Ultrahigh‐resolution quantitative spinal cord MRI at 9.4T

To present the results of the first human spinal cord in vivo MRI scans at 9.4T.


| INTRODUCTION
Today MRI is an indispensable method for clinical diagnostic and research investigations of the human spinal cord (SC). For SC-related diseases such as multiple sclerosis, 1-3 amyotrophic lateral sclerosis, 4 spondylosis, 5 or SC injury 6 MRI is the technique of choice for primary diagnostics, patient stratification, and therapy monitoring. Furthermore, it has been reported that in patients with multiple sclerosis, gray matter (GM) lesions in the SC may be detectable more readily than GM lesions in the brain, which illustrates the importance of SC imaging investigations in general. 7 Small structures within the SC are inherently challenging to image at lower B 0 field strengths; hence, field strength is of important relevance for SC measurements. Utilizing higher field strengths provides a higher signal-to-noise-ratio (SNR), facilitates higher spatial resolutions, and potentially improves diagnostic power. For example, it has been shown that the detection of small lesions in patients with multiple sclerosis at 3T MRI is improved versus 1.5T MRI. 8,9 Sigmund et al 10 and Zhao et al 11 have published comparisons of anatomical SC images, which display excellent improvement of the image quality at 7T versus 3T. Additionally, it has been shown that in patients with multiple sclerosis, on average 4.7 white matter (WM) lesions can be detected at 7T in the SC, whereas at 3T only 3.1 WM lesions can be found. This increase of 52% emphasizes the advantages of higher field strengths for SC MRI. 12 Barry et al 13 gave an excellent review on the current state of "Spinal Cord MRI at 7T" included also the advantages of ultrahigh-field strength for different imaging modalities beyond anatomical MRI. To take it a step further, this study presents in vivo MR measurements of the SC at a field strength of 9.4T, which could theoretically offer about twice the SNR compared with MRI at 7T with similar radiofrequency (RF) coils and sequences. 14,15 To our knowledge, with the exception of our own preliminary work published in a conference abstract, 16 these are the first human in vivo SC MRI results obtained at this field strength. Previous work at 9.4T has focused on human postmortem samples, 17 injured SC in squirrel monkeys, 18,19 and SC in mice 20 and rats. 21 A comprehensive overview on 7T SC coils 13 demonstrated that the majority of these designs had only posterior elements on a holder adapted to the neck and utilized only surface-loop receive elements. 10,11,12 In this work, a new approach to SC RF coil design is introduced that combines transceiver surface-loop elements with additional receive-only vertical loops and contains anterior as well as posterior coverage to increase the central SNR. This coil design 22 was originally suggested for improved central SNR in human brain scans. The increase in central SNR is also beneficial for imaging the cervical SC because of its central position in the head and neck.
Anatomical images of the cervical SC were measured with a T 2 *-weighted GRadient-Echo (GRE) sequence. 23 To address the SC MRI inherent issue of periodic B 0 inhomogeneities caused by the intervertebral discs, 24,25 the literature provides solutions as "electrocardiogram-triggered, higher order, projection-based B 0 shimming" 26 or z-shim gradient pulses. 25 To minimize the influence of these B 0 inhomogeneities, four different B 0 shimming routines were tested and compared during this work. Intrinsic SNR maps were calculated and the impact of different echo times (TEs) on image quality and contrast was investigated to find optimal sequence parameters.
Recently, anatomical SC images with an in-plane resolution of 0.18 × 0.18 mm 2 have been published, 10,27 which is the highest resolution of in vivo human SC measurement in the literature. In this study, the resolution could be further improved and anatomical images with an in-plane resolution of 0.15 × 0.15 mm 2 are presented.
Additionally, algorithmic SC detection and GM/WM segmentation was performed, as well as calculation of T 2 *-relaxation-time maps of the human SC at 9.4T.

