted at the start of each session. Magnetic Resonance in Medicine detected by the camera (tracking accuracy $0.01 mm in translation and $0.01 8 in rotation 9 ) was sent in real time to the MRI host computer to adjust the MRI acquisition. 9 Before each experiment, the participants were asked to move in all directions to ensure optimal coverage of the marker position by the camera. As a precautionary measure, the MRI acquisition sequence was programmed to suspend data acquisition when the marker position was not available (indicating that the marker was out of the FOV). The largest translations and rotations observed during the experiments with respect to the initial pose were x 5 9.9 mm, y 5 6.8 mm, z 5 7.3 mm, Rx 5 9.2 8, Ry 5 8.1 8 and Rz 5 9.1 8. Adjustment of the MRI scanner components was implemented according to the latest head position information available from the camera prior to each RF pulse (frequency 5 1/ TR). Note that RF excitation was maintained during periods of suspension of data acquisition to preserve the steady-state equilibrium of the magnetization. Data acquisition was suspended when the marker was out of the FOV of the camera to ensure optimal motion correction of all data points.

| Experiment 1
The volunteers were instructed to move during periods of 44 s, that is, 10% of the image acquisition time. The motion experiments were conducted using 5 conditions, which covered the first, second, third, fourth, and fifth 10% of k-space. As a result, condition 5 involved motion during the acquisition of the center of k-space, leading to maximum image degradation. This experiment was performed using a 64channel head coil (see Table 1).

| Experiment 2
The volunteers were instructed to move during periods of 3 seconds (jerks) to replicate motion behaviors reported from epileptic patients. 17 Data acquisition was repeated for 4 conditions, with 10, 20, 30, and 40 jerks per acquisition. The onset times of the jerks were randomized for each acquisition. This experiment was performed using a 64-and 20channel head coil (see Table 1).
For both motion experiments 1 and 2, the order of the motion conditions was randomized. No instructions were given to the participants on the type of motion that they should carry out. Both experiments were first conducted with no interruption of data acquisition (but with PMC) to study the F IGUR E 3 Example of R Ã 2 maps acquired without (A) and with (B) suspension of data acquisition during periods of head motion (experiment 1). In (B), the threshold value used for suspension of the acquisition was DMew th 5 0 and the points of suspended acquisition are highlighted in red (11% and 12% of the total number of points). The aggregate measures of motion impact, calculated from the motion trajectories, are indicated on top of each R Ã 2 map

| Magnetic Resonance in Medicine
relationship between motion history and degradation of the R Ã 2 maps (SDR Ã 2 ), despite conventional PMC. From this relationship, a predictor of the impact of motion on image quality was identified. Both experiments were then repeated with interruption of data acquisition when different threshold values of this predictor were exceeded. Note that 1 subject did not take part in the experiments with acquisition suspension.

