A numerical human brain phantom for dynamic glucose‐enhanced (DGE) MRI: On the influence of head motion at 3T

Dynamic glucose‐enhanced (DGE) MRI relates to a group of exchange‐based MRI techniques where the uptake of glucose analogues is studied dynamically. However, motion artifacts can be mistaken for true DGE effects, while motion correction may alter true signal effects. The aim was to design a numerical human brain phantom to simulate a realistic DGE MRI protocol at 3T that can be used to assess the influence of head movement on the signal before and after retrospective motion correction.


INTRODUCTION
Characterization of the microvasculature and quantification of microcirculatory parameters such as cerebral perfusion and capillary permeability by MRI is important in several clinical applications. 1 One major application is brain cancer, a disease that may cause multiple disabilities and death. Imaging therefore plays a crucial role for diagnosis and therapeutic monitoring. 2,3 By intravenously injecting a contrast agent bolus and tracking it with dynamic T 2 /T 2 *-or T 1 -weighted MRI, quantitative and qualitative information on cerebral perfusion and/or blood-brain barrier permeability can be obtained. 4,5 However, these methods often use a gadolinium (Gd) -chelate contrast agent and, while these are safe for most patient populations, it has been shown that Gd accumulates in bone and tissue. [6][7][8][9] Some of these agents are also potentially toxic for patients with kidney disease, leading to nephrogenic systemic fibrosis. 10,11 Therefore, it is of interest to investigate non-Gd contrast agents. Dynamic glucose-enhanced (DGE) MRI employs chemical exchange saturation transfer (CEST) or exchange-enhanced relaxation to study tissue uptake of glucose analogues, which, for D-glucose, is determined by tissue perfusion, transport and metabolism. [12][13][14][15][16][17][18][19][20][21] DGE MRI explores the change in signal intensity of several sugar hydroxyl protons, located approximately between 0.6 and 3.0 ppm relative to the proton resonance of free water in the water saturation spectrum (Z-spectrum). The inherently slow (minute time scale) infusion and uptake of D-glucose requires long scan times of around 10-20 min. The calculation of difference images as a function of time is therefore susceptible to image artifacts related to head motion. In addition, B 0 changes due to motion and field drift are problematic if a single frequency selection is used for RF saturation. Voxel displacement due to motion can render the shim settings determined at the beginning of the scan invalid, 22 and heating of the iron plates in the room temperature shim trays due to eddy currents may cause field drift. 23,24 Other sources of signal change may be dilatation of the lateral ventricles caused by D-glucose uptake and ventricular pulsation related to, for example, the cardiac pulsation. 20,[25][26][27] A previous study investigated the effect of motion using difference images of Z-spectra without glucose infusion to create synthetic data before and after simulated motion, also taking B 0 -fluctuations into consideration. 28 Hypo-and hyperintense signal changes of the order of 1% were observed in the vicinity of tissue boundaries, causing pseudo-DGE effects of the same order of magnitude as the DGE signal changes. Another study concluded that standard motion correction algorithms reduced pseudo-DGE effects. 29 While such retrospective motion correction may be a necessary step in post-processing, interpolation and smoothing can erroneously alter the true DGE signal entangled with residual motion artifacts at tissue interfaces. 28 The aim of this study was to implement a realistic customizable framework for DGE MRI to simulate dynamic Z-spectra under the influence of motion and dynamic B 0 -changes. As a first application of this numerical human brain phantom, we simulated a DGE experiment that used simulated dynamic contrast responses for different tissue types as ground truth. Rigid body movement patterns and physiological volumetric changes of the lateral ventricles were included in the simulations and their influence on the DGE effects before and after retrospective motion correction were studied.

