Single voxel spectroscopy (SVS) relies on a homogeneous B0, a consistent frequency, and assumes that the localization remains valid for the duration of the scan. For a restless subject, who is unable to maintain a consistent pose during the scan, these do not hold true. We present a method that provides real-time (once every pulse repetition time [TR]) B0 and frequency measurements in addition to real-time correction of the volume of interest (VOI) position.
Current motion and artifact correction methods in magnetic resonance spectroscopy can be divided into two categories: phase and frequency adjustments and localization correction. Phase and frequency adjustments refer to a group of techniques that measure the signal phase and frequency by using either the residual water signal (1–4) or a secondary navigator (5–7). These methods correct both a velocity-induced phase error and frequency changes that result from either scanner drift or pose change. Phase and frequency adjustments can be applied both retrospectively and prospectively, but only prospective methods are able to correct the change in water saturation frequency.
Localization correction techniques in magnetic resonance spectroscopy have been demonstrated using an optical tracking system (7) and an imaging navigator technique called PROspective MOtion correction (PROMO) (8). The technique presented by Zaitsev et al. (7) provides both frequency and localization correction by combining optical tracking with navigator based frequency correction in addition to reacquisition of free induction decays (FIDs) with velocity induced phase errors. The disadvantages of an optical device are that they require: additional hardware, a marker to be rigidly affixed to the head, a clear line of sight between camera and marker, and the calibration of a camera to scanner transform.
There are several navigator-based motion tracking methods available, which take advantage of the k-space properties of rigid body transforms to subsample k-space in a time efficient manner. These include orbital (9), spherical (10), and cloverleaf (11) navigators. Although these techniques can be particularly fast, an imaging navigator is better suited to magnetic resonance spectroscopy because of its long repetition times (on the order of 1.5–3 sec) and lack of anatomical information. One such navigator is PROMO (12), which uses a set of three perpendicular, single slice, low resolution spiral images to register the head position to a reference map. This was demonstrated in SVS by Keating et al. (8).
In this work, the effect of changing head pose on zero-, first- and second-order B0 homogeneity was investigated in four different VOIs for a single volunteer. The use of an echo planar imaging (EPI) volume navigator (vNav) to correct in real time for both VOI position and zero- to first-order B0 inhomogeneity changes is demonstrated in six healthy volunteers. Finally, we demonstrate that this navigator minimally affects the metabolite signal and maintains spectral quality, when a subject moves during the scan.
The relationship between linewidth and B0 inhomogeneity can be expressed in terms of first- and second-order B0 changes. The signal from a single substance (n) can be described by Eq. 1 (13).
where an is the relative amplitude of the signal, T2n* depicts the inherent linewidth, and for a voxel with B0 inhomogeneity expressed as the magnitude of first- (g to g) and second-order (g to g) shim correction terms. Figure 1 demonstrates the theoretical effect of B0 inhomogeneity on linewidth based on Eq. 1. Figure 1a shows the change in linewidth as a function of a first-order B0 gradient, for which the magnitude ranges from 0 to 20 μT/m, for three different metabolite linewidths (T2n* = 40, 80, and 160 msec) and a voxel size of (2 cm)3. Figure 1b demonstrates how first-order B0 inhomogeneity affects different voxel sizes for a metabolite linewidth of T2n* = 80 msec. The effect of a second-order inhomogeneity is more complicated as the linewidth and signal amplitude are not proportional to one another. Figure 1c plots the linewidth as a function of the magnitude of the five second-order shim currents, whereas Fig. 1d plots the spectral amplitude for the same relative to its amplitude in a homogeneous VOI.
A field map can be used to optimize the shim currents of the scanner (14). A field map is generated by the complex division of two images with differing echo times (TEs). The difference in TE is typically chosen such that fat and water are in phase. This occurs for a TE difference of roughly 2.2–2.5 msec for a gradient echo at 3T. The best fit of the shim gradients to the spatially varying B0 field can be determined by minimum square error regression over the VOI, taking care to exclude voxels that do not have adequate signal-to-noise ratio. Reese et al. (14) suggests that a least square error cost function in the regression is sufficient for this application. Finally, image distortions resulting from the B0 field variations can be corrected using the known frequency offset of each voxel.
