Facilitating diffusion tensor imaging of the brain during continuous gross head motion with first and second order motion compensating diffusion gradients

To demonstrate the feasibility of motion compensating diffusion gradient schemes in the acquisition of quality diffusion tensor images (DTI) of the brain during continuous gross head motion.


INTRODUCTION
DTI provides useful information in neurological conditions, such as multiple sclerosis, Parkinson's disease, stroke, and developmental disorders. 1,2However, diffusion-weighted (DW) imaging techniques are susceptible to macroscopic motion artifacts because of the strong diffusion gradients used. 3,46][7][8][9] As a result, patients prone to head motion are largely excluded from DTI, despite its potential as a powerful diagnostic tool.Many methods of motion prevention and correction have been proposed to mitigate signal dropout in DW acquisitions. 4,10Although motion prevention is ideal, it can be difficult to achieve.This is particularly true in many patient populations for which DTI promises to provide significant diagnostic value, including children and patients with neurologic conditions (e.g., stroke, Parkinson's disease). 1Many motion correcting techniques rely on the acquisition of multiple volumes and the elimination or reacquisition of images containing signal dropout. 8,11,12Although successful for overcoming brief and occasional head motion, such techniques rely on the ability to eventually acquire motion free images, which may not be possible for continuously moving patients.][15][16][17][18] The prevalence of motion artifacts in DW acquisitions has made DTI of the living heart exceedingly difficult.To combat these challenges, Nguyen et al. [19][20][21] have developed motion compensated diffusion gradient schemes that are robust to the bulk motion of the beating heart and have reproducibly quantified cardiac DTI parameters.To our knowledge, this approach has not yet been applied to the brain. 22n this study, we investigated the feasibility of acquisitions containing first and second order motion compensated diffusion gradients, based on the work of Nguyen et al., [19][20][21] for DTI of the brain during continuous gross head motion.Motion compensating gradient schemes were integrated into DTI acquisitions without specialized hardware or additional postprocessing steps.The resulting sequences were tested using a clinical MR scanner and compared to a standard DTI acquisition and a popular retrospective correction for extreme motion.

METHODS
0][21] Motion compensated gradients aim to refocus gradient-induced phase accumulated by moving spins.A spin in a gradient magnetic field, ⃗ G, accumulates phase according to 23 where ⃗ r 0 , ⃗ v 0 , and ⃗ a 0 are the spin's initial position, velocity, and acceleration, respectively, ⃗ M n (n = 0, 1, 2, … ) are the n th order gradient moments, defined by and  is the gyromagnetic ratio.The traditional Stejskal-Tanner (M0) diffusion gradients null the zeroth-order gradient moment, so that stationary spins are rephased under on-resonance conditions. 24From Eq. ( 1) it is evident that moving spins will have non-zero residual phase, causing signal fallout.In this work, bipolar diffusion gradients were used to provide M1 motion compensation, nulling zeroth and first order gradient moments, and B1-resistant diffusion gradients proposed by Nguyen et al. 20 were used to provide M2 motion compensation, nulling zeroth, first, and second order gradient moments.Pulse sequence diagrams of the M0, M1, and M2 sequences used in this work can be found in Figure S1.
Five healthy volunteers were scanned using a 3 T Siemens Prisma (Siemens Healthcare) and a 32-channel head coil under a Cleveland Clinic Institutional Review Board approved protocol.Whole-brain DTI datasets were acquired using diffusion prepared EPI sequences (TR/TE = 12 600/89 ms; FOV = 220 × 220 mm 2 ; 110 × 110 matrix; 2-mm slice thickness; 60 slices; 30 volumes with b = 1000 s/mm 2 , 4 volumes with b = 0).DTI data was acquired with M0, M1, and M2 motion compensating diffusion gradients for one volunteer (Study 0), and with M0 and M2 motion compensating diffusion gradients in four additional volunteers (Study 1-Study 4).Echo and repetition times were matched for all diffusion gradient schemes, resulting in an acquisition time of 7 min, 46 s for each sequence.
In the first volunteer (Study 0), two datasets were acquired for each motion compensating gradient scheme (M0, M1, M2) in a single session.The volunteer was instructed not to move during one acquisition, yielding the no-motion datasets.In the other acquisition, the volunteer moved his head throughout the duration of the scan, yielding the motion datasets.The motion was approximately a periodic rotation of 20 • around the axis of the spine with a frequency of 1 Hz.
In the four additional volunteers (Studies 1-4), M0 datasets were acquired without motion to serve as the reference data.M0 and M2 datasets were also acquired during continuous head motion consisting of 5 • to 20 • periodic rotation around the axis of the spine with a frequency of approximately 0.1 to 1 Hz.
Each dataset was corrected for motion and eddy current artifacts with eddy 13 from FMRIB Software Library (FSL) 25 (M0, M1, M2 datasets).In addition, the eddy package provides specialized options to correct for motion-associated signal dropout with outlier replacement 14 and slice-to-volume misregistration. 26Because these are widely available options for dealing with extreme motion, we also generated datasets with these additional corrections (M0+, M1+, M2+), hereafter referred to as datasets with extreme motion correction.For each dataset, fractional anisotropy (FA), longitudinal diffusivity (LD), mean diffusivity (MD), and transverse diffusivity (TD) maps were calculated using the dtifit package of FSL.The maps were then coregistered to a common space using Advanced Normalization Tools. 27I map analysis was performed in python.For whole brain analysis, five superior and inferior edge slices were discarded because of registration errors resulting from extreme head motion.In all comparative analyses, the parameter maps derived from the no-motion-M0 data were used as references.For each dataset, density plots of the per voxel values of DTI-derived indices and per voxel differences between DTI-derived indices calculated from the motion and reference data over the whole brain were computed with Seaborn 28 and Matplotlib. 29