MRI data acquisition
Experimental data were acquired for deriving different realistic patterns of motion and to aid in modeling the anatomical basis for the numerical human head phantom. Six subjects (three male, three female; ages 39-74 y), including five patients (four gliomas, one meningioma) and one healthy volunteer, were scanned at a 3 T MAGNETOM Prisma with a 20-channel head coil system (Siemens Healthcare, Erlangen, Germany). The project was approved by the local ethics committee (The Regional Ethical Review Board in Lund, EPN 2017/673) and written informed consent was obtained from all participants. All participants fasted for 4-6 h before examination to stabilize the baseline blood glucose and insulin levels. D-glucose solution (50 ml, 50% dextrose) was administered intravenously in one arm for a duration of 4 min using a power injector. The glucose infusion was followed by a saline flush of 30 ml included in the total infusion duration.
Four of the patients (three gliomas, one meningioma) and the healthy volunteer were scanned with a prototype CEST sequence with a 3D gradient echo (GRE) readout consisting of 12 partitions (TR/TE/FA = 10 ms/3 ms/12 • ). 30 Saturation was accomplished using 10 Gaussian-shaped pulses (B 1 = 1.6 μT) of 100 ms duration and 10 ms interpulse delay at a single saturation offset of 2.0 ppm from the water resonance frequency. One non-saturated (S 0 ) image was acquired by saturating at 1500 ppm at the start of the CEST image acquisition. Morphologic images were acquired using a MPRAGE sequence with 1 mm 3 isotropic voxel size and examined by an experienced neuroradiologist (P.C.S.).

F I G U R E 1
Steps of the numerical head phantom framework illustrated in a chronologically ordered pipeline (A) Segmentation of high-resolution anatomical images results in different tissue components, that is, air, bone, fat, and intracranial tissue, in which veins (V), arteries (A), gray matter (GM), white matter (WM), CSF, and tumor (Gb) were later identified. Each voxel has the option to add dynamic contrast. Motion patterns representing a least severe pattern (LSP) and a most severe pattern (MSP) of movement in the subject group can be included followed by adding Rician noise, smoothing and downsampling to a voxel size of 2 × 2 × 3 mm 3 . The arrows indicate the order of the simulation steps. (B) Simulations steps over time for the ground truth without motion (left) and for the cases with motion (right) One volunteer and one patient were examined using a temporal resolution of 5 s, a voxel size of 3.3 × 3.3 × 4.0 mm 3 , an acquisition time of 599 s with a 180 s baseline. Two patients were scanned with the same temporal resolution and voxel size as previously mentioned, but with a total acquisition time of 798 s including a baseline of 120 s. One patient was examined using a temporal resolution of 9.4 s, a voxel size of 1.7 × 1.7 × 4.0 mm 3 and an acquisition time of 468 s including a 180 s baseline. Finally, one patient was scanned with a 3D turbo spin echo (TSE) CEST sequence, but these data were not included for the retrieval of motion patterns for the phantom. Instead, the MPRAGE images acquired from this patient were selected for anatomical modeling of the phantom.

Simulation pipeline
The phantom was based on an MPRAGE brain image segmentation, used to model dynamic Z-spectra with or without D-glucose infusion under the influence of B 0 -changes and head motion. An overview of the simulation pipeline is shown in Figure 1. An MPRAGE brain image from a patient with a glioblastoma was used to segment gray matter (GM), white matter (WM), cerebrospinal fluid (CSF), other soft tissue and bone using SPM12. 31 In addition, regions of interest (ROIs) were manually drawn to delineate soft tissue in veins (V), arteries (A), and tumor tissue (Gb for glioblastoma) as well as CSF in lateral ventricles (LVs).