MATERIALS AND METHODS
All scans were performed on a Siemens Allegra 3T (Siemens Healthcare, Erlangen, Germany) in Cape Town, South Africa, according to protocols that had been approved by the Faculty of Health Science Research Ethics Committee of the University of Cape Town.
Investigation of the Effect of Motion on B0
To demonstrate the change in B0 because of pose variations, a single volunteer was scanned. Twelve high- resolution field maps were acquired with the head in different positions. The volunteer moved his head incrementally, first about the X-axis (chin-down to chin-up) and then about the Z-axis (rotate left to right). Six field maps were acquired during the X-axis rotation, ranging from 7.2° to −14.4°, and a further six field maps for the Z-axis rotation, ranging from −19° to 16°. Resultant rotations were assessed offline. The subject was trained before scanning as to how much to move his head.
A gradient echo sequence was used for the field map acquisitions with the following parameters, 48 slices, matrix 64 × 64, field of view (FOV) = 192 mm, slice thickness = 3 mm, TR = 502 msec, TE1 = 4.59 msec, TE2 = 7.05 msec, bandwidth = 260 Hz/pixel, and a slice separation of 0.6 mm. No shim adjustment was performed before each field map.
Each field map was registered to a reference three-dimensional multiecho magnetization prepared rapid gradient echo (15) using SPM5 (16) and resliced to match the three-dimensional multiecho magnetization prepared rapid gradient echo resolution of 1.0 × 1.3 × 1.0 mm3 using linear interpolation. This process facilitated the extraction of a chosen anatomical VOI based on the multiecho magnetization prepared rapid gradient echo. Four (2 cm)3 VOIs were selected; one in medial frontal gray matter anterior to the corpus callosum, one in right frontal white matter, another in right central white matter, and finally, one in the right inferior occipital brain region above the cerebellum. These VOIs are depicted in Fig. 2. For each VOI and head orientation the zero-, first-, and second-order B0 inhomogeneity in this VOI was calculated by transforming the voxel coordinates into the scanner frame of reference.
Using the mean frequency, linear B0 gradients and second-order terms, we investigated the effect of head pose on B0 in the four VOIs in our volunteer. The mean frequency (zero-order shim term) was calculated without fitting the first-order and second-order shim terms, and the first-order shim estimates were calculated without fitting the second-order terms.
The EPI vNav
To measure head pose and B0 inhomogeneity in real time, we implemented a three-dimensional multishot EPI vNav with a resolution of 8 × 8 × 8 mm3, an acquisition matrix of 32 × 32 × 28, and 256 × 256 × 224 mm3 FOV, so as to completely cover the FOV of the Siemens 3T Allegra scanner used in this study. Two contrasts were acquired with interleaved partition acquisitions, TE1 = 6.6 msec and TE2 = 9.0 msec, TR = 16 msec, and bandwidth 3906 Hz/pixel. The two contrasts were acquired interleaved in 58 shots, each with 2° flip angle. The first two shots collect a navigator used in N/2 ghost reduction for each contrast and the remaining 56 acquire 28 partition encodes (k-space slices), interleaved, for each contrast giving a total navigator duration of 928 msec.
The navigator sequence is highly customizable on the scanner console allowing for navigators to be tailored to a subject, sequence, and VOI. For example, the number of partitions could be reduced to 16 and still cover the full brain (128 mm), reducing scan time to only 544 msec. However, this would require the operator to position the navigator to ensure that it overlaps with the brain. As the SVS sequence used in this study has a sufficiently long magnetization (M0) recovery period, a navigator covering the complete FOV was chosen.