RESULTS
Sample DW images from Study 0, acquired during head motion using M0, M1, and M2 diffusion gradients, with and without retrospective motion correction for extreme motion, are shown in Figure 1B-G.Reference images acquired without deliberate motion using traditional M0 diffusion gradients are shown in Figure 1A.Results of visual signal dropout assessment in all DW images across the whole brain for all volunteers are summarized in Table 1.
In Figure 1B

T A B L E 1
Number (%) of diffusion-weighted images from the whole brain (1500 total images from the central 50 of 60 total slices and 30 diffusion directions) corrupted by continuous gross head motion for each volunteer.

No. (%) of images corrupted by head motion M0 M0+ M1 M1+ M2 M2+
Study 0 662 (44) 671 ( 45) 109 ( 7 of the motion-M0/M0+ images were corrupted by signal dropout.In Figure 1D, some signal dropout is also present in the motion-M1 DW images, although it is less frequent and severe than that observed in the motion-M0/M0+ images.Applying retrospective extreme motion correction to the motion-M1 images further reduced signal dropout from 7% to 1% in the motion-M1+ data (Figure 1E).As is represented in Figure 1F,G, no substantial signal dropout was observed in the motion-M2/M2+ DW images.These results were consistent across all volunteers, with no significant signal dropout observed in any of the images acquired with M2 sequences (see Table 1).
Figure 2 shows representative directionally encoded color fractional anisotropy (DEC-FA) maps from Study 0, calculated with motion and no-motion M0, M1, and M2 datasets, with and without retrospective correction for extreme motion.Voxelwise absolute differences between motion and reference maps are also shown.A similar analysis of the Study 0 LD, MD, and TD maps can be found in the Figures S2-S4.
Visual assessment of the no-motion DEC-FA maps in Figure 2 indicates consistency among the M0, M1, and M2 data in the absence of motion.In contrast, the motion-M0/M0+ DEC-FA maps are largely corrupted by head motion, with both maps visibly overestimating FA in the white and gray matter compared to the reference.The motion-M1/M1+ maps demonstrate markedly improved consistency with the reference.It is noted that the retrospective corrections used to yield the motion-M1+ data improved the fidelity of the resulting DEC-FA map, with reduced blurring and overestimation of FA throughout the white and gray matter.The motion-M2/M2+ DEC-FA maps are the most consistent with the reference, demonstrating the robustness of the M2 diffusion gradients to motion during data acquisition.
Representative DEC-FA maps for all five volunteers and reference no-motion-M0, motion-M0/M0+, and motion-M2 data are shown in Figure 3. Difference maps with respect to the reference data are also presented for all motion datasets.Similar results for the LD, MD, and TD maps acquired in Studies 1 to 4 can be found in Figure S5.

F I G U R E 2
Directionally encoded color fractional anisotropy (DEC-FA) maps from a representative slice in the corpus callosum region, calculated using DTI data acquired without and with intentional motion, and the differences between motion and reference no-motion-M0 maps.Results are presented for Study 0 acquisitions using M0, M1, and M2 motion compensating gradient schemes, both without (left) and with (right) added retrospective correction for extreme motion.