2.2.1
Dynamic Z-spectra Using a custom script written in Python Z-spectra as a function of B 1 (0.6, 1.2 and 1.6 μT) were derived using the Bloch-McConnell equations for 10 Gaussian-shaped saturation pulses (total saturation time of t sat = 1 s) and a field strength of 3 T. It was assumed that five protons exchange with bulk water (three protons at 1.2 ppm and one proton each at 2.2 ppm and 2.8 ppm, respectively) and that magnetization transfer contrast (MTC) originated from the semisolid macromolecular protons. Six types of tissue were used in the simulations, that is arterial blood, venous blood, GM, WM, CSF, and tumor. The change in D-glucose concentration was assumed to be 10 mM in arterial and venous blood and 1 / 4 of this concentration change was assumed in GM and WM tissue. 32,33 These levels were added to the typical baseline concentration of 5 mM in vessels and 1 / 4 of that in GM and WM. To obtain the Z-spectra in CSF, we assumed that the D-glucose concentration in CSF was 60% of that in arterial blood. 34 For the Z-spectra in the tumor, we used a vessel volume of 8% since angiogenesis will result in a larger vessel volume than in normal tissue. 35 Together with the extravascular extracellular space (EES), which can account for as much as 25% of the tumor volume, one can assume that the D-glucose increase is 33% of 10 mM in the tumor. 36 In addition to the CEST effect, the presence of hydroxyl protons that are chemically shifted induces an exchange-related transverse relaxation component that can be described by the Swift-Connick equation, giving a transverse relaxivity r 2ex = 0.012 s −1 mM −1 . 18 Tumors show an increased lactate production (i.e., the Warburg effect) 37 leading to a lower pH in the EES and thereby a change in the exchange rates of the hydroxyl protons. For the tumor, we assumed an EES pH of 6.8 leading to a reduction in the exchange rates from k 1.2 ppm = 4000 to 3100 Hz, k 2.2 ppm = 8000 to 2500 Hz, k 2.8 ppm = 10 000 to 6300 Hz approaching the intermediate exchange regime. 38 Taking this pH into account, we also have the tumor D-glucose relaxivity increased to r 2ex = 0.014 s −1 mM −1 . Protons on the semi-solid macromolecules resulting in an MTC effect were assumed to have concentrations of 5500 and 15 400 mM for GM and WM, respectively. For the tumor, a concentration of 2750 mM was assumed because of its location between GM and WM and having a higher water content. Time-dependent glucoCEST uptake curves were simulated, based on the overall shape of DGE response curves previously observed at 3 T. 13,20,32,39 A dynamic time series with an acquisition duration of 690 s and a temporal resolution of 5 s was simulated. The time series consisted of S 0 (non-saturated signal) data acquisition (20 s), baseline (20-140 s), D-glucose infusion (140-380 s), signal increase and decay (380-565 s). The previously simulated Z-spectra as a function of D-glucose concentration were used to create dynamic Z-spectra under the influence of D-glucose infusion. The tissue specific dynamic Z-spectra were combined with the corresponding tissue segmentations, to obtain dynamic Z-spectral maps.

Motion parameters
Using Elastix, 40,41 rigid body motion estimates were derived from all five participants in the glucoCEST DGE study using the first saturated image S Sat as positional reference. Two patients were selected to represent two different motion patterns, that is, a least severe pattern (LSP) of movement and a most severe pattern (MSP) of movement, within this group of subjects ( Figure S1D, in Appendix S1). An arbitrary motion pattern was also created to investigate the effect of non-continuous motion ( Figure S1E, in Appendix S1). Thus, a total of six different motion patterns were generated.
Additional motions exists because LVs undergo an oscillating volume change of approximately 1% due to cardiac pulsation, while D-glucose loading leads to ventricular dilatation by about 2.4%. 20,[25][26][27] Therefore, non-rigid LV movement was simulated by adding an affine transformation with a scaling coefficient changing over time to the LV segmentation. The scaling coefficient was varied randomly to mimic undersampled pulsation patterns due to, for example, cardiac pulsation. This is superimposed on a linear increase from the subject specific start of infusion, simulating an increasing D-glucose uptake ( Figures 2E,F). Increases in LV volume were offset by reductions in WM volume and vice versa.