| R ES ULT S
Without suspension of data acquisition, the degradation of the R Ã 2 maps with head motion follows the same linear dependence on DMew for all subjects and under both block and jerk-like motion conditions (Figures 2A, B). This was observed for both head coils, although less robustly for the 64-channel coil (R 2 5 0.51, slope 5 0.68 mm 21 ) than for the 20-channel coil (R 2 5 0.9, slope 5 0.39 mm 21 ). Figure 2C shows example R Ã 2 maps for different values of DMew. The degradation of the R Ã 2 maps increases when motion takes place near the center of k-space (experiment 1) or with an increasing number of jerks (experiment 2). The good agreement of the relationship between DMew and SDR Ã 2 for both motion experiments establishes DMew as a summary measure of motion impact, aggregated over the whole acquisition, regardless of the motion trajectories.    Figure  2 for reference. In Figure 3B, the time points corresponding to suspension of data acquisition are highlighted in red and accounted for 11% and 12% of the total number of points. The suspension of data acquisition leads to the largest improvements in data quality when motion takes place during the acquisition of the center of k-space. The quality of the R Ã 2 maps under the motion conditions remains close to that of the R Ã 2 maps acquired without motion. Table 2 shows the increase in scan time due to the suspension of data acquisition and the corresponding values of SDR Ã 2 (experiment 1). With DMew th 5 0 mm/s, optimal image quality is preserved for all motion conditions at the cost of a large increase in acquisition time (45 s-50 s, the duration of the motion period in experiment 1). When DMew th 5 7e -4 mm/s, increasing image degradation is observed as motion takes place near the center of k-space. The additional scan duration is short/long when motion takes place at the periphery/center of k-space (conditions 1 of 5). Figures 4A and 4B show the effect of suspending data acquisition for DMew th 5 3.5e -4 , 1.4e -3 , and 2.8e -3 mm/s. For a given motion condition, the maximum DMew and SDR Ã 2 values decrease when DMew th is reduced, illustrating a gain in image quality. As for Figure 2, the robustness of the linear dependence of SDR Ã 2 on DMew is less pronounced for the 64-channel coil (R 2 5 0.46) than for the 20-channel coil (R 2 5 0.81). Figure 4C shows the R Ã 2 maps with the highest SDR Ã 2 for each threshold value (20-channel coil and experiment 2). The improvement in image quality when DMew th is reduced is clearly apparent. For DMew th 5 3.5e -4 mm/s, the quality of the R Ã 2 maps (SDR Ã 2 5 4.35 s 21 ; DMew51e -1 mm/ s) is close to that obtained under the no motion condition (SDR Ã 2 5 3.08 s 21 ; DMew 5 1e -3 mm/s). Figures 5A and 5B show the relationship between the residuals E and DMew for DMew th 5 3.5e -4 , 1.4e -3 , and 2.8e -3 mm/s. Figure 5C shows the maps of E corresponding to the highest SDR Ã 2 values of DMew th . Without suspension of the acquisition, E follows a similar linear dependence on DMew as SDR Ã 2 ( Figure 4). For low values of DMew th , the values of E tend to remain near the values of the no-motion condition. The range of E values increases with increasing DMew th but remains smaller than those obtained without acquisition suspension.