F I G U R E 3
Directionally encoded color fractional anisotropy (DEC-FA) maps from all five volunteers using data acquired with M0 diffusion gradients without head motion, M0 diffusion gradients during head motion, and M2 diffusion gradients during head motion.Results are presented for data processed with (M0+) and without (M0) retrospective corrections for extreme motion for data acquired with M0 diffusion gradients.Difference maps (ΔFA) taken with respect to the reference no-motion-M0 data are presented for the motion M0, M0+, and M2 data.Note that some data from Figure 2 (Study 0) is repeated for ease of comparison between the five studies.
Although four of the five motion-M0 DEC-FA maps are noticeably corrupted by head motion compared to the reference maps, all motion-M2 maps are consistent with the reference data.In cases where <15% of DW images acquired with M0 gradients are corrupted by head motion (Studies 1, 3, and 4), retrospective corrections for extreme motion were able to improve the quality of the motion-M0+ DEC-FA maps, but in the cases where >15% of the motion-M0 DW images were corrupted by motion (Studies 0 and 2), the retrospective corrections for extreme motion were unable to improve the quality of the resulting motion-M0+ DEC-FA maps.In all cases, the motion-M2 data is most consistent with the reference, as is evident in the difference maps.
Density plots of the FA distributions and voxelwise error in FA over the whole brain for each acquisition and all volunteers are shown in Figure 4.The results of Study 0 are shown in Figure 4A-C, and the results of Studies 1 to 4 are shown in Figure 4D.Similar analysis for LD, MD, and TD can be found in the Figures S2-S5.Study 0 DTI parameter analysis from a region of interest in the corpus callosum is presented in Table S1.
From Figure 4A, parameter distributions derived from Study 0 no-motion-M1/M2 datasets demonstrate high levels of agreement with data acquired using standard M0 diffusion gradients in the absence of motion.
In Figure 4B, Study 0 FA distributions derived from motion-M0/M0+ datasets (blue) are significantly right-shifted compared to the reference (dashed black).In comparison, the motion-M1/M1+ distributions (green) have significantly improved consistency with the reference and the motion-M2/M2+ distributions (orange) are the most consistent.
Figure 4C presents distributions of the per voxel differences in FA maps calculated from Study 0 data acquired during motion compared to reference maps.Both without (left) and with (right) retrospective motion correction for extreme motion, the motion-M0 datasets resulted in the largest bias and variance in the voxelwise difference in FA.Although both the motion-M1 and motion-M2 difference distributions yielded low bias, the motion-M1+ datasets generally achieved the lowest bias, and the motion-M2 datasets yielded the lowest variance.
A comparison of the standard (dashed black), motion-M0 (blue), motion-M0+ (orange), and motion-M2 (green) FA maps from Studies 1 to 4 can be found in Figure 4D.Whole brain FA (left) and voxelwise FA error (right) distributions are consistent with the results in Figures 3 and 4A-C.The degree of corruption of the motion-M0/M0+ data varies between different studies, with very significant variance and bias observed in Study 2 and minimal variance and bias observed in Study 3. In all cases, however, it is observed that the motion-M2 FA maps are consistent with the reference data, with FA distributions that align with the reference (left) and narrow difference distributions that are sharply peaked around zero (right).