2.2.3
Dynamic B 0 -changes The previously described segmentations were used to create a model incorporating the volume magnetic susceptibility of air (χ ≈ 0.36 ppm) 42 in the paranasal cavities and head surroundings, brain tissue and fluids (χ ≈ -9.05 ppm) 43 surrounded by the skull (χ ≈ −11.56 ppm), 44 and fat/skin (χ ≈ −8.50 ppm). 45,46 Smaller inter-tissue susceptibility differences (ranging from about −0.05 to 0.25 ppm) 47 were neglected under the assumption that a partial volume effect of a few percent will not affect the frequency shift substantially. The tissue contributions to the magnetic field were then numerically estimated by convolution with a magnetic dipole kernel, 48,49 after expanding the FOV by a factor of two (zero-padding) in each dimension to increase computational accuracy. 50 The effect of a static field in form of zero-and first-order static shims was obtained by using the first volume in the series of the simulated B 0 -field, and then adding these shim fields to retrieve a final B 0 map. In this case of motion, the rigid body motion parameters were then applied to the susceptibility model with the Matlab (The MathWorks, Natick, MA, USA) function imwarp using a rigid body 3D transformation and cubic interpolation, creating a time series with reassigned fields in each voxel for each time point. At each time point, the original shims were added to produce the final motion-corrupted B 0 -field time series. Finally, to return to the initial frame of reference for Z-spectral analysis, the inverse of the previously applied rigid body motion was applied. Subtracting the map for the first time point scan (baseline) from the individual maps for each time point in the time series then provided the field change per time point, thus providing insight into possible B 0 -changes that shift the Z-spectrum due to motion. Figure 1 illustrates how the effects of the dynamic B 0 -changes were applied to the dynamic Z-spectra for each voxel leading to a shift along the Z-spectral frequency axis. An offset of 2 ppm and a B 1 of 1.6 μT were selected to create a single offset time series of S Sat (t). Afterwards, the rigid body motion parameters and non-rigid LV movement were applied with the Matlab function imwarp using cubic interpolation, and the DGE image time series affected by movement was down-sampled spatially to 2 × 2 × 3 mm 3 .

Generation of DGE time series affected by motion
Smoothing using a Gaussian kernel was applied to obtain an image quality similar to in vivo conditions. Rician noise with a standard deviation of 0.05% was added to the magnitude signal. In the normalized signal (S/S 0 ), this corresponds to 0.10% standard deviation. As the effect size in DGE is about 0.50% in terms of normalized signal, 0.10% noise corresponds to a noise level of about 20% in the DGE signal change.

Numerical phantom experiments and analysis
The ground truth was set to be an experiment with D-glucose infusion, without motion, and thus no dynamic B 0 -field changes, but a static B 0 -field combined with a first-order shim, with a total duration of 565 s. Subsequently, experiments with and without infusion, and with six rigid head motion patterns (LSP, MSP, 2 × MSP, 2 × MSP only rotations, 2 × MSP only translations and arbitrary motion, summarized in Figure 2 and Figure S1A-E, in Appendix S1) including ventricular pulsation, and with and without ventricular expansion, were simulated while keeping the zero-and first-order static shim constant, that is, adding it to each time point after motion occurred ( Figure 1B).

Retrospective motion correction
Retrospective motion correction was applied to the simulations using Elastix. 40,41 The details of the parameter settings are provided in the Supporting Information document. The first S Sat image was assigned as positional reference.

DGE maps and statistics
The DGE signal S DGE (t) was calculated using the following equation where S base is the average baseline signal (over 24 time points in this study), S Sat (t) is the saturated signal at time t and S 0 is the average non-saturated signal (over four time-points in this study). A positive S DGE value can be interpreted as an increased D-glucose concentration. ROIs were drawn in tumor, WM, GM and CSF to extract DGE signal curves before and after rigid motion correction. Maps of the mean area under curve AUC mean were calculated using: where For WM and tumor tissue, the change in DGE AUC mean signal over the time frame 240-305 s compared to the time frame 0-240 s was evaluated using the two-sided Wilcoxon signed-rank test. This statistical test was performed for the motion patterns "ground truth", "LSP" and "MSP", with and without motion correction (for LSP and MSP), and with and without D-glucose infusion.