The vNavs are reconstructed immediately online to create a field map and two magnitude volume images. Pose estimation is performed using a single vNav contrast (TE1) by coregistering subsequent vNavs to the first vNav after the preparation (“dummy”) TRs. This registration is performed using an optimized Prospective Acquisition CorrEction (PACE) (17) algorithm that is an established method for registering whole-head EPI. The image reconstruction and PACE registration is performed online immediately after the navigator block in <80 msec.
Field map phase unwrapping is performed online using Phase Region Expanding Labeller for Unwrapping Discrete Estimates (PRELUDE) (18) with a mask created by including all voxels with a magnitude greater than max(|all voxels|)/15 that form part of the largest connected region of such voxels. This threshold was chosen as it ensured the inclusion of all voxels with sufficient signal-to-noise ratio. The connected regions are found using routines in the PRELUDE package, with the largest such region selected as the mask. Two frequency and first-order shim estimates are calculated online, one for the selected SVS VOI and one for the navigator FOV. The shim estimate for the navigator FOV is calculated using an unweighted least squares regression, whereas the shim estimate for the chosen VOI uses a weighted least squares regression, where the weighting of each navigator voxel is according to its intersection with the SVS VOI. The final two adjustments performed during shim estimation are to correct for B0 distortion of each voxel (14) and to shift the VOI position according to the motion estimate for the current TR thus ensuring that the voxel coordinates are mapped to the scanner coordinates taking into account the current pose. This ensures that SVS is acquired from the correct anatomical region with optimal shim setting in that VOI. Hence, shim estimation can only be performed after completion of prospective acquisition correction.
The complete online block, including transmission of the current estimates back to the sequence, occurs in under 170 msec, enabling the sequence to update the spectroscopy VOI according to current pose and apply the appropriate shim estimate to that VOI within the same TR. Figure 3 summarizes the flow and operation of the vNav block.
Insertion into SVS PRESS Sequence
The navigator block was inserted before water suppression in a SVS point resolved spectroscopy (PRESS) (19) sequence, occupying a portion of the TR used for M0 relaxation. The timing is illustrated in Fig. 4. As the navigator has a flip angle of 2°, we hypothesized that it would minimally affect the M0 relaxation process. This was explored in the in vivo experiments described below. The navigator real-time shim estimates were applied from the first preparation or dummy TR, whereas the pose estimates were calculated and applied from the second TR after preparation to allow the vNav shim time to stabilize. These estimates were applied synchronously immediately following the vNav block and before the water suppression.
In Vivo Validation
Six SVS PRESS scans were acquired with different protocols for each of six healthy volunteers. The aim was: (i) to investigate the impact of the navigator on the M0 relaxation process, (ii) to compare the navigated real-time shim to that of a manually optimized shim, (iii) to investigate the impact of shim- and motion-correction in the presence of pose changes, and (iv) to decouple the effects of motion correction, frequency correction, and shim correction. A VOI was chosen in the right central white matter as this region is expected to have minimal interaction between pose change and second-order B0 inhomogeneities. Of the six SVS PRESS acquisitions, the first three were baseline scans without movement and included the original Siemens sequence, a sequence with our navigator but no feedback, and a fully shim- and motion-navigated sequence. These were acquired in a random order. For the remaining three SVS PRESS acquisitions, the volunteers had been trained to lift their chin by approximately 8° on receiving a cue at 20 sec, to drop it to its rest position and rotate their head left by approximately 10° at a cue 48 sec later, and finally return to their rest position a further 52 sec later. To avoid spectral dephasing, volunteers were instructed to pace each movement to occur slowly and smoothly over a two to three TR window. The first of the SVS PRESS sequences with motion was fully shim- and motion-navigated, the second was only motion-navigated, and the third had no navigator feedback applied. For all of the above acquisitions, the shim was optimized first using the scanners automatic “advanced” shim adjustment and then further manually adjusted to acquire a T2* ≥ 43 msec and a water linewidth <8.5 Hz. These six SVS acquisitions are summarized in Table 1.
Table 1. Summary of the Six SVS Protocols Acquired for Each Volunteer in the Right Central White Matter
Real-time shim update
ShMoCo, shim- and motion-corrected; MoCo, motion corrected; NoCo, no correction applied.