DISCUSSION
In this work, motion compensating diffusion gradients were proposed to mitigate artifacts due to large degrees of motion.Experiments in healthy volunteers demonstrated considerably improved consistency in DTI parameters from data obtained during head motion when M1 and M2 diffusion gradients were used in place of traditional M0 diffusion gradients.
6][17][18] In contrast, the proposed use of motion compensating diffusion gradients during data acquisition inherently reduces the occurrence of motion corruption in the acquired images, even in the presence of continuous, large-scale head motion.In the present study, M1 gradients reduced and M2 gradients completely eliminated significant signal drop out in DW images acquired during head motion, allowing for quantification of DTI parameters despite the motion.
M1 and M2 diffusion gradient sequences were directly compared to FSL eddy's corrections for outliers and slice-to-volume misregistration, which are widely used to address significant subject motion. 10,13,14Although FSL eddy with extreme motion correction worked well when DW image corruption was moderate (e.g., Study 0 motion-M1+ and Study 4 motion-M0+) it was unable to correct for severe dropout (e.g., Study 0 and Study 2 motion-M0+).This is consistent with the work of Andersson et al., 14 in which the FSL framework was reliable only when no more than 10% of the DW images were affected by significant signal dropout.As a result, the motion-M1/M2 datasets dependably resulted in more consistent estimation of DTI parameters than the M0+ datasets (Figures 2-4 and S2-S5).
From an examination of the data acquired during head motion in Study 0, M2 gradients were observed to eliminate signal dropout in the DW images, whereas some signal dropout remained in the images acquired with M1 gradients.The motion-M1 results presented in Figures 2,  4, S2-S4, and Table S1 show elevated quantification of DTI parameters when additional retrospective motion correction was not used.In contrast, the motion-M2 DTI parameters generally had less bias and variance compared to the reference parameters without the need for additional retrospective corrections.
Motivated by the results of Study 0, Studies 1-4 focused on the use of M0 and M2 diffusion gradients, with and without retrospective extreme motion correction.Across the five volunteers scanned in this work, the motion-M2 DW images consistently exhibited no significant signal dropout and the resulting DTI parameter maps were found to have superior agreement with the reference data compared to M0+ data.In contrast, the degree of corruption in the motion-M0 images and DTI parameter maps was observed to depend on the details of the motion exhibited by the individual volunteers (see Table 1, Figures 2 and 3).To understand this intersubject variability, motion parameters and outliers were studied, but were found to be unreliable because of the extreme corruption of the motion-M0/M0+ data, (see Figure S6).Further investigation of motion details in relationship to the corruption of M0 DTI data, although outside the scope of this study, is warranted in future work.Despite the varying quality of DTI results obtained using traditional methods, when a subject is prone to head motion this work finds that it is possible to obtain quality DTI parameter maps with motion compensated diffusion gradients.
One cost of using motion compensated diffusion gradients is the resulting increase in TE and total scan time compared to traditional sequences, which also increases opportunities for motion.Although TE and TR were matched at TE/TR = 89/12 600 ms for all sequences in this study, the minimum possible echo and repetition times on our system were TE/TR = 47/5900 ms, 76/10 400, and 88/11 400 ms, yielding minimum acquisition times of 3 min 38 s, 6 min 25 s, and 7 min 2 s, for M0, M1, and M2, respectively.][32] Although DTI parameters from no-motion M0, M1, and M2 data, in this work, were generally consistent and data obtained with M1 and M2 gradients during head motion was significantly improved compared to the motion-M0/M0+ data, some bias was observed in no-motion-M1/M2 data relative to no-motion-M0 data.In a region of interest in the body of the corpus callosum for Study 0, the percent error in the mean for FA, LD, MD, and TD was 8% or less for no-motion-M1 and 12% or less for no-motion-M2, compared to 6% or less in rescan data acquired with the standard no-motion-M0 sequence (see Table S1).Diffusion time has been shown to affect DTI measures and could contribute to the observed bias. 335][36] A detailed study on the effects of M1 and M2 gradients on perfusion signals is warranted in future work, potentially investigating the use of low, but non-zero b-value reference images as has been proposed in cardiac DTI. 37,38

Study limitations
One limitation of the present study arises from its focus on gross head motion.Slight signal loss because of physiological motion (e.g., cardiac pulsatility, respiratory motion), can be easily missed, but might cause biases in quantitative analyses that can lead to false inferences.Future work should investigate the ability of motion compensating diffusion gradients to eliminate subtle artifacts associated with physiological motion.Another limitation of this work is the controlled range and frequency of head motions tested.Throughout the motion scans, volunteers moved their heads with consistent rhythms and in regular patterns.Future work testing a larger variety of head motions is required to ensure the continued robustness of motion compensated diffusion gradients to motion.
This work was additionally limited in its participant population.Future studies should be designed to test the large-scale application of motion compensating diffusion gradients to DTI for a range of patient populations that are susceptible to head motion, including children and adults with neurological conditions such as Parkinson's and stroke.
Finally, it will be important to perform studies in future work that test the generalizability of the results presented here.Although this work primarily focused on the use of M2 diffusion gradients because of their ability to eliminate signal dropout without additional retrospective corrections for extreme motion, additional work is necessary to provide guidelines regarding the optimal degree of diffusion gradient motion compensation (i.e., M1, M2), possibly in combination with additional retrospective motion corrections, for given populations.In such studies, the benefit of reduced signal dropout gained with increasing gradient motion compensation will need to be weighed against increased TE and scan time, as well as any potential bias in quantified DTI parameters.