| Radiofrequency coil
All measurements in this study were performed with an in-house-built 16-channel tight-fit array coil. 22 This array consists of eight transceiver surface loops and eight receiveonly loops. The eight transceiver surface loops are placed on a cylindrical holder. Each of the eight vertical receiveonly loops is positioned along the central axis of a transceiver loop, perpendicular to its surface. Figure 1A consists of a photograph of the coil and some of the transceiver and receive-only loops.
This coil was originally constructed for human brain scans with an emphasis to enhance the SNR in deep brain structures and high-transmission efficiency. The addition of the vertical loops has no measurable effect on the Tx efficiency of the array. However, the 16-channel set-up provides an improvement of the SNR in vivo of approximately 30% in the center of the brain, compared with the surface-loops-only set-up.
The central SNR improvement was the motivation for testing the coil for SC data acquisition. The coil array and its housing geometry is cylindrical, with openings at top and bottom. This geometry makes it possible to place the subject further inside the coil than originally intended and allow the cervical SC to be imaged. Figure 1B depicts the subject positioning for brain scans versus that for SC scans. For brain scans, the subject's brain is placed in the center of the loop array; however, for cervical SC measurements in this study each subject was positioned further into the coil until the shoulders contacted the coil housing. In that position, the upper cervical SC is located in the center of the loop array.

GELDSCHLÄGER Et aL.
During RF transmission, the array was driven in the circular polarized mode with a 45° phase shift between adjacent transceiver loops. This was achieved with an eight-way splitter box, with corresponding phase shifts incorporated.
A comprehensive set of safety tests of the RF coil has been performed according to our internal standard operational procedures approved by the local institutional review board; the coil was certified for in vivo use in human subjects as described earlier. 28 The specific absorption rate was continuously monitored and supervised by means of the k-factor, 28 which is incorporated in the coil file. Sequence parameters that would exceed the specific absorption-rate limit are hence not allowed, and the parameters have to be adjusted accordingly prior to the scan. In case the specific absorption rate is still exceeded during the scan, it is stopped automatically by the vendor-integrated specific absorption-rate-monitoring system.

| Volunteer scans
Ten healthy volunteers (eight men, two women, age: 27.4 ± 4.72 years, weight: 75.4 ± 11.77 kg) were scanned on a 9.4T whole-body MR scanner (Siemens Healthcare, Erlangen, Germany) equipped with a SC72 whole-body gradient system with a maximum amplitude and slew rate of 40 mT/m and 200 mT/m/ms, respectively.
All experiments were performed with the approval of the local ethics committee. Informed signed consent was obtained from each volunteer before each MR experiment.

| Study design
Each in vivo scan session was performed with the following routine. A localizer sequence was applied and used to position the B 0 shimming volume ( Figure 1C). The tested B 0 shim approaches are explained in the B 0 shimming comparison subsection below.
After B 0 shimming, a flip-angle (FA) map using the AFI (Actual Flip-angle Imaging) sequence 29   Autocalibrating Partially Parallel Acquisitions (GRAPPA)]) acceleration) was acquired to check the achieved mean FA at the SC position (see FA map in Figure 1D). If the mean FA was unequal to the desired one, the transmit voltage was adjusted according to the proportion of desired and measured FA. After adjustment of the transmit voltage, the AFI sequence was applied and the mean FA was evaluated again to ensure correct FA adjustment. The FA homogeneity within the SC was assessed visually in the FA map.
All anatomical images were acquired with an axial T 2 *weighted GRE sequence. 23 The sets of parameters used during this study are listed in Table 1. Every time sequence A, B, C, D, E, F, G, H, or I is mentioned in the text, respectively, a GRE sequence with the parameters displayed in Table 1 is referred to. All sequences acquired one average, and phase stabilization was active. Each parameter set was tested on at least three subjects. Sequence B was applied with a slice oversampling of 66%.