RESULTS
The experimentally measured dynamic time series for LSP (Figure 2A,C) and MSP ( Figure 2B,D) show an overall continuously increasing displacement over time with random fluctuations. Larger motion is seen from the start of the glucose infusion initiating a diverging motion pattern with superimposed sudden jumps. Especially the negative pitch in MSP ( Figure 2D) shows that the head falls back into the head coil, with a large magnitude. For the LV, the scale coefficients in the two examples in Figures 2E,F show a steady increase after the start of the glucose infusion, simulating a volume dilatation of up to approximately 2% in the LV. The curve is superimposed with a randomly varying constant interval simulating an oscillatory expansion and reduction of approximately 1% in the volume of the LV due to, for example, cardiac pulsation, but not limited to this. Oscillatory deformation of LVs is complex and cannot be fully captured with the low temporal resolution (5 s) in this phantom study. 27 The simulations illustrated in Figure 3 define the ground truth. Baseline Z-spectra for the tissues are shown at three different B 1 values in the top row of Figure 3A, while the differences between spectra before and after infusion are shown in the bottom row. The estimated DGE signals at an offset frequency of 2 ppm and a B 1 of 1.6 μT are S DGE,A = 1.11%, S DGE,V = 0.51%, S DGE,GM = 0.28%, S DGE,WM = 0.06%, S DGE,CSF = 0.90%, and S DGE,Gb = 0.38%. The simulated normalized signal curves for the different tissues are shown in Figure 3B (left) and the resulting DGE signal in 3B (right). About 1 min after the end of the infusion, the DGE signal reached a peak followed by a slow steady decrease. Anatomical reference images for an axial slice in the mid-section of the brain, including the MPRAGE image used for segmentation, are shown in Figure 3C. Gadolinium-enhanced MPRAGE, FLAIR, and segmented images are added for visual comparison. Applying the signal curves to each separate tissue segmentation created the DGE images in Figure 3D, visualized in form of AUC mean maps for 4 different time intervals. The AUC mean map series starts with the baseline, showing no signal change, followed by time intervals after D-glucose infusion showing the individual tissue responses. In particular, changes in CSF and tumor are clearly visible.
AUC mean maps for the ground truth and different cases without/with infusion and before/after motion correction are shown for the different motion patterns in Figures 4A, 5A, and Figures S4, S6, S8, S10, in Appendix S1. The ground truth shows a clear dynamic tissue response ( Figures 4A and 5A). For all motion patterns without infusion, added motion clearly affects the AUC mean -maps by introducing artifacts (pseudo-DGE contrast) seen as hyper-and hypo-intensities at tissue interfaces such as the tumor rim. Infusion tends to highlight especially CSF over other tissues, while a small signal increase is seen within the motion-induced signal pattern in the tumor rim. These changes are amplified with increasing motion (cf. Figure S4, in Appendix S1), showing increased pseudo-DGE effects, again at tissue interfaces such as the tumor rim, and especially some hyper-and hypointense signal changes in the cortex and within and around the ventricles. The AUC mean -maps show that motion correction reduced the overall appearance of pseudo-DGE effects but could also introduce new effects, especially in regions with many tissue interfaces such as GM and the lateral ventricles ( Figures 4A, 5A, and Figures S4, S6, S8, S10, in Appendix S1). For all motion patterns, the tumor rim enhancement was reduced when motion correction was applied and the tumor itself became visible after infusion, both with and without motion correction. The extracted dynamic response curves for the ROIs, both before and after motion correction, are displayed in Figures 4B, 5B, and Figures S5, S7, S9, S11, in Appendix S1. Without infusion, the DGE signal in the tumor showed a pattern ( Figure S9E1-E2, in Appendix S1) similar to the purely translational motion (cf. Figure S1D, in Appendix S1). For the motion pattern using only rotations ( Figures S7E1-E2, in Appendix S1), there were hardly any pseudo-DGE effects visible in the tumor. In particular, pseudo-DGE effects were not present in WM and the resulting DGE signal change in WM was within the noise fluctuations. However, GM and CSF showed clear pseudo-DGE effects for all motion scenarios, except in CSF when only rotations were applied (cf. Figures S7H1-H4, in Appendix S1). Adding infusion led to increased DGE signal, and motion correction was able to decrease the DGE signal in tumor and WM to levels closer to the ground truth. This can be seen, for example, in the MSP case ( Figure 5B), where in tumor tissue, the situation without infusion shows an erroneous large positive signal. After motion correction, this signal is reduced but still higher than in the ground truth.
The effect of only B 0 -changes on the DGE signal for LSP and MSP is illustrated in Figure 6. The AUC mean signal is shown in the transverse view for LSP (Figures 6A-C) and MSP (Figures 6D-F) without infusion and with infusion in the late post-infusion time interval (305-425 s). The dynamic B 0 -changes ( Figures 6A,D) induced a gradual spatial intensity change on the AUC mean DGE signal ( Figure 6B,E), that gave decreases as large as 0.5% in the  Figure 6G.
Box-and-whisker plots of the AUC mean signal of all voxels in the previously defined ROIs in tumor tissue and WM are shown in Figures 7 and 8, respectively. Three cases are shown for LSP and MSP: (1) only dynamic B 0 -changes without tissue mixing, (2) B 0 -changes and movement without motion correction and (3) B-changes and movement followed by motion correction, with and without infusion. Tumor tissue without infusion, influenced only by B 0 -changes, showed slight hyperintense signal for MSP (cf. Figure 7). Addition of the infusion led to a signal change higher than the ground truth. Motion led to hyperintensities, which were increased after motion correction in MSP. These effects were enhanced by infusion. Motion led to a wider quartile range and larger total range, with a larger variability for MSP. Motion correction led to a slightly wider quartile range, and the median signal after infusion was higher for MSP than for LSP and for ground truth. White matter influenced only by dynamic B 0 -changes showed hyperintensities for the MSP case without infusion, and adding infusion led to a larger signal intensity increase compared to the ground truth. Comparing the AUC mean DGE signal before and after motion correction showed no significant differences, but the signal trended to be larger than the ground truth due to B 0 -changes (Figure 8, bottom, Figure 5B).
The p-values from the statistical test for LSP and MSP are listed for tumor tissue (Table 1) and for WM ( Table 2). Both tissues showed significant DGE AUC mean signal change when infusion was added in all cases (ground truth, with/without motion). In tumor tissue, DGE AUC mean signal changes were non-significant without infusion for LSP before and after motion correction. For white matter, DGE AUC mean signal changes were non-significant without infusion before and after correction. For both tissue types, DGE AUC mean changes for MSP were significant without infusion and before and after motion correction.

DISCUSSION
In order to understand the influence of motion on DGE data, we developed a numerical human brain phantom. Availability of such a ground truth that takes only the DGE effect into consideration, without any motion, opens up possibilities for testing and validation of different motion patterns and correction and analysis approaches to account for these. The introduced DGE response curve shapes were based on experimental glucoCEST time series at 3 T and 7 T. 13,20,39 Experimental DGE response curves reported in the literature usually show high variability in curve shape between different scans and tissues, due to low contrast-to-noise ratio and artifacts mainly related to motion. In this study, response curves were simulated with peak signal based on the results of Bloch-McConnell simulations at 3 T, giving a sufficient approximation of experimentally measured DGE signal changes due to the gluco-CEST effect. 32 Inclusion of a full description of the kinetics for D-glucose uptake for the human brain would be required for a more realistic approach in future studies. 51 The availability of a numerical human brain phantom is likely to increase the understanding of the effect of motion during a DGE experiment, and an actual Bloch-simulator may enable addition of additional MR-related artifacts, for example, intra-volume motion-induced blurring. 52,53 On the other hand, bulk head motion, which could be simulated in the present approach, can be considered the main source of intensity artifacts, as shown, for example, in Figures 4 and 5.
Even though the number of participants was small, the LSP and MSP motion patterns (Figure 2) are likely to represent rigid motion occurring during a DGE experiment with a slow infusion duration of approximately 240 s. We interpret that the slowly varying characteristics

F I G U R E 7
Box-and-whisker plots showing the AUC mean DGE signal for tumor tissue (purple ROI in Figure 4C) under the influence of motion patterns LSP and MSP, with and without infusion. Top row: Effect of only dynamic B 0 -changes without tissue mixing (in the presence of a static first-order shim) on the DGE signal without motion correction (NoMoco) (cf. Figure 6B, C, E, F). Middle row: Effect of dynamic B 0 -changes (in the presence of a static first-order shim) with tissue mixing and without motion correction (NoMoco) (cf. Figures 4B, 5B). Bottom row: Dynamic B 0 -changes (in the presence of a static first-order shim) with tissue mixing and with motion correction (Moco) (cf. Figures 4B, 5B). The ground truth with infusion for tumor is shown separately beneath. The upper and lower borders of the blue box indicate the upper and lower quartile, respectively, the red line shows the median value, and the whiskers represent the maximum and minimum AUC mean values. The '+' denotes outliers, defined as values that are more than 1.5 times the interquartile range away from the bottom or top of the box originate from the involuntary relaxation of the subject's neck muscles, leading to a translation in Z-direction and a pitch rotation similar to nodding. [53][54][55] Ventricular volume variations are caused by cardiac pulsation and loading of D-glucose 20,25,26 These effects were simulated together in Figures 4 and 5 for cases with and without glucose infusion, leading to signal variations similar to those reported in vivo at 3 T (0.25%-1.5% in tumors). 20,28,29,32 Note that development of new pulse sequences 56 or use of other sugars such as 3-ortho-methylglucose (3OMG) 57,58 or D-glucosamine (D-GlcN) 59 is expected to increase these effects. The example motion patterns showed different amounts and types of movement, and different motion scenarios lead to individual pseudo-DGE patterns. Larger motion changes led generally to stronger pseudo-DGE effects, as demonstrated when comparing LSP, MSP and 2 x MSP in the Supporting Information. In addition, the resulting dynamic response curve depended on the pattern of the motion. This could be confirmed using the arbitrary motion pattern demonstrated in Figure S11, in Appendix S1. In this case, we did not observe the typical dynamic response obtained using the other motion patterns. This

