Pathologies of the spine and spinal cord can result from neurodegenerative and vascular diseases, trauma, and cancer, all of which can induce severe functional disability and neuropathic pain. While precise diagnosis is required to determine the optimal intervention for treating patients, clinicians lack objective tools to assess the microstructural damage to the tissue. For example, what is the extent of the lesion and what are the types of axons that are damaged/spared? These factors can have a dramatic impact on the functional recovery of the patient and knowing them would guide the choice of interventional therapy. Moreover, in order to test the efficacy of new treatments, it is essential for researchers/practitioners to know the actual effect of the treatments on the tissue (regeneration and/or remyelination?). While conventional MRI cannot specifically identify the type of axonal damage, recent developments in multi-parametric MRI (mpMRI) have shown a strong potential for characterizing tissue microstructure [1]. Although mpMRI techniques have been used successfully for the brain, their application to the spine remains challenging. Reasons include the poor signal-to-noise ratio (SNR), susceptibility artifacts inducing image distortion and signal drop-out and high level of physiological noise (particularly problematic for functional MRI studies).

In this review we will go through different aspects of mpMRI in the spine and spinal cord, including hardware, pulse sequences and images processing techniques. Each section aims to present the state-of-the-art mpMRI techniques, also touching on their pros and cons. The goal of this review is to give an overview of the field with some of the current developments. For more detailed information about the “how-to” of spinal cord mpMRI see [2]. While this review focuses on the spinal cord, most of these techniques are also applicable to the spine and plexus.

RF coils

Typically, signal transmission is done via the single-channel body coil integrated into the MRI system, to achieve good homogeneity, while signal reception is done via multi-channel coils close to the imaged structure, to achieve good sensitivity (i.e., signal-to-noise ratio, SNR). More sensitive and highly-parallelized radiofrequency (RF) receive coils have the potential to provide better image quality in fMRI. Higher sensitivity brings more SNR and thus better temporal SNR, which is directly related to the statistical power in fMRI. Highly-parallelized coils (i.e., high number of channels) enable higher acceleration factors for parallel imaging, which directly translate to reduction in susceptibility distortions (see section 4). While standard commercial coils for the neck only have 2-4 channels, advanced coil designs have been proposed that cover the neck region with more coils and smaller coil elements (see Figure 1).

Shimming

As mentioned in the introduction, one challenge of spinal cord fMRI is the presence of susceptibility artifacts, which appear as image distortions in ultra-fast imaging sequences (echo planar imaging) typically used in fMRI and diffusion-weighted imaging acquisitions. The main cause of these artifacts are large inhomogeneities in the main magnetic field (B0), because of the presence of structures with different susceptibilities: the bony vertebral structure, the intervertebral disks and the air-filled lung, giving raise to small-scale and large-scale variations in the field (see Figure 2). These inhomogeneities hamper MR imaging and spectroscopy of the spinal cord, e.g. by causing geometric distortions in echo-planar imaging, yielding signal dropouts and echo time shifts in T2*-weighted acquisitions, and broadening lines in MR spectroscopy. Thus, it is essential to perform a shim[1] of the magnetic field to minimize such inhomogeneities in the acquisition volume. However, even with a careful adjustment of the standard shim procedure, the image or spectrum quality may be insufficient because not all inhomogeneities can be compensated. [1] Active shimming is performed for every acquisition, and employs dedicated coils in the magnet through which current is passed to generate an extra corrective magnetic field to improve the B0 homogeneity. The standard shimming procedure consists in modelling the magnetic field with spherical harmonic functions of up to the 2nd or 3rd order (based on field maps), then injecting the associated current into each shim coil during acquisition.

Another problem that limits the quality of the shim in the spine region is the variation of B0 with time. The respiratory cycle induces field fluctuations through motion of the chest and therefore fluctuations in local oxygen concentration. The associated MR frequency changes then lead to temporal signal fluctuations, inducing image artifacts such as ghosting and blurring. These variations are of course critical when imaging close to the thoracic region, given that their impact is already substantial in the brain at 7T [4]. The typical signature of these respiratory-induced variations of B0 is a time-dependent voxel shift along the phase-encoding direction with EPI acquisition. The amount of voxel shifts—which can be more than 10 mm—not only varies with time, but also spatially along the spine because it depends on position relative to the lungs [5]. Hence, these displacements are extremely difficult to correct a posteriori, rendering difficult the interpretation of DTI and fMRI data. Respiration-induced change in B0 is one of the main reasons very few people dare to venture into fMRI experiments below the cervical level.