| B 0 shimming comparison
Along the SC the B 0 field varies dramatically because of transitions between bone and intervertebral discs 24,25 ; these "B 0 field distortions interfere with encoding of contrast and spatial origin of the MR signal, commonly causing artifacts in the form of signal loss and geometric distortion, but can also contribute to ghosting, blurring and distorted excitation volumes, amongst other effects." 13 Thus, as input for the scan protocol optimization for anatomical SC imaging, multiple B 0 shimming algorithms were compared to investigate the influence of the B 0 shim performance on SC image quality. In particular, the primary focus evaluating B 0 shim performance by the severity of signal dropouts that occur periodically in the intervertebral disc-located slices.
For all tested B 0 shimming routines, the same threedimensional cuboid B 0 shimming volume was set on the localizer image. The B 0 shimming volume was planned to encompass a region including the SC and approximately 3 cm of the surrounding tissue ( Figure 1C).
The first shimming routine was the Siemens implemented image-based shim routine using up to second-order spherical harmonics. Within this routine the B 0 map is measured with a Double-Echo Steady-State (DESS) sequence 30 ; based on that, the shim values for each transmit channel were calculated and applied. Frequency adjustment was performed afterwards. The whole B 0 shim procedure was repeated three times to achieve robust convergence, which is necessary because the Siemens shim algorithm implementation does not directly consider B 0 shim field imperfections.
The next tested routine was the Fast, Automatic Shim Technique using Echo-planar Signal readouT for Mapping Along Projections (FASTESTMAP). 31 Again, the process from B 0 mapping to frequency adjust was performed three times before the image acquisition.
The third routine was the ConsTru (constrained TSVD inversion) method. 32 After acquiring the B 0 map (with all shim values set to zero), the ConsTru shim values were calculated and applied to the scanner. Before the anatomical image was measured, the frequency adjust was performed.
Lastly, the straightforward TuneUp shim was applied. The TuneUp shim is the shim setting the vendor determines to deliver a homogeneous magnetic field in a spherical phantom after installing the MR scanner. It characterizes the B 0 shim quality and is often used as an initial point for further B 0 shim calculations.
After adjusting each set of B 0 shim parameters, the anatomical images were acquired with sequence A.

| Calculation of the intrinsic signal-to-noise ratio
To create intrinsic SNR maps ("the SNR that would be obtained with a homogeneous FA of 90°, an infinite repetition time and an echo time of zero" 14 ), signal measurements were acquired using sequence B. Afterwards a corresponding noise scan was performed by applying the same sequence with the transmit voltage set to zero. Based on the noise data, the noise covariance matrix 33

| Sequence parameter optimization and high-resolution anatomical images
To show that the RF coil can be used to acquire high-quality anatomical images of the SC, single low-resolution axial slices were acquired with sequence C, utilizing each time another TE (5,9,14, or 19 ms) as initial input for scan parameter optimization. Following the feasibility test, multiecho sequences, where the signal is measured at four different TEs, were applied (sequence D and E). Within these sequences, 12 different transversal slices with an in-plane resolution of 0.23 × 0.23 mm 2 were acquired along the upper cervical SC (level C1, C2, C3, and partly C4). The parameters of sequence E were set analogously to those set in Massire et al, 27 because these TEs produced good results at 7T. The data from both sequence D and E were combined to create MEDIC (Multi Echo Data Image Combination) 36 images and to calculate T 2 * maps. MEDIC images were obtained by computing the sum of squares of the image from each TE. To assess the GM/WM contrast for different TEs, the contrast-to-noise ratio (CNR) for each TE and each slice was calculated. We defined the CNR (analogously to Bachmann et al, 9 Fushimi et al, 37 and Constable and Henkelman 38 ) as the difference between the mean signal of GM and WM, divided by the standard deviation of artifact free background noise. To average the signal intensities, the GM and WM tissue masks resulting from image segmentation (see Segmentation subsection) were utilized.
By  38 Hendrick, 39 Edelstein et al, 40 and Brown et al, 41 were applied. Following these rules, the experimental SNR for the parameters of sequence D and E was derived.
To observe the benefits of very high-resolution images, measurements (sequence F, G, and H) with the highest axial in-plane resolution (0.15 × 0.15 mm 2 ) achievable with the proposed sequence settings and gradient timings, were performed to make the small SC structures visible. The scanner did not allow the measurement of higher resolutions because the gradients could induce too severe peripheral nerve stimulation. Although sequence F is a single-echo sequence, the sequences G and H are multiecho sequences, with two echoes, respectively. For that high resolution, the scanner did not permit the measurement of four echoes in one sequence because of the risk of too strong nerve stimulation. The data from both sequences G and H were combined to create MEDIC images.

| Segmentation
Using the Spinal Cord Toolbox (v.4.0; sourceforge.net/projects/spinalcordtoolbox), 42 algorithmic SC detection, as well as GM/WM segmentation within the SC, was performed. The Spinal Cord Toolbox is a comprehensive software program used to process MRI SC data. From the wide range of applications and algorithms implemented in the Spinal Cord Toolbox, the propagated cord segmentation (function sct_propseg) 43 for detecting the SC within the full FOV and the deep-learning algorithm with dilated convolutions (function sct_deepseg_gm) 44 to segment the GM within SC were employed. Once the SC detection and the GM segmentation were executed successfully, the WM segments were detected by subtracting the GM mask from the full SC mask. To test the SC detection algorithm, a GRE sequence acquiring 128 slices (sequence I, slice thickness: 1 mm) was applied. With these 128 transversal slices, it was possible to create sufficient sagittal and coronal coverage to visualize the SC detection result in all three dimensions. The SC detection algorithm was also tested on the data acquired with sequence D and E. Additionally, on this data the GM segmentation algorithm was applied.

| T 2 *-time calculation for gray matter and white matter
To perform pixel-wise T 2 *-time calculations, the multiecho results from sequences D and E were employed. In addition to the four TEs adopted from Massire et al 27 (sequence E), another four TEs (sequence D) were measured. This yielded eight, instead of four, sample points per pixel to calculate a more accurate mono-exponential fit. The fitting was performed with a nonlinear least squares algorithm using MATLAB (MathWorks, Natick, MA).
To calculate the average T 2 *-relaxation times of GM and WM separately, the segmented GM and WM tissue masks were applied to T 2 * maps. All the fitted T 2 * values of GM and WM pixels from three subjects were pooled to calculate mean GM and WM specific T 2 *-relaxation times and a two-sample t test was performed with MATLAB to investigate statistically significant differences. Figure 2 presents the results acquired with sequence A. The Siemens shim routine offered the images with the best quality and was consequently utilized for all of the following measurements. Most notably, the Siemens shim performed much better in the first and second slice. Signal dropouts, which are visible for the other three set-ups, arose just marginally for the Siemens shim. For slices three to six, all routines provided excellent image quality. The Tune-Up shim, which consists of a fixed set of shim values, delivered similar image quality to those of other shimming routines. For slice seven, the Siemens shim again offered the best image quality. Because this slice touched a disc, a slight signal dropout occurred in the lower area of the SC and cerebrospinal fluid (CSF). This dropout worsened for Tune-Up, FASTESTMAP, and ConsTru shim relative to the Siemens shim. For slices eight and nine, again all routines provided similar, sufficient image quality. Figure 3A shows a map of the intrinsic SNR, represented by 6 out of the 12 measured slices. For the two presented subjects, the highest SNR in the SC was achieved in slices 3 to 10, with values mostly observed between 6600 and the maximum of 8060 (subj.1, slice 4). For slices positioned further in the head direction than slice 3 and placed further in the feet direction than slice 10, the SNR dropped dramatically ( Figure 3B).