F I G U R E 8
Box-and-whisker plots showing the AUC mean DGE signal for white matter (yellow ROI in Figure 4C) under the influence of motion patterns LSP and MSP, with and without infusion. Top row: Effect of only dynamic B 0 -changes without tissue mixing (in the presence of a static first-order shim) on the DGE signal without motion correction (NoMoco) (cf. Figure 6B, C, E, F). Middle row: Effect of dynamic B 0 -changes (in the presence of a static first-order shim) with tissue mixing and without motion correction (NoMoco) (cf. Figures 4B, 5B). Bottom row: Dynamic B 0 -changes (in the presence of a static first-order shim) with tissue mixing and with motion correction (Moco) (cf. Figures 4B, 5B). The ground truth with infusion for white matter is shown separately beneath. The upper and lower borders of the blue box indicate the upper and lower quartile, respectively, the red line shows the median value, and the whiskers represent the maximum and minimum AUC mean values. The '+' denotes outliers, defined as values that are more than 1.5 times the interquartile range away from the bottom or top of the box observation is similar to what has been described previously in the fMRI literature, that is motion may mimic activated voxels when the motion pattern is correlated with the stimulus. 60,61 It has previously been reported that B 0 -changes due to motion can affect the detection of DGE signal changes. 20,28 The reported variations are around 0.03 ppm for 1 mm translations and 1 degree rotations at 3 T. These effects could be reproduced, inducing global signal variations across tissue boundaries. The images in Figure 6, and Figures S2, S4, S6, S8, S10, in Appendix S1 demonstrate that the amount of motion and the brain location govern where erroneous signal occurs. Dominant positive field alterations could be found in the anterior and posterior parts, and negative field changes were observed in the mid-section of the brain, corresponding to the location of some air-filled cavities and the particular motion pattern. We previously recommended using an offset frequency at 2 ppm and a B 1 of 1.6 μT at 3 T. 32 Saturation closer to water (as commonly done at 1.2 ppm) has a higher T A B L E 1 p-values for AUC mean DGE signal change between the time intervals 0-240 s and 240-305 s in tumor tissue (for ROI colored in purple in Figure 4C) glucoCEST effect, even for a B 1 of 0.6 μT, as expected based on the merged water and D-glucose resonances in this intermediate to fast exchange regime. However, it will also introduce more pseudo-DGE signal changes, especially those caused by dynamic B 0 -changes, because the signal slope in the Z-spectrum is steeper when closer to the direct water saturation ( Figure 3A). One approach to correct for B 0 effects may be dynamic shimming, 62,63 which would be technically simple to add to the phantom. However, this approach is complex in practice in vivo, where dynamic shimming of each time point may not help because shimming is global. As a consequence, it is unclear if the field in the voxel after the new shimming will be the same as before the motion (i.e., shimming is just improved summed over the whole brain unless one would do slice-by-slice shimming) and the Z-spectrum can therefore still be shifted. In addition, motion correction would then be applied to the brain following the shim correction, which may again change the average field shift in many of the voxels. The easiest way for field correction would be acquisition of the Z-spectrum as a function of time, which would allow the field to be corrected by simply shifting the spectrum at each time point, which is our goal for DGE in the future. DGE contrast was slightly improved (closer to the ground truth) after motion correction. However, residual motion artifacts, especially at the gray matter/CSF tissue interfaces, were still prominent both with and without glucose infusion. The correction algorithm seemed to introduce both hypo-and hyperintensities in the LVs due to their non-rigid motion characteristics. The signal in the tumor tissue was small and strongly affected by motion, but the ring enhancement at the tissue interface (tumor border) could be reduced by motion correction. Figure 7 (top row) shows that a small additional hyperintensity was introduced by the B 0 shift for the MSP case. Without motion correction (Figure 7, middle row), tissue mixing leads to additional increases or decreases in signal, spreading the range of pseudo-DGE effects. The different motion patterns also confirmed that pseudo-DGE effects can originate both from motion (tissue mixing) and by applying motion correction. These effects originate from regions where tissue interfaces exist, and a large contrast is seen. This is similar to what has been observed in fMRI experiments, where motion correction can introduce false activiations. 60,61 However, motion correction will have no effect on the signal changes introduced by B 0 shifts. In our study, the simulated tumor was homogenous, while, in reality, a tumor is heterogeneous, which likely would lead to more pseudo-CEST effect due to the presence of more tissue interfaces.
Extracted response curves in tumor tissue, corrupted by motion, generally showed strong signal variations, and the pseudo-DGE effects using LSP and the different MSP movement patterns varied in the range up to 1% signal change, in agreement with a previous study. 28 The signal variation depended on how heterogeneous the structure was, and WM contrast was thus less affected. Furthermore, gradual small local field changes caused by B 0 -changes due to motion cannot be corrected when a single frequency point is acquired. In addition, static B 0 inhomogeneities initially reduced by shimming will affect the overall contrast before and after motion. The negative pitch movement used in the different MSP movements (MSP, 2 × MSP, 2 × MSP only rotations), which can be interpreted as leaning backwards in the head coil, is often seen in practice due to relaxation of the neck muscles over long scanning durations. This situation leads to movement of the air-filled nasal sinuses towards the transversal slice of interest. Ultimately, fast acquisition of full Z-spectra is required to improve this situation. Contaminated dynamic contrast after retrospective motion correction is known from other fields of varying contrast MRI. We attribute these effects to smoothing and interpolation errors, and this problem has been reported in, for example, quantitative T 1 mapping in a previous study. 64 Aimed to be used for studying different scenarios of motion, the proposed phantom can be applied to study different regions of the brain individually. The modular simulation design enables the use of Z-spectra and curve shapes that reflect different tissues and effect sizes (e.g., due to different concentrations of infusions). DGE MRI using D-glucose is only one of several dynamic contrast modalities that can be simulated using this simulation framework. Modalities such as dynamic susceptibility contrast MRI, 4,65 dynamic contrast enhanced MRI 5,66 and dynamic CEST imaging using other sugars, such as 3OMG, 57,58 D-GlcN, 59 and dextran 67,68 that have different uptake curves, can also be assessed using this phantom.

CONCLUSIONS
A numerical human brain phantom simulating a DGE experiment in combination with rigid-head movement and dynamic LV dilatation was developed. Motion-induced pseudo-DGE effects were reduced in some areas and increased in other areas by rigid motion correction. This was especially observed at tissue interfaces, where the induced pseudo-CEST effects were tangled with true CEST contrast complicating image interpretation. B 0 -shifts due to dynamic field changes as well as added shim settings for different positions after motion affected the spectral intensities. These effects cannot be corrected by rigid motion correction in DGE without acquiring large parts of the Z-spectra including the water proton frequency. Ultimately, advanced post-processing methods using retrospective motion correction, and perhaps dynamic B 0 -measurements and motion corrections using dual navigators, 69 are needed when implementing the DGE technique into the clinical practice. The phantom can be used for testing these approaches for dynamic CEST experiments, as well as for other types of dynamic scans affected by motion, for example, functional MRI.