There are, however, solutions to static and dynamic field inhomogeneities. Very-high order external shimming systems have been proposed that can better compensate spatially small-scale variations, thereby producing higher quality spectra in MR spectroscopy [6] and less susceptibility artifacts in EPI data. Such technology has recently been applied to the spine, demonstrating an improvement in field homogeneity in the thoracic spine by more than 30 % [7]. Moreover, since these additional shim coils are controlled independently from the scanner, it is possible to synchronize the current delivered to the shim coil with the respiration of the subject in order to compensate the dynamic variations [4].

Another technology uses RF coils as shim coils, enabling higher order shimming by virtue of the smaller elements and their placement nearer the region of interest as compared to the built-in shim coils of the scanner. Promising results were demonstrated in the brain at 3T [8]. For more information about shimming in the spinal cord see [9].

Ultra high field

Ultra-high field MRI (7T and above, as commonly defined in 2016) is gaining popularity among researchers for its capability to probe even finer details of the anatomy and function. The increased field strength provides a nearly proportional boost in SNR as well as an increased susceptibility contrast that can be exploited for fMRI studies [10] and for susceptibility-based structural imaging such as susceptibility-weighted imaging or susceptibility tensor imaging [11].

Despite the attendant difficulties of ultra-high field systems (greater inhomogeneity, higher specific absorption rate), the first studies of 7T MRI of the spinal cord were quite encouraging. The higher resolution showed finer details of the anatomy, notably exquisite details of the gray matter [12]. Clinical applications of 7T MRI of the spinal cord anatomy include spinal cord injury [13], amyotrophic lateral sclerosis [14] and multiple sclerosis [15]. Recently, 7T has been applied for investigating resting state networks in the spinal cord [16] and more studies are expected to follow from this group and other groups equipped with a 7T system. A review on ultra-high field MRI of the spinal cord is available in [17].

Reduced field-of-view

As mentioned previously, the difficulty in obtaining a homogeneous B0 field directly translates to image distortions along the phase-encoding direction in EPI data. Since the amount of image distortion (i.e., pixel displacement) is directly proportional to the time spent filling the k-space, one strategy is to fill the k-space faster. A straightforward approach is to reduce matrix size (or number of phase-encoding lines) although this comes at the cost of reduced spatial resolution. Another strategy is to decrease the echo spacing (time between two phase lines), which is done by increasing the bandwidth. In addition to these strategies, one can also reduce the field-of-view (FOV) in the image domain, which corresponds to under-sampling the k-space along the phase-encoding direction and therefore image faster (reduction of 2x or more, depending on the acceleration factor). This approach is particularly well-adapted to imaging the spine given its thin and elongated structure, hence yielding the possibility to reduce the FOV along the right-left or antero-posterior directions. However, reconstructing under-sampled k-space data will induce aliasing artifacts on the image. To avoid these, several techniques exist that are discussed below.

Outer volume suppression is one technique for suppressing the region that falls outside the reduced FOV in the phase-encoding direction. This can be achieved by placing saturation bands anteriorly and posteriorly to the FOV; however the cons of this method are that it comes at a cost of higher specific absorption rate (SAR), higher TR and imperfect saturation (slight aliasing can sometimes be seen on images). Another strategy, only applicable to spin echo sequences, is to rotate the 90° excitation plane with respect to the refocusing plane (180° flip angle), so that only spins that fall at the intersection of both planes will produce a spin echo. This technique, called ZOOM (or CO-ZOOM if contiguous slices are used), has shown promises for spinal cord diffusion imaging [18, 19]. An advantage of this technique is that it provides efficient outer volume suppression when combined with additional saturation bands. Cons are that (i) it only works for spin echo sequences, (ii) the profile is not perfectly sharp due to the parallelogram effect at the edge, hence requiring additional saturation bands and reordering of adjacent slices with sufficiently high TR to avoid T1 saturation effects and (iii) it is currently not a product sequence for all popular manufacturers.