| Sequence parameter optimization and high-resolution anatomical images
First, a single slice with a low in-plane resolution of 0.47 × 0.47 mm 2 was measured four times, each time with a different TE. Figure 4 displays the FOV of one volunteer. Anteriorly, the teeth can be seen and on the posterior side, the neck muscles are visible. In the center of the FOV, the SC is surrounded by CSF (yellow box) and the vertebrae. In the four surrounding images in Figure 4, enlargements from the F I G U R E 3 A, Intrinsic signal-to-noise ratio (SNR) map. Six of 12 measured slices are displayed by skipping every second slice. From left to right, the number of the displayed slice, the full field of view of that slice, and an enlargement of the spinal cord (SC) is visible. The slice positioning is similar to Figure 2B. The white circle in slice three marks the values that were averaged. The same circle (adjusted for the SC position, respectively) was used in each slice. B, Slice-wise averaged SNR values for two subjects. The SNR map from subject 1 is shown in A FOV at the SC location at different TEs are displayed. At a short TE of 5 ms, a high signal with low GM/WM contrast is visible. With increasing TE, the signal has partially decayed and GM/WM contrast has increased. The "butterfly" shape of the GM is clearly visible and distinguishable from the WM for a TE of 9 ms and 14 ms; furthermore, nerve roots at the left and right side of the SC are apparent. A TE of 19 ms seems to be too long, as GM/WM contrast has diminished and magnetic susceptibility effects appear, especially in the CSF.
In Figure 5A, the results of the two multiecho GRE sequences D (TEs: 4/10/16/22 ms) and E (TEs: 5.1/9.4/13.8/17.7 ms) with an in-plane resolution of 0.23 × 0.23 mm 2 are depicted. Figure 5B shows the corresponding CNR for each slice and TE. Figure 5C illustrates the averaged experimental SNR in the SC for each TE recalculated from the intrinsic SNR depicted in Figure 3B (Subj.1).
As Figure 4 has already shown, as TE increases, the signal decays and GM/WM contrast grows. While a TE of less than 9.4 ms seems too short to achieve a strong visual GM/WM contrast, a TE of between 9.4 ms and 13.8 ms provided the best compromise between signal intensity and tissue contrast as illustrated in Figure 5A. Confirming this observation, the CNR between GM and WM is increasing from a TE of 4 ms until 13.8 ms for most slices. After 13.8 ms, the CNR is decreasing or rises only marginally.
Slice five is located at an intervertebral disc location. With increasing TE, a signal dropout occurred here. While this dropout does not or only marginally occurs for the first four TEs, it became severe for a TE of 13.8 ms and longer.
To not adulterate the CNR of that slice, the GM and WM masks were corrected manually to exclude the pixels where the dropout occurs (see Supporting Information Figure S1 for the CNR values including signal dropout).
The experimental SNR is decreasing along an exponential curve with increasing TE. In slice three (which had one of the highest intrinsic SNR values of 8000 as seen in Figure 3) the highest experimental SNR was acquired.
The MEDIC images offer excellent GM/WM contrast, as it can be seen in the images in Figure 5A, as well as from the CNR values in Figure 5B. For instance, slice three provides a CNR of 4.95 at a TE of 13.8 ms and a CNR of 7.95 for the MEDIC image.
The results given in Figure 5 were acquired with a TR of 500 ms and a FA of 50°; the results given in Figure 4 were acquired with a TR of 250 ms and a FA of 25°. After a proper scaling, there is no visible difference in the image quality.
As TEs between 9 and 13.8 ms are good compromises between tissue contrast and minimal magnetic susceptibility effects, a TE of 12 ms was chosen for the single-echo measurements with an in-plane resolution of 0.15 × 0.15 mm 2 (sequence F). In all depicted slices ( Figure 6A), high-quality images of the SC are visible. GM and WM can be distinguished. Nerve roots and blood vessels are shown in excellent detail in most of the slices. The anterior median fissure and the posterior median septum are recognizable. The same is valid for the MEDIC images ( Figure 6B

| Segmentation
The white-colored pixels in Figure 7 mark the SC, as it was masked by the SC detection algorithm. All white-colored pixels mark the location of the spinal cord in each perspective correctly, except for the bottom slices, where the SC was not detected. On the slices acquired with sequence D and E, the SC detection algorithm and the GM segmentation algorithm were applied. Figure 8 shows three segmented example slices. The algorithm detected the SC and segmented GM and WM correctly. The typical butterfly shape is observed as depicted by white pixels in the GM mask.