SVS PRESS with vNav
SVS PRESS with vNav (ShMoCo)
SVS PRESS with vNav (ShMoCo)
SVS PRESS with vNav (MoCo)
SVS PRESS with vNav (NoCo)
The VOI was positioned using a multiecho magnetization prepared rapid gradient echo. The SVS PRESS voxel was (2 cm)3 with a TR of 2000 msec, TE of 30 msec, 512 readout (ADC) sample points, bandwidth of 1000 Hz, frequency offset −2.6 ppm, water suppression bandwidth 35 Hz, 64 averages in addition to four dummy or preparation acquisitions. For each volunteer, a water reference FID with the same parameters, apart from TR = 4000 msec and a single average, was acquired using the manually optimized shim for further processing in LCModel (20). The LCModel measures of linewidth and signal-to-noise (S/N), as well as the spectra themselves, were compared for spectra acquired with the different protocols.
To demonstrate the versatility of the vNav and its use in a VOI with higher B0 inhomogeneity, three additional SVS PRESS scans were acquired in volunteer 5. The VOI chosen was the medial frontal grey matter, as depicted in Fig. 2. The vNav protocol was changed before the first scan in this VOI using a 1 sec “set” scan. This “set” scan sets both the new vNav EPI protocol, and the subject specific vNav position. This protocol had an increased resolution of 5 × 5 × 5 mm3, reduced FOV of 220 × 200 × 110 mm3, matrix 44 × 40, 22 partitions, TEs of 8 and 12.8 msec, TR of 21 msec, and a bandwidth of 3906 Hz. The three scans had the same SVS parameters as above and varied as follows: (i) stationary baseline scan, without any correction, (ii) shim and motion corrected scan with movement, and (iii) only motion corrected scan with movement. The subject was asked to move in the same manner as described above.
Investigation of the Effect of Motion on B0
Figure 5 shows the effects of motion on B0 homogeneity in four VOIs for a single volunteer. Figure 5a shows the volunteer's motion about the scanner's isocenter for the relevant axes for both the chin down–up and left–right trajectories. Field maps were acquired at six different head poses along each trajectory. Figure 5b shows the mean frequency change in each VOI as the head moves. To ease comparison, the plots were offset to cross zero at the neutral head pose (pose 3 of chin up–down) by 17.1, 45.4, 37.8, and −90.7 Hz for the medial frontal, right frontal, right central, and lower occipital regions, respectively. Figure 5c shows how the magnitude of the first-order shim estimates change in each VOI. As the second-order shim requirements are the greatest in the frontal lobe, the changes in the five second-order shim estimates for only the medial frontal VOI are plotted in Fig. 5d. The first- and second-order shim estimates are also offset to cross zero at the neutral position of the chin down–up trajectory.
The magnitude of the second-order shim terms required in the neutral head position of the chin up–down trajectory are compared for the different VOIs in Fig. 6. This shows that, for most of the terms, the frontal lobe requires a significantly higher second-order B0 shim when compared with the central and inferior occipital regions.
In Vivo vNav Validation
Figure 7 is a single navigator volume from a single TR showing the magnitude volume of the first echo and the field map derived from the navigator. The linewidth and S/N, as measured by LCModel, were compared for the different sequences. The VOI for all the scans was in the right central white matter. Figure 8a and b present the mean (±SD) of the S/N and linewidth for each of the different SVS protocols calculated over all the volunteers. There is no loss in S/N because of the navigator and no increase in linewidth, when using real-time shim- and motion-correction for both stationary and moving scans. When no shim correction was applied to the scans with movement, the linewidth increases on average by ∼ 2 Hz, whereas the variability of linewidths increases dramatically with no motion correction. The S/N is 33% lower in the scans with movement when no shim correction is applied (significance P < 0.05). Figure 9 shows the spectra for all the scans acquired with movement (fully shim and motion corrected, motion corrected only, and uncorrected scans) for each of the six volunteers superimposed on top of the respective baseline scan without movement. These plots demonstrate that the spectra are affected by the pose change, when no correction and only motion correction are applied.