2D-RF excitation is a somewhat more direct approach to reduced FOV techniques, where instead of exciting the whole slice only spins that fall within the FOV are excited. Hence, this method does not require additional outer volume suppression. Encouraging results of 2D-RF EPI sequences were demonstrated in the spinal cord for fMRI and DTI studies. Pros are that (i) contiguous slices can be acquired (in comparison with the CO-ZOOM technique), (ii) no signal is visible outside the FOV, hence no aliasing. Cons are that (i) this sequence is currently not widely available in most vendors, and (ii) the excitation pulse needs to be longer than the regular sinc pulse, yielding longer TE. It should be noted however that the introduction of parallel transmit systems in recent clinical product lines will solve the problem of a longer pulse.

Parallel imaging is another technique that consists in using the shared information across receiving coil channels to remove the aliasing. Two popular methods are GRAPPA [20], which reconstructs missing k-space lines, and SENSE [21], which reconstructs the un-aliased image using the prior that each pixel is a linear combination of all coils, weighted by their sensitivity profile. Two notable pros to this approach are that it is available with most MRI systems and compatible with other approaches (e.g., minimize echo spacing, reduce matrix size, etc.). However acceleration factor needs to be carefully chosen based on the receive coil to be used and region to be imaged. More information is available in [3].

Segmented EPI is another technique that consists in acquiring several chunks of sub-sampled k-space and then recombine it before Fourier transform. This technique is not “reduced FOV” per se although it uses the same principles to address susceptibility artifacts by decreasing the effective echo spacing of the acquired k-space. A notable challenge of this technique is to make sure that motion did not occur between chunks (or if so, correct for it using e.g. navigator echoes). This technique has shown promising results with segmentation performed along the phase [22] and readout [23] direction. More information can be found in [24].

Multi-resolution MRI

There is a desire to image brain and spinal cord function simultaneously. Applications include the study of the pain matrix and motor learning. However, optimal sequence parameters are different for both organs. The brain asks for full brain coverage, with isotropic voxel resolution (typically 2x2x2mm3), while the spinal cord asks for highly anisotropic resolution (1x1x5mm3) with axial orientation and small FOV (in order to minimize susceptibility distortions). To overcome this issue, a sequence was developed that allows the user to prescribe two different sets of imaging slices: one for the brain and one for the spinal cord [25]. Promising results were obtained for the study of pain processing with fMRI [26].

Dynamic shimming

As discussed previously in the “Hardware” section, shimming is another way of avoiding susceptibility artifact by directly minimizing B0 inhomogeneities in the first place. While it is difficult to obtain a perfectly homogeneous B0 over a large region using standard 2nd or 3rd order shimming systems—especially when using a large FOV, such as in brain-spine protocols—it is possible to assign specific shim coefficients for each imaging slice. This technique is called dynamic shimming and has shown promising applications in the brain at 7T [27]. Because most fMRI acquisitions of the spinal cord are done axially with thick slices (typically, 3-5 mm), Finsterbusch et al. proposed to only update the z-gradient as a way to compensate for intravoxel dephasing [28] and showed promising results for simultaneous brain-spine fMRI studies [26]. Figure 3 shows an example of image quality improvement when using this approach on T2*-weighted EPI of the spinal cord. Note that dynamic shimming is different than real-time shimming, where shimming coefficient are updated continuously and independently from imaging slices.

Not only is it difficult to acquire good quality mpMRI data for the reasons discussed earlier, but it is also challenging to process them. In contrast to the brain, fewer efforts have been tailored towards image processing software for the spinal cord [29]. Unlike popular software packages for brain neuroimaging (e.g., FSL, SPM, BrainVoyager, FreeSurfer, AFNI, MINC Toolkit), spinal cord researchers have only been developing specific tools (e.g., spinal cord segmentation algorithm) without providing an integrative framework that is compatible with most applications (i.e., adapted to a large variety of mpMRI protocols) and that can reach the community at large (i.e., open source, license-free scripting environment, extensive documentation, support forum, continuous integration service for integrity testing). At the same time, brain software packages are often not suited for spinal cord images because (i) the spine is an articulated structure, therefore standard motion-correction algorithms assuming rigid or affine transformation are inadequate, (ii) brain extraction and segmentation tools are not adapted to the spinal cord because of different shape, contrast-to-noise ratio, etc., (iii) common brain MRI templates [30] and atlases [31] are de facto not usable for spinal cord imaging, (iv) useful features are missing, such as spinal cord cross-sectional area (CSA) measurement. The lack of a standard processing platform has had negative impacts on the spinal cord neuroimaging community and has limited the ability of researchers to compare and reproduce published results, as well as to conduct collaborative and multi-center studies.

To address this problem, the Spinal Cord Toolbox (SCT) [32] has been developed, which is a software package specifically designed to process spinal cord mpMRI data and to perform atlas-based analysis. SCT includes most of the state-of-the-art tools that have already been validated and published (see list of publications below). It is generic (i.e., works for any scanner brand, sequences parameters, modalities or contrasts), ensures cross-platform compatibility, and it is license-free and open source. The primary objective of SCT is to provide a common pre-processing platform, which supplements what is missing from the common brain software package. Hence, the goal of SCT is not to replace entirely what other software already offers (e.g., first- and second-level analysis of fMRI data), but to provide the necessary tools to pre-process mpMRI data, to perform group analysis within standard space, and/or to perform cord-specific quantification such as quantification of cross-sectional area across vertebral levels. The main features of SCT include: Automatic segmentation of the spinal cord [33], automatic segmentation of the spinal canal [34], automatic segmentation of the white and gray matter using multi-atlas algorithm [35], automatic vertebral labeling [36], MRI template of the human spinal cord [37], probabilistic template of white and gray matter [38], atlas of white matter pathways [39], probabilistic map of spinal levels [40], pipeline for registering data with the template [37] and robust motion correction methods for diffusion and functional MRI time series [41]. Figure 4 illustrates the existing templates and atlases available in SCT. SCT can be downloaded at: http://sourceforge.net/projects/spinalcordtoolbox/.

Figure 5 illustrates a typical user-case example. First, the spinal cord is segmented on the anatomical image, then registered to the template. DTI data are then registered to the template using information from the anatomical scan (for greater robustness). Gray matter segmentation could also be performed, provided image contrast and resolution are sufficient. Following registration, several metrics can be calculated, such as the fractional anisotropy in the white matter (or in specific tracts), cross-sectional area, white/gray matter ratio, etc. All these metrics can be automatically calculated at specific vertebral levels, e.g., between C2 to C5 levels, and are corrected for partial volume effect using maximum a posteriori estimation [39].

In this syllabus we reviewed the main challenges associated with spine and spinal cord MRI and presented some of the current strategies to tackle them. While no mention of the plexus was done in the current review, most of these techniques could be equally applied to this region (except for the image processing techniques which are specific to the spinal cord).

Regarding hardware, RF coil arrays are critical for improving sensitivity (higher SNR) and higher acceleration factors (less distortions). Shimming is another exciting avenue for hardware development, with room for improvement in terms of adapted design for the spine, higher order shimming and real-time shimming capabilities. It is hoped that MR vendors will continue setting priority on these hardware aspects. While ultra-high field MRI has lots of potential for fMRI through an increase in SNR and susceptibility contrast, it is also a less “turn-and-key” system compared to lower fields, notably because of the increased SAR, inhomogeneous excitation profile and increased susceptibility artifacts.

Regarding pulse sequences, the techniques that are most useful for imaging the spine include reduced field-of-view approaches (2D RF excitation, outer-volume suppression methods, parallel imaging) and z-shimming. Small additional features can also be mentioned: phase stabilization methods (based on navigator echoes), advanced shimming techniques (e.g., prescribing a shim box inside the spine, FASTMAP), respiratory and cardiac gating, advanced fat saturation methods, etc.

With the emergence of international initiatives for setting up multi-center imaging studies, there is an increasing need for standardizing hardware, acquisition protocols, and image processing methods. Each of these steps is critical and can make a difference between two sites in terms of accuracy and precision. It is hoped that the community will continue working together with manufacturers to converge towards more efficient shimming, RF reception, and mpMRI sequences for the spine and plexus, while ensuring cross-compatibility across vendors.