| T 2 *-time calculation for gray matter and white matter
For each of the three subjects that were measured with multiecho sequences D and E (12 slices acquired per sequence), a T 2 * map was calculated. Figure 9 shows an example slice, where the CSF and the SC are distinguishable; to some extent, the GM and WM within the SC are also discernible. T 2 * slices, in which these tissue types were not recognizable because of midvertebra disksrelated signal dropouts, were excluded from averaging. Based on the created GM and WM masks, mean T 2 * values were calculated. For the cervical SC at 9.4T, the calculated T 2 * value for GM is 24.88 ms ± 6.68 ms and for WM 19.37 ms ± 8.66 ms.
A two-sample t-test results in p value of 1.6233*10 −271 for the null hypothesis ("The mean of the GM sample is equal to the mean of the WM sample."). For that reason, the null hypothesis needs to be rejected. Thus, the means of GM and WM T 2 *-relaxation times differ significantly.

| DISCUSSION
This work presents the first in vivo MR images of the human SC measured at the B 0 field strength of 9.4T. An RF coil optimized for central SNR in the human brain was employed to acquire in vivo cervical SC images.
A TE between 9 and 13.8 ms was found to offer the best image quality with an excellent contrast between CSF, GM, and WM (Figures 4 and 5). That TE recommendation is slightly higher in comparison with what Sigmund et al 10 suggested at 7T with their SC-dedicated coil (GRE sequence with TE = 4.91 ms). In comparison with the noncombined data, the combined MEDIC images 36 show improved image quality in terms of tissue contrast. Because the acquisition of multiecho data does not require additional scan time, MEDIC is an attractive alternative to single-echo measurements. Slice one in Figure 5A shows high contrast between CSF and the SC largely independent of TE, which possibly occurs because of the reduction in transmit B + 1 field at the edge of the FOV of the RF coil in that slice position.
The highest axial in-plane resolutions (in vivo) acquired prior to this work were 0.18 × 0.18 mm 2 by Sigmund et al 10 ( Figure 6C) and Massire et al 27 ( Figure 6D) at a field strength of 7T with SC-dedicated coils. We presented an improved resolution of 0.15 × 0.15 mm 2 . That means in the 9.4T images ( Figure 6A,B) the area of each pixel is decreased by approximately 30% from 0.0324 mm 2 to 0.0225 mm 2 compared with the 7T acquisitions. The 9.4T results show the small SC structures in excellent detail. The quality of these images seems to outperform the 7T results, as the nerve roots are also visible at 7T, but are not as sharp as at 9.4T. Moreover, the anterior median fissure and the posterior median septum are recognizable in T 2 *weighted images at 9.4T, whereas they are hard to detect at 7T. 10,27 In the most recent work from Massire et al, 45 these two structures are recognizable on 7T T 1 maps of the SC, but again difficult to detect on T 2 *-weighted images, as is possible at 9.4T. Compared with the acquisitions with an in-plane resolution of 0.23 × 0.23 mm 2 ( Figure 5A), the 0.15 × 0.15 mm 2 images present the SC in slightly improved acuity. However, the enhancement may not be as strong as expected from the difference between both spatial resolutions. A possible reason for this is the physiological SC motion (between 0.40 and 0.50 mm in superior/inferior, 0.60 ± 0.34 mm in anterior/ posterior, and 0.17 ± 0.09 mm in right/left direction 46,47 ), which possibly blurs images, especially in cases of very high spatial resolutions. Cardiac triggering or prospective motion correction 48 could be an option to mitigate the influence of motion; however, it was not applied during this study. Future work should test whether image acuity at a spatial resolution of 0.15 × 0.15 mm 2 can be improved and even higher spatial resolutions are feasible by including optimized motion compensation methods and encoding schemes that prevent excessive peripheral nerve stimulation. Even if a higher spatial resolution would be possible technically, it remains to be clarified whether this can further improve diagnostic specificity and sensitivity. T 2 * maps were derived from multiecho GRE images and showed clear distinction between GM, WM, and CSF. By means of the Spinal Cord Toolbox, masks for GM and WM were calculated, and the T 2 * times for all