The motion and shim estimates measured as a function of time by the vNav for one of the acquisitions with motion, but with only motion-correction applied, is presented in Fig. 10. As no shim correction was applied during the scan, the plotted VOI frequency shift was computed using an offline spectral cross-correlation of each FID acquired in the scan to the first FID. The first-order shim estimates have been offset by their values at data point 3 for ease of comparison. As frequency and shim were measured regardless of whether such feedback was applied, Fig. 11a and b present scatter plots of the frequency and Y-axis shim change for all three scans with movement in all six volunteers as a function of the angle of rotation about X because of the chin up movement. The Y shim gradient is plotted here, as it was affected most by this movement as seen in Fig. 10c. These frequency, shim, and head rotations were measured for the maximum chin-up rotation, averaged over the duration that the subject maintained that pose. These scatter plots demonstrate a correlation of frequency and shim with head pose. For the Y-axis shim, the trend is roughly a 1° to 1 μT/m (y = 0.99 θ + 0.6) relationship between the head up down angle and Y shim gradient.
To decouple the effect of frequency shift from first-order shim changes on line broadening, the three scans with movement for three of the volunteers (4, 5, 6) were processed offline to remove frequency shifts by cross-correlating the spectrum of each FID to that of the first FID and producing a new frequency-coherent average for each scan. These frequency-coherent averages had a mean linewidth (±1 SD) averaged across the three volunteers of 4.9 ± 0 Hz with shim- and motion-correction, 6.8 ± 0.9 Hz with only motion correction, and 6.5 ± 0.5 Hz when no real time correction is applied. This demonstrates a loss in linewidth purely as a result of first-order shim changes, independent of frequency shifts, of just <2 Hz, when no shim correction is applied.
For the three additional SVS scans acquired in the medial frontal lobe of volunteer 5, the scan with motion and shim- and motion-correction applied had the same linewidth and S/N as the baseline stationary scan, namely 4.9 Hz and 20, respectively. By comparison, the moving scan with only motion correction had a linewidth of 6.8 Hz and S/N of 14. The chin-up rotation produced a shim change in Y of 10 μT/min, whereas the chin-left rotation produced the same shim change in Y of 10 μT/min and additionally a 15-μT/m change in X, both calculated as the mean during the respective pose.
SVS is a technique that inherently lacks anatomical information and thus accurate volume localization cannot be guaranteed. Furthermore, the accuracy of the spectra may be adversely affected by artifacts induced by pose change, including phasing errors, line broadening, and frequency drifts over time that may or may not be observable.
The changes in B0 resulting from changes in pose are dependent on the region of interest. As Fig. 5 demonstrates, all four VOIs exhibited a frequency and first-order B0 change with both chin-down to chin-up and left to right movements; however, first-order B0 changes were not significant in the right central region for left to right pose change. The second-order B0 shimming requirements are significant in the frontal lobe and as such the ability to adapt these second-order shim terms in relation to pose changes would be beneficial, although it is not possible on the current hardware.
In this work, a vNav capable of measuring head pose and B0 shim correction factors within a single TR was demonstrated. This vNav is ideal for spectroscopy, as it provides a series of anatomical images with sufficient resolution to provide online pose estimation and offline registration of the spectra to an anatomical image. The accuracy of the volume-to-volume registration performed by PACE was not scrutinized in this study; however, the baseline fluctuations in position estimates were well below 1 mm and 1°. Two sources of image artifact in the vNav are: the presence of dark bands resulting from the three PRESS slice selection planes and image distortions because of the use of a spectroscopy-specific shim. The impact of the dark bands is minimized by the use of SVS water suppression; the water suppression perturbs globally the magnetization of the entire volume, thus minimizing the dark bands. The second is inherently corrected by the navigator after the first TR by applying the appropriate first-order shim for the vNav as calculated from the vNav itself and thereafter alternately switching the shim between the best calculated values for the navigator and the best values for the spectroscopy VOI.