CONCLUSION
This study has shown that quality diffusion tensor images can be acquired in the presence of large degrees of head motion with motion compensating diffusion gradient schemes.Unlike most motion correction techniques, which rely on the limitation of head motion duration and scope to be effective, diffusion encoding using first or second order motion compensation can reduce and even eliminate the incidence of signal dropout in DW images acquired during continuous head motion.This technique can be applied on a conventional clinical MR scanner using only modifications to the pulse sequence.As a result, the use of motion compensating diffusion gradients has the potential to make DTI routine to a larger patient population.S1.In Study 0, seven datasets were collected in a single session: Reference no-motion-M0, Rescan no-motion-M0, no-motion-M1, no-motion-M2, motion-M0, motion-M1, and motion-M2.The rescan no-motion-M0 data was collected at the end of the scanning session to assess variability in the reference scan.
The subject was brought out of the MRI and allowed to get up off the table for approximately five minutes before re-entering the MRI to collect the rescan data.
As described in the Methods, all data was corrected for motion and eddy current artifacts with eddy from FSL. DTI parameters (FA, LD, MD, TD) were calculated using dtifit from FSL and the resulting maps were coregistered with the reference data using ANTs.For comparison to a popular retrospective motion correction technique, FSL eddy's outlier replacement and slice-to-volume misregistration corrections were additionally applied, yielding the motion-M0+, motion-M1+, and motion-M2+ datasets.(Figure 3), an investigation of the per-subject motion was performed using the RMS movement and outlier report generated by FSL Eddy.RMS motion parameters for the motion-M0+, and motion-M2+ data for all subjects are shown in (A).A count of the number of outliers detected by FSL Eddy compared to the number of images identified as having signal dropout through visual inspection for the motion-M0+ and motion-M2+ data for each subject are reported in (B).
How to cite this article: Kara ,C, severe signal dropout because of head motion is evident in the motion-M0/M0+ DW images.Based on visual inspection of the Study 0 data, >40% F I G U R E 1 Diffusion-weighted images (b = 1000 s/mm 2 ) from Study 0 showing a representative slice and 10 representative diffusion directions.(A) Reference images acquired with traditional M0 diffusion gradients, no deliberate head motion, and no retrospective corrections for extreme motion.(B-G) Diffusion weighted images acquired with (B,C) M0, (D,E) M1, and (F,G) M2 motion compensated diffusion gradients, acquired during continuous gross head motion without (B,D,F) and with (C,E,G) retrospective corrections for extreme motion.

F I G U R E 4
Distributions of fractional anisotropy (FA) in the whole brain, calculated with Study 0 DTI data acquired (A) without intentional motion and (B) with intentional motion, both without (left) and with (right) retrospective correction for extreme motion.In the distributions for data acquired with motion, the reference distribution calculated from no-motion-M0 data is shown in dashed black.(C) Distributions of the per voxel differences in FA maps calculated with motion data and the reference (no-motion-M0) data.(D) FA distributions (left) and errors (right) calculated with respect to the reference data for Studies 1-4.

Figure S5 .
(A-C) DTI parameter maps for a representative slice in the region of the corpus callosum for Studies 1-4, showing (A) LD (×10 −3 mm 2 /s), (B) MD (×10 −3 mm 2 /s), and (C) TD (×10 −3 mm 2 /s) calculated using reference data (no-motion-M0), motion-M0 data, motion-M0+ data, and motion-M2 data.Differences between the reference data and each of the other datasets are also shown.(D-F) DTI parameter distributions in the whole brain from studies 1-4 are shown for (D) LD, (E) MD, (F) TD. (left) Distributions of DTI parameters, calculated with motion-M0 data (blue), motion-M0+ data (orange) and motion-M2+ compared to the reference no-motion-M0 data (dashed black).(right) Distributions of per voxel differences in DTI parameter maps calculated with motion data and the reference data.

Figure S6 .
Motivated by the intersubject variability observed in the motion-M0 and motion-M0+ FA maps D, Koenig K, Lowe M, Nguyen CT, Sakaie K. Facilitating diffusion tensor imaging of the brain during continuous gross head motion with first and second order motion compensating diffusion gradients.Magn Reson Med.2024;91:1556-1566.doi: 10.1002/mrm.29924