GM and WM voxels inside the respective masks were averaged. Although an uncertainty remains when using this pixel-by-pixel fit method (see noise in Figure 9 and the resulting relatively high SD) the first estimate of T 2 * times in the human SC at 9.4T has been presented (GM 24.88 ms ± 6.68 ms and for WM 19.37 ms ± 8.66 ms) and a statistically significant difference was found between GM and WM. These results are consistent with the literature. While Massire et al 27 presented the F I G U R E 8 Spinal cord (SC) and gray matter (GM) segmentation applied on the results acquired with gradient echo sequence E (echo time: 9.4 ms). Enlargements of three transversal example slices are shown in the left column of images. The central column of images depicts where the SC-detection algorithm determines the SC (white-colored pixels). On the right hand side, one column depicts where the GM-segmentation algorithm determines the GM (white-colored pixels) within the SC. The other column presents the white matter (WM) pixels F I G U R E 9 Enlargement at the spinal cord position of an example slice from a T 2 * map expected slightly higher T 2 *times from the cervical SC at 7T (GM: 29.3 ms ± 4.5 ms, WM: 23.5 ms ± 5.7 ms), Pohmann et al 14 published T 2 *times at 9.4T for the human brain (GM: 23.8 ms ± 1.0 ms, WM: 19.2 ms ± 0.9 ms), which are very similar to the values calculated for the SC at 9.4T.
Periodic signal dropouts were observed in slices near intervertebral discs, possibly due to strong magnetic susceptibility between soft tissue and bone. 24,25 although among the herein investigated B 0 shimming approaches, the vendor-implemented second-order B 0 shimming offers the best image quality for slices next to intervertebral disks; in general, the differences within the four different shimmed images are not very prominent (Figure 2). Even the TuneUp shim, which is a fixed set of shim values, does not have a substantially poorer performance in most slices in comparison with FASTEST MAP and ConsTru. Although the vendor shim mitigates the dropouts, it did not obliterate it. Tackling the B 0 inhomogeneities for SC imaging remains a future challenge. A possible solution could be dedicated B 0 shim hardware. [49][50][51][52][53] In addition, a 3T study showed improved B 0 homogeneity within the neck by placing pyrolytic graphite foam behind the neck 54 ; this could also be worth investigating at ultrahigh field.
The simplicity of the utilization of a 16-channel transceiver array RF coil driven in circular polarized mode for SC imaging that was originally optimized to yield high SNR in central brain structures and high transmit efficiency for brain MRI and MRS is worth noting. Utilizing the same coil for both brain and cervical SC applications is cost-efficient and offers the opportunity to tackle "the lack of commercial SCdedicated coils" as reported in Massire et al. 27 There is no costly development of a SC-dedicated coil needed to achieve comparable results for the cervical SC. Two of the 10 subjects reported that the positioning inside the coil was not as comfortable as for brain scans; however, no subject needed to interrupt or abort a scan session.
We chose to investigate the intrinsic SNR of the SC (ie, the SNR in the situation of a homogeneous FA of 90°, an infinite TR and an TE of zero) to achieve a normalized SNR value that allows for the comparison between different field strength and hardware set-ups by correcting for the influence of the sequence parameters. It is worth noting that the reported intrinsic SNR values featured in Figure 3 are not achievable in experiments.
The intrinsic SNR reported in Pohmann et al 14 (approximately 3000 in brain center at 9.4T, but smaller voxel size as in sequence B) can be recalculated with respect to the parameter values from sequence B by multiplying it with a factor of 3.2 (eq. 7.12 from Hendrick 39 ). With the coil 55 used in Pohmann et al 14 and the settings from sequence B, the intrinsic SNR in the brain center theoretically would be 9600. This value is higher than the intrinsic SNR that was acquired in the SC during this study (Figure 3; SNR between 6600 and 8060 in slices 3 to 10). The lower SNR in the SC compared with the brain center at 9.4T is most likely caused by the different coil set-ups utilized. The dedicated brain coil 55 has 31 receive channels arranged on a close-fitting helmet to maximize SNR. The coil employed in this study was a single row eight-channel transceiver coil with eight additional receive-only elements and has, therefore, a limited coverage, which negatively impacts central SNR. 15 On the other hand, the combination of transceiver loops and dipolar receive-only elements was designed to enhance the SNR in deep-brain structures. 22 The design of the coil used herein allows SC applications, which is not the case for the dedicated brain coil design in Pohmann et al. 14 However, the coil loading differs between the subject position for SC imaging and brain imaging ( Figure 1B), which may also lead to SNR losses. All these influence factors could be reasons for the slightly lower SNR in the SC as compared with the central brain at 9.4T. 14 In Sigmund et al, 10 the authors report an average experimental SC SNR in vivo of 227 at 7T. If we recalculate the intrinsic SNR from this study (Figure 3, slices 3 to 10) with respect to the sequence parameter from Sigmund et al 10 (utilizing eq. 7.12 from Hendrick 39 and eq. 1 from Pohmann et al, 14 acquisition bandwidth was not mentioned; therefore, we assumed the same bandwidth as in sequence B) and we achieved an experimental SNR between 265 and 324. The SNR increase was expected; however, it is not as high as the change from 7T to 9.4T may imply. 14,15 In general, an increase of the static magnetic B 0 field strength leads theoretically to an almost quadratic increase of SNR and thus doubles the SNR at 9.4T versus 7T for similar sequences and coils. 14,15 Furthermore, at high field a combination of surface loops and dipolar elements is needed for SNR optimization in central regions. 15 In addition, wrap-around coils (such as the coil in Zhang et al 56 ), tend to provide higher SNR compared with the neck coils used by Sigmund et al 10  The decreased SNR in slices 1, 2, 11, and 12 ( Figure 3), as well as the failed SC detection in the lowest positioned slices in Figure 7, indicates a drawback of the proposed coil: the relatively small longitudinal coverage for areas that are not placed in the center of the loops. Vertebras C1, C2, C3, and to some extent, C4 can be covered excellently, but reaching any lower vertebra seems to be problematic. In addition, measuring the brain and the SC within the same scan is not possible. To achieve extended coverage, other array architectures (eg, with dipole receive elements instead of loops 57 or SC-dedicated coils (eg, see Sigmund et al, 10 Zhao et al, 11 Kraff et al, 58 or Barry et al 13 for an overview of all 7T SC coils) need to be employed. Having a coil design that considers the principles of the proposed coil, but is further optimized for SC imaging could improve results with respect to longitudinal coverage, SNR, and subject comfort. With an extended longitudinal coverage, the brain and the SC could be imaged in the same experiment, as it is necessary to assess the functional connectivity between both regions.
With this work, we have taken the first step into SC research at a B 0 field strength higher than 7T. By testing and optimizing other sequences (eg, magnetization-prepared 2 rapid gradient echoes [MP2RAGE], 59 T2W, echo planar imaging, MRS, or the AMIRA approach 60 ) in the future, anatomical image quality could be further improved and multimodal SC imaging at 9.4T can be enabled. Furthermore, T 1 and T 2 relaxation-time measurements need to be performed to complete the set of tissue relaxation times as input for sequence optimization and quantitative MRI for the SC at 9.4T.

| CONCLUSION
With this work, the first step into SC research at a B 0 field strength of 9.4T was performed. With an RF coil, originally optimized for maximal SNR and transmit efficiency in deep brain structures, we were able to acquire high-resolution (inplane: 0.15 × 0.15 mm 2 ) anatomical SC data in excellent detail and quality. Algorithmic SC detection, as well as GM/ WM segmentation, reliably works on the data. The T 2 *times of GM and WM in the human SC at 9.4T were presented. Such high-resolution images at ultrahigh field of the human SC might open new possibilities in the field of SC research and clinical patient care. Knowing the T 2 *-relaxation time allows optimization of imaging parameters and potentially enables improvements in sensitivity and contrast.