The first question investigated in this study was whether the navigator impacted the signal of the spectra. S/N measurements in LCModel demonstrated that the vNav had no effect on the S/N when compared with acquisitions with no navigator (shown in Fig. 8). It should be noted that a change in TR from 2 sec to 1.5 sec would have a noticeable effect on this S/N. The accuracy of the navigator's shim estimate is demonstrated by the linewidth of the baseline acquisitions with real-time shim correction, where the mean and SD did not exceed that of the manually optimized shim.
This vNav technique provides three real-time adjustments: VOI location, frequency, and first-order shim correction. The benefit of applying all three is clearly demonstrated by the narrow linewidth and high S/N in the presence of pose change (Fig. 8). The effect of shim adjustment and VOI localization correction was decoupled by acquiring an acquisition with only motion correction. To separate the effects of frequency shifts from first-order shim errors on linewidths, offline frequency correction was applied to motion corrected data of three subjects. Although the linewidth improved with each adjustment, it was only fully regained by applying all three.
In the data presented, each subject's chin was raised for approximately 1/3 of the acquisition. This resulted in a mean line width increase from 4.9 to 6.8 Hz when only offline frequency correction was applied. This is consistent with Eq. 1, which suggests that for the duration of a 10-μT/m shim change, the linewidth will increase from 5 to 10 Hz. As this only occurred for 1/3 of the acquisition in the present case, the mean linewidth is expected to be 5 Hz + (5 Hz/3) = 6.7 Hz. The use of offline frequency correction demonstrated an improvement in linewidth solely due to shim correction and not due to frequency correction. This is because a single frequency shift, as is present in our data, results in a secondary peak, rather than line broadening.
The real-time shim data and results presented here are the raw estimates from a single navigator. One could further increase the stability of the navigator by taking into account the large time course of data available. This type of temporal filtering would prove beneficial should an accelerated version of the navigator be implemented.
The three scans in the medial frontal grey matter of volunteer 5 demonstrate the application of the vNav in a region of higher B0 inhomogeneity. In such regions, a higher vNav spatial resolution provides the specificity required for first-order shim measurements. This higher resolution vNav protocol has a limited FOV that requires the operator to position the vNav over the subject's brain. This additional step should be taken into account when choosing the appropriate vNav protocol for the VOI. This scan demonstrated significant first-order shim changes for both chin-up and chin-left rotations indicating the importance of this type of navigator in regions of high B0 inhomogeneity.
As already discussed the navigator is dynamically configurable; if a shorter TR is necessary, simply reducing the number of partitions acquired, and subsequently manually positioning the navigator, will enable TRs of 1.5 sec to be achieved. Acquiring the complete scanner FOV with the navigator simplifies the user interaction and was achievable with our TR of 2 sec. The duration of the complete navigator block presented is ∼ 1.1 sec and is faster than that achieved using the PROMO technique in SVS (8) of 1.5 sec. Further navigator optimization is possible by employing acceleration techniques like parallel imaging.
Finally, this work has not addressed phasing errors brought about by subject movement during the PRESS localizing gradients. Reacquisition, as presented by Zaitsev et al. (7) would be one possible solution; however, as single voxel spectra are accumulated over repeated measurements, it may be more appropriate to provide an offline tool to simply exclude dephased measurements.
Changes in B0 homogeneity were demonstrated in four different SVS VOIs in a single volunteer for different head poses. A vNav capable of measuring and adjusting, in real time within each TR, head pose, VOI frequency, and VOI first-order shim has been demonstrated. For restless subjects, whose head pose cannot be assumed to be constant, this provides a useful addition to the SVS sequence. The first-order shim estimates calculated by the vNav result in linewidths equal to those achieved with manual first-order shim optimization and maintain spectral quality in the presence of pose changes.
Thomas Benner, Michael Hamm, and Charles Harris have provided valuable assistance in this project. Resources necessary in the project were provided by University of Cape Town, Martinos Center, and Cape Universities Brain Imaging Centre.