| DI S CU S S IO N
In this study, we improve the quality of MRI data acquired with prospective motion correction by suspending data acquisition during periods of detrimental motion. We present F IGUR E 5 Relationship between mean residuals of the R Ã an aggregate measure of motion history during the acquisition that exhibits a high correlation with the degradation of the resulting image ( Figure 2). From this predictor of motion impact, a threshold parameter is defined to suspend image acquisition during motion periods (Figure 3). This threshold guarantees a minimum level of data quality in the final images, given the motion characteristics of patient populations (Figure 4). This threshold also governs the tradeoff between image quality and the prolongation of scan time due to the suspension of the acquisition (Table 2) and can be set by the user depending on the desired level of image quality and patient tolerance for prolonged scan durations.
The relationship between the aggregate measure of motion and image degradation (Figure 2) was examined for 2 types of head movements: continuous motion periods of 44 s with predefined onset times (experiment 1) and short motion periods of 3 s (jerks) at random times during the scans (experiment 2). The comparability of the participant's motion behavior between motion conditions was not a prerequisite for this study: no instructions were given to the participants regarding their motion behaviors, and in experiment 2, the onset times of the jerks were randomized across repetitions. Despite this variability, the relationship between image degradation and the motion impact remained remarkably consistent, suggesting that the predictor of motion impact may hold for a large range of motion behaviors (this relationship was preserved when the TR duration of the acquisition was extended to 50 ms; data not shown). However, the correlation between image degradation and motion impact was reduced with the 64-channel coil. The numerical simulation results presented in Supporting Information Figure S2 suggest that head motion in the sharp sensitivity profile of this coil might be the origin of the effect, which is in line with recent literature. 30,31 The delay between tracking of the marker position and acquisition update on the measures of DMew ($ twice our TR value 6 ) was estimated to lead to an average (SD) change in DMew of 0.017 (0.030) mm/s, small compared to the range of DMew values encountered due to motion (see Supporting Information Figure S3).
Aside from head speed, the predictor of motion impact accounts for the time of occurrence of the motion in k-space. As a result, periods of strong motion may be tolerated at innocuous times of the image encoding order to minimize the increase in acquisition time at a minimal cost in terms of image quality. The prolongation of scan time due to the suspension of data acquisition was estimated from experiment 1, for which the motion behavior may be assumed reproducible across multiple repetitions of the same condition. The results are presented in Table 2. Suspension of data acquisition based on head speed only (DM ewth 5 0 mm/s) led to systematic prolongations of scan time by the duration of the motion periods. With DM ewth 5 7e -4 mm/s, the prolongation of scan time varied by a factor 5, depending on the time of occurrence of the motion.
Although the presented framework for suspension of data acquisition during periods of detrimental head motion was designed for FLASH, it might be extended to other types of image acquisitions. However, suspension of data acquisition might in general not be recommended for functional MRI, for which the synchronicity of image acquisition and functional stimulation is paramount. The proposed framework is particularly well suited for acquisitions, for which each head motion information used for prospective correction can be associated with data of similar amplitude. However, specific adjustments might be required for each type of acquisition (e.g., FLAIR, MPRAGE). In the context of EPI acquisitions, this would require new motion information at the frequency of the EPI readout sampling ($1 kHz)-out of range of our system-and would also require careful implementation to prevent the occurrence of image ghosting. However, an implementation based on head speed only, prior to RF excitation, is likely to bring significant improvements in data quality (e.g., diffusion MRI) in noncompliant populations. Note that keeping the time interval between consecutive RF excitations constant for a given slice is critical for 2D acquisitions. A direct implementation of the proposed framework, for which RF excitation is maintained during periods of acquisition suspension, might violate this principle depending on the acquisition scheme used. An alternative strategy, as discussed below, is to complete data acquisition despite the occurrence of motion and to reacquire the data points affected by motion post-hoc while preserving the slice ordering of the acquisition.
In the presented framework, the weights used to calculate the predictor of motion impact DMew were taken as an average across an independent population. In Supporting Information Figure S4, these DMew estimates are compared to those obtained from the individual weights of the participants to the motion experiments. Although individual weights cannot not be used in practice to suspend the acquisition, these results provide a measure of the accuracy of the DMew estimates calculated with the proposed framework. The mean (SD) difference between average and individual weights was 215.94(23.04)%. The average (SD) deviation between the DMew values obtained from the average and individual weights was 0.031(0.062) mm/s. The largest deviations were obtained for high values of DMew th , which preferentially targets data points at the k-space center (see Supporting Information Figure S4).
The use of weights averaged across a reference population requires acquisition of data on a compliant cohort prior to implementation of the proposed framework. Although this task needs to be completed only once, an alternative approach based on the relationship between the amplitude of the weights and their distance to the origin of k-space might | 2423 Magnetic Resonance in Medicine be preferable. Based on the observed relationship (see Supporting Information Figure S5A), alternative weights were computed and used in offline analysis to calculate new estimates of DMew from the motion trajectories recorded during the experiments. The results are presented in Supporting Information Figure S5B-D. The average (SD) difference between the DMew values calculated from the individual weights (our gold standard for accuracy assessment) and these alternative weights was 0.048(0.087) mm/s. This is $50% larger than the accuracy estimates reported above for the proposed framework but remains small compared to the range of DMew values encountered due to motion.
This study illustrates the improvements in data quality when low spatial frequencies are preferentially targeted to trigger the suspension of data acquisition (Figure 4). These results open the way for follow-up studies comparing the real-time data suspension presented here with other strategies to minimize motion impact. On the model of previous works, 32-34 1 candidate approach is to adjust the k-space trajectory of the acquisition in real time depending on the motion behavior. With this approach, regions of k-space with low signal amplitude (high k-space values) would be acquired during periods of head motion, with little effect on image quality. This approach would alleviate the need for suspension of data acquisition, keeping the acquisition time minimal for all motion behaviors. Another alternative, using the theoretical framework presented here, would be to complete the acquisition of the whole image without suspension and re-acquire post-hoc the k-space lines most affected by motion until a target DMew is reached. This strategy has been implemented in previous studies by identifying k-space lines corrupted by motion based on motion score 12,13 or magnitude/phase data. 15,16 This approach would circumvent the need for assumptions on the expected amount of head motion to determine threshold values from aggregate measures of motion impact (see real-time suspension of data acquisition during periods of motion section above), and would adjust naturally to the different motion behaviors encountered. Note that this strategy implies that the acquisition of neighboring data points might be separated by long time intervals, which could degrade the phase consistency of neighboring k-space lines and exacerbate image degradation (e.g., ghosting) when dynamic effects are present during the acquisition. 32,33 Here, we use the SD of R Ã 2 maps in white matter (SDR Ã 2 ) as a marker of image quality. We chose this parameter because it can be computed with ease, because of its sensitivity to motion-related image degradation, 10,20 and because quantitative maps of MRI parameters do not suffer from the sources of image bias that commonly affect anatomical MRI data. 18 Other markers of motion artefacts might be equally suitable. 35 R Ã 2 maps are sensitive to the orientation of the head relative to the main magnetic field 6 and to the inhomogeneities of this field near the air-tissue interface. 37 The resulting variability in the SDR Ã 2 values can be estimated to be up to 1 s 21 from the data acquired across multiple scanning sessions in the absence of head motion. This is about 5 times smaller than changes in SDR Ã 2 due to head motion, which shows a remarkable level of correlation with DMew nonetheless. Note that the contribution of voxels near the air-tissue interface was minimized by eroding the white matter masks by 2 voxels prior to the calculation of the SDR Ã 2 estimates.
On top of motivating the use of DMew for suspension of data acquisition in real time, the linear increase in SDR Ã 2 with DMew also suggests that SDR Ã 2 might be used as an objective measure of data quality, potentially more robust than visual assessment. However, it should be emphasized that SDR Ã 2 depends on the level of noise in the acquired data (image resolution) and on the robustness of the R Ã 2 fitting (number of echoes). It should also be noted that because R Ã 2 is a biomarker of the microstructural properties of brain tissue, 22-24 SDR Ã 2 values might be affected by microscopic changes in the healthy and diseased brain.

| CON CLU S IO N
We present a framework that allows the suspension of MRI data acquisition during periods of head motion. In this framework, head motion information is provided by a prospective motion correction system during the scans and is processed in real time to compute a predictor of the impact of motion on the MR images. Data acquisition is suspended when excessive motion degradation is predicted. The predictor of motion impact accounts for both head speed and the amplitude of the k-space signal acquired at the time of the motion, which allows for a more efficient reacquisition strategy than that solely based on head speed. The framework was tested for 2 different kinds of motion behavior. Suspending data acquisition during impactful periods of head motion led to significant improvements in the quality of the MRI images. The tradeoff between image degradation due to motion and prolongation of scan time due to the suspension of data acquisition may be tuned depending on the desired MRI image quality and subject tolerability.

ACKNOWLEDGMENTS
This work was carried out on the MRI platform of the D epartement des neurosciences cliniques-Centre hospitalier universitaire vaudois, which is generously supported by the Roger De Spoelberch and Partridge Foundations. A.L. is supported by the Roger De Spoelberch Foundation. B.D. is supported by the Swiss National Science Foundation (NCCR Synapsy, project grant Nr 32003B_159780), Foundation Parkinson Switzerland, and Foundation Synapsis. The research leading to these results has received funding from the European Union's Horizon 2020 research and innovation program under grant agreement No. 720270 (HBP SGA1). The authors would like to thank Dr. David Carmichael (Institute of Child Health, University College London, UK) for helpful discussions on the movement characteristics of epileptic patients.