Diffusion tensor imaging (DTI) is an emerging technique in the work-up of peripheral
nerve pathologies. As an extension of diffusion-weighted imaging (DWI), DTI is a specialized
magnetic resonance imaging (MRI) technique that can provide information about the
microstructural organization and connectivity of peripheral nerves and complement
the information gained from anatomical non–nerve-specific sequences. It is particularly
useful for examining the integrity and orientation of nerve fibers within a nerve.
DTI provides information about the orientation and connectivity of nerve fibers within
a nerve bundle, further explained in this article. By analyzing diffusion tensor data,
it is possible to reconstruct nerve pathways and visualize the three-dimensional (3D)
trajectory of nerve fibers, as in nerve tractography. For example, this helps referring
physicians understand the anatomical relationships and identify the specific nerve
components involved in a tumor or injury. Studies suggest that tractography may be
suitable to identify a “window” from which to approach the tumor resection preoperatively.[1]
More importantly, DTI allows for quantitative evaluation of peripheral nerves by calculating
several important parameters. Fractional anisotropy (FA) exploits the fact that nerves
have a very structured architecture with several intrinsic barriers composed of myelin,
endoneurium, perineurium, and epineurium, facilitating water movement primarily along
the long axis of the nerve and restricting free water diffusion along the short axis
of nerves. Thus FA describes the physiologic anisotropy of nerves, and values range
between 0 and 1, with healthy nerves demonstrating higher FA values and pathologic
nerves having lower values. This reflects the normal close-to-directional diffusion
in healthy nerves, and the abnormal, more isotropic diffusion in pathologic nerves.[2]
Mean diffusivity (MD) reflects the average of three diagonal eigenvectors of the diffusion
tensor, used for apparent diffusion coefficient (ADC) maps.[3] By definition, MD is the average of (1) the direction of the largest eigenvector,
axial diffusivity (AD), and (2) the average of the two smaller eigenvectors, expressed
as radial diffusivity (RD). AD is considered a reflection of axonal integrity, whereas
RD and FA reflect myelin sheath integrity.[4]
Defining normative values of FA, MD/ADC, RD, and AD is challenging due to considerable
variability, even intra-individually or when comparing obtained FA and ADC values
from the scanners of different vendors.[5] There is age-dependent variation in the FA: It tends to decrease with older age,
maybe reflecting a decrease in the number of myelinated fibers in the peripheral nerves
with advancing age.[4] There is also variation in absolute FA numbers between men and women, but the difference
is negligible when corrected for body weight, because FAs are inversely associated
with height, weight, and body mass index.[4] Similarly, an early study of FA and ADC values in median nerves found age- and location-dependent
variation of the values.[6]
Using these principles, DTI has been used to evaluate nerve pathologies in trauma,
tumors, inflammation, degeneration, and physiologic processes, such as nerve regeneration
following nerve injury or surgical interventions. Changes in FA values, AD and RD,
can indicate the progression or lack thereof in nerve healing. Any nerve injury causes
a decrease in FA by disturbing fiber integrity and physiologic anisotropy of the involved
nerve. This is caused by an increase in extracellular water and/or damage in the outer
connective tissue layers of the nerve, depending on injury severity. Changes in RD
provide additional information on the integrity of the nerve sheath structures. In
contrast, increasing FA and decreasing RD values may suggest successful nerve regeneration,
indicating recovery of the microstructural organization of the nerve fibers. A decrease
in RD in regenerating nerves is considered most specific for successful regeneration,
whereas persistently low FA and high RD may suggest the failure of a therapeutic approach.[7] Accordingly, in subclinical ulnar nerve neuropathy from a chronic injury, DTI was
shown to be more sensitive than just relying on T2-weighted imaging alone.[8]
DTI can be particularly useful in evaluating peripheral nerve sheath tumors (PNSTs)
by providing additional information about the tumor and its relationship with the
surrounding nerves by localizing the tumor within/relative to the peripheral nerve
and differentiating it from surrounding tissues ([Fig. 1]). By analyzing the directionality of water diffusion, an analog to nerve microstructure,
DTI can provide information about the microstructural organization of tumors and help
characterize the tumor and understand its extent. It can aid in differentiating between
an infiltrative (neurofibromas or invasive malignant PNSTs) and displacing nature
(schwannoma) of a nerve sheath tumor.
Fig. 1 Tractography in peripheral nerve sheath tumors. (a) Maximum intensity projection of a motion-sensitized driven-equilibrium Cube sequence
in a 58-year-old patient with a hybrid peripheral nerve sheath tumor (schwannomatous
and neurofibromatous differentiation) (arrows) of the left brachial plexus. (b) The tumor mainly involved C5, but partially also C6, and mainly displaces C7 on
tractography images of those three nerves, derived from a multiplexed sensitivity
encoding multi-shot echo planar imaging diffusion tensor imaging sequence, confirmed
during surgery.
Moreover, precise analysis of minimum diffusivity values (MDmin/ADCmin) can offer insight into the invasiveness of a PNST. It can be difficult to differentiate
confidently between benign and malignant PNST by morphological criteria alone. Especially
long-standing chronic (ancient) schwannomas present challenges. These tumors can be
large and have complex features, including cystic degeneration and/or hemorrhagic
areas that mimic features of malignant PNST. Low MDmin/ADCmin values correlate with higher cellularity of the tumor tissue, a common feature of
malignant tumors. Tumor intrinsic MDmin/ADCmin values < 1 (×10−3 mm2/s) were shown to yield a specificity of 94% while maintaining a sensitivity of 100%,
whereas fluorodeoxyglucose positron emission tomography had a lower specificity of
83% when using a maximum standardized uptake value > 3.2.[9]
The imaging information on the potential functional impact and tumor histology that
DTI can offer is important for treatment planning. It provides insights into the potential
functional impact of the tumor on the nerve.[9]
[10]
[Fig. 2] demonstrates an apparent peripheral nerve tumor case in which DTI helped prove a
chronic impingement rather than neoplastic etiology of the nerve lesion.
Fig. 2 Magnetic resonance neurography images of a 43-year-old patient, referred for the
work-up of a suspected peripheral nerve sheath tumor of the right common peroneal
nerve (CPN) detected on an outside magnetic resonance image. (a) Axial short tau inversion recovery image at the level of the tibial tuberosity of
the right leg demonstrates an enlarged CPN (short arrow) at the fibular neck with
mildly enlarged but markedly hyperintense fascicles. At close examination, it was
noted that the myotendinous origin (long thin arrow) of the peroneus longus muscle
was slightly more posterior reaching and more strongly developed than in most patients,
causing long-standing impingement on the CPN at this level and subsequent moderate
neuropathy. The normal tibial nerve (TN) is seen at the same level (arrowhead). (b) Second echo diffusion-weighted double-echo steady-state (DESS) image at the same
level demonstrates the markedly signal-increased CPN (short arrow) as opposed to the
normal TN (arrowhead), indicating neuropathic changes of the CPN. (c) Accordingly, tractography images derived from multiplexed sensitivity encoding multishot
echo planar imaging diffusion tensor imaging demonstrated the enlarged CPN without
disruption of fibers. Although the deep branch of the CPN (thin arrow) was easily
followed in tractography, the superficial branch (short arrow) was more difficult
to trace but appeared normal on images of the anatomical sequences.
DTI has been explored in numerous studies of patients with neuropathies including
autoimmune-mediated, inherited, and systemic entities. To describe just two examples:
In patients with demyelinating type 1 Charcot-Marie-Tooth disease, DTI of the sciatic
nerves showed significant differences between healthy controls and patients, with
decreased FA and increased MD, RD, and AD.[11] A case-control study exploring DTI in diabetic polyneuropathy showed that, as expected,
there was a significant decrease in FA and an increase in ADC of both tibial and common
peroneal nerves in diabetic polyneuropathy patients with an overall moderate, but
significant correlation between DTI and nerve conduction studies. There was a significant
positive correlation between FA and nerve conduction studies and a negative correlation
between ADC and nerve conduction.[12]
When assessing patients with neuropathies or nerve injuries, it is currently still
the norm to assess the physiologic status of the affected nerves with electrophysiologic
examinations. Nerve conduction studies (NCS) and electromyography (EMG) help differentiate
between demyelinating and axonal involvement.[13] In entities such as carpal tunnel syndrome, the diagnosis is typically made with
NCS and EMG. Imaging, including MRI, used to be mainly employed in conjunction to
determine the cause in atypical carpal tunnel cases, although it seems to have gained
more popularity. There may be a certain lag time for electrophysiologic examinations
to detect nerve injuries or changes in nerve regeneration. In addition, electrophysiologic
studies are also more operator dependent than magnetic resonance neurography (MRN)
with DTI.
Although larger scale studies are still needed to prove the value of DTI in the diagnostic
work-up of peripheral nerve pathologies, MRN with DTI has a definite advantage over
electrophysiologic examinations in the diagnostic work-up of plexuses that are in
deeper body locations, and thus more challenging to access reliably for NCS, for example.
A recent smaller scale study in patients with ulnar nerve neuropathy and healthy volunteers
showed that T2 contrast-to-noise ratio in MRN and MD of DTI-based MRN delivered better
results in depicting the precise site of ulnar nerve entrapment,[14] indicating added diagnostic value also in more superficially located nerves.
DTI even seems sensitive to subtle, transient nerve pathologies, such as in neurapraxia,
causing measurable, albeit variable changes in peripheral nerves after compression
with a tourniquet.[15] Similarly, changes induced by subtle stem cell therapy after a crush injury were
detectable in experimental rat model studies. Animals undergoing stem cell therapy
demonstrated measurably higher FA and lower RD in the treatment course, indicating
therapy effects.[16]
Next, we discuss the physics of diffusion MRI, technical considerations, and new developments
as potential solutions to the technical challenges of DTI.
Diffusion Magnetic Resonance Imaging Physics
Diffusion refers to the random motion of molecules resulting from their thermal energy,
which is also called Brownian motion. The fundamental principles to sensitize MRI
signals to the diffusion of water molecules were presented by a group of works in
the 1980s.[17]
[18] This early nuclear magnetic resonance research showed that the diffusion of water
molecules leads to the nonzero phase accumulation on the water spins when a bipolar
gradient waveform is applied. This phase incoherence of spins results in the signal
cancellation and decrease of the net signal.[19]
The random spatial displacement of water molecules by diffusion is usually modeled
to follow a zero-mean Gaussian distribution. The water molecules with fast (or freer)
diffusion have a wider Gaussian distribution function than those with slow (or more
restricted) diffusion. Because molecules naturally displace farther as time passes,
the variance of the Gaussian distribution normalized by the time period during which
spins are dephased and rephased by diffusion encoding gradient waveforms is used to
quantify the diffusion, or diffusion coefficient (in the unit of distance squared
divided by time).
Diffusion Tensor Imaging
Diffusion of water molecules in the human body usually does not occur equally in all
directions because the random molecular motion is bound by the surrounding tissue
structures. DTI[20]
[21] accounts for this directional variation of diffusion, by modeling the diffusion
with a 3D random Gaussian distribution. In the DTI signal model, diffusion is determined
by a symmetrical 3 × 3 diffusion tensor matrix. The elements of the diffusion tensor
matrix are determined by fitting MRI signal measurements with at least six noncollinear
diffusion encoding gradient directions. Then, eigen decomposition of the fitted tensor
matrix is conducted, where the resulting eigenvalues form the AD (largest eigenvalue)
and the RD (mean of the other two eigenvalues), and corresponding eigenvectors indicate
the associated diffusion directions. FA and MD are secondary measures of diffusion
derived from the eigenvalues to indicate the degree of diffusion anisotropy (0 = isotropic;
1 = single directional diffusion) and the averaged diffusion, respectively. The sensitivity
of DTI to tissue microstructure has provided a pathway to probe tissue microstructures
beyond the nominal spatial resolution in the acquired MRI images.
Diffusion Magnetic Resonance Imaging Pulse Sequence
The major workhorse MRI pulse sequence for diffusion-weighted contrast and quantitative
diffusion coefficient mapping is the single-shot spin echo diffusion-weighted echo
planar imaging (DW-EPI) sequence. The DW-EPI sequence has been used for many different
body parts, but its application to peripheral nerves has not yet been widely accepted.
In essence, the many challenges of DW-EPI for peripheral nerves are because these
nerves are quite small, whereas DW-EPI images have a coarse resolution, low signal-to-noise
ratio (SNR), and are fraught with distortions. The following section introduces technical
considerations about sequence parameter tuning and additional techniques for DW-EPI
of peripheral nerves.
Receive Coil Selection
DWI/DTI from the DW-EPI sequence typically suffer from poor SNR, and therefore any
measure to compensate for the low SNR without a substantial increase in scan time
should be adopted. Thus the choice of the optimal receive coil for the target imaging
volume is quite important because these coils can make a tremendous impact on SNR.
A high-channel surface coil array closely fitting the imaging target body part is
the best choice if available. Recently, major vendors released a large blanket-like
flexible coil with a high number of coil elements (GE: AIR Coils; Siemens: Contour
Coils). It has good potential to improve SNR of the brachial plexus where customized
surface coil arrays have not been widely available.
Fat Suppression
Single-shot EPI is highly susceptible to off-resonance-induced distortion artifacts
along the phase encoding direction because the pixel bandwidth along the phase encoding
direction is quite small compared with the resonance frequency of fat. As a result,
fat appears to be largely displaced along the phase encoding direction, and thus fat
signal should be suppressed to avoid a pileup of artifacts on water signals. The spectral
saturation of fat is probably the simplest solution without a major impact on scan
time, but it does not work well in the area with strong off-resonance, such as the
neck or torso. If local shimming to mitigate the off resonance is insufficient, nulling
the fat signal based on the inversion recovery using short tau inversion recovery
can be adopted, although it also partly reduces the water signal, depending on the
T1 relaxation time of tissues. Another solution for fat suppression is to adopt Dixon-based
techniques[22] for running fat-water separation through postprocessing,[23]
[24] although current implementations require longer acquisition and reconstruction times.
The b-Value Selection
The reciprocal value of the diffusion coefficient of the target tissue is a good candidate
for the b-value to detect small changes around the assumed diffusion coefficient.
However, the SNR should also be considered to prevent the DWI/DTI from being dominated
by noise. A wide range of b-values (400–1,400 s/mm2) was reported for diffusion MRI of various peripheral nerves, and there is no consensus
yet about the optimal b-value.[25]
[26] A few studies recommended 600 s/mm2 for a reasonable balance between SNR and sensitivity.[2]
[25] A recent study by Foesleitner et al[27] reported the non-Gaussian behavior of diffusion-weighted signals when they measured
them over 16 b-values ranging from 0 to 1,500 s/mm2. Interestingly, non-Gaussianity appeared beyond b-values of 600 s/mm2 in the axial diffusion direction while it did so for b-values of 800 s/mm2 in the radial diffusion direction, which led to a recommendation of 700 s/mm2 for the b-value. They also demonstrated that for higher b-values, the bi-exponential
and kurtosis model better fit in the measurement than the conventional mono-exponential
signal model. Further investigations are necessary to identify specific nerve diseases
that can benefit from this high b-value kurtosis model.
Diffusion Encoding Direction Selection
The number of diffusion encoding directions is another design parameter. In brain
imaging, the high angular resolution in the axial and radial diffusion directions
of the white matter tracts requires a large number of diffusion encoding directions.
However, in peripheral nerve imaging, except for some pathologic cases such as nerve
sheath tumors, the orientation of the nerves (or axial/radial diffusion directions
of nerves) can also be identified using structural, non–diffusion-weighted images.
Therefore, the number of diffusion encoding directions is rather considered for finding
an acceptable balance between the scan time and conditioning (or accuracy) of diffusion
coefficient estimation. As in the case of b-values, there is no agreement on the optimal
number of diffusion encoding directions, but 15 to 20 were reported to be a good candidate.[2]
[28]
Off-resonance-induced Distortion Correction
There have been localized shimming approaches to correct for the center frequency
offset and linear component of the off-resonance frequency pattern in a slice-by-slice
manner. Real-time B0 inhomogeneity[29] correction applies a different center frequency for each slice after calibration
measurements to correct for the global translation artifact and imperfect fat saturation
due to the center frequency mismatch. Adjustment of linear shimming for individual
slice locations, so-called dynamic shimming,[30] demonstrated the improvement of off-resonance correction in body diffusion MRI.[31]
[32] High-order shimming[33] to correct for beyond zeroth (constant offset) and first-order (linear) components
of the off-resonance pattern has been well adopted in brain diffusion MRI, but it
has not yet been thoroughly investigated in body diffusion MRI or peripheral nerve
diffusion MRI.
Off-resonance-induced signal translation in one spatial direction is proportional
to the ratio of the off-resonance frequency over the pixel bandwidth determined by
the encoding gradient amplitude in the direction. In single-shot EPI, the pixel bandwidth
along the frequency encoding gradient direction is usually larger than the off-resonance
frequency, and thus associated signal translation is insubstantial. The pixel bandwidth
along the phase encoding direction, proportional to phase encoding step size over
echo spacing, is usually much smaller than the off-resonance frequency, resulting
in significant signal translation artifacts along the phase encoding direction. Therefore,
pulse sequence techniques for off-resonance correction in DW-EPI have attempted to
increase the phase encoding step size and/or decrease the echo spacing.
Readout-segmented EPI[34] decreases the echo spacing by using a reduced frequency field of view (FOV) that
consequently needs shorter frequency encoding. A tailored radiofrequency (RF) pulse
to excite signals selectively in both slice and phase directions can allow encoding
a reduced phase FOV, which increases the phase encoding step size and reduces the
off-resonance artifact.[35] Parallel imaging[36]
[37] facilitates undersampling along the phase encoding direction, which leads to the
increased phase encoding step size and thus achieves a reduction of signal translation
artifacts in addition to the accelerated scan time. [Fig. 3] shows a coronal DW-EPI image of the lumbosacral plexus where the combination of
reduced phase FOV and parallel imaging further mitigated the off-resonance-induced
distortion. Development of techniques to resolve shot-to-shot phase incoherence[38] enabled the use of multi-shot EPI for DW-EPI instead of single-shot EPI, where each
shot acquires a segment of EPI undersampled along the phase encoding direction and
as a result, mitigates the associated signal translation artifacts.
Fig. 3 Effects of parallel imaging on single-shot diffusion-weighted echo planar imaging
(DW-EPI). (a) Coronal single-shot DW-EPI with reduced field of view (FOV). (b) Reduced FOV plus parallel imaging. The spatial distortion due to off resonance (red
arrows) is further reduced by combining the two techniques.
Numerous postprocessing methods have been developed for off-resonance artifact correction
of EPI images. Reverse polarity gradient acquisition[39] is arguably one of the most popular methods and has been applied for DW-EPI of a
wide range of body parts.[40]
[41]
[42] The method compares the two non–diffusion-weighted (b0) EPI images, each acquired
with an opposite polarity of the phase-encode gradient and then estimates the off-resonance
frequency for each voxel, exploiting that signal translation occurs in the opposite
direction between two compared images. The estimated off-resonance map is later used
to undo the distortion of subsequentially acquired diffusion-weighted EPI images,
assuming there is no major change in the underlying off-resonance pattern. The calibration
scan with reverse polarity is usually simple to configure in the MRI scanner. The
method works on the image domain requiring no raw k-space data and is supported by
a major postprocessing tool (FSL software).[43]
Denoising
Denoising is probably one of the most well-accepted applications of machine learning,
and the limited SNR issue of diffusion MRI makes it a very suitable application. A
group of reports has been published using various training approaches for denoising
DTI,[44] accelerated multi-shot DWI,[45] and multi b-value DTI.[46] As in the cases of many machine learning techniques, validation with disease cases
should follow to establish the techniques in routine clinics.
Non–Diffusion-weighted Echo Planar Imaging Sequences
There have been developments of non–DW-EPI sequences for acquiring diffusion-weighted
images that attempt to overcome the limited image quality by EPI. Using a single-shot
or multi-shot fast spin-echo readout instead of an EPI readout[47] has demonstrated promise in achieving improved image sharpness with almost no distortion
in imaging of various body parts.[48]
[49]
[50]
The major challenge with fast spin-echo readout approaches is that the bulk phase
incoherence due to motion during diffusion encoding can cause instability in the readout
fast spin-echo train, resulting in severe signal loss artifacts. Diffusion-weighted
magnitude preparation schemes adopted a stabilizer gradient[48] to create a linear phase across the voxel to override a motion-induced phase offset,
and it achieved decent image qualities. However, the linear phase by stabilizer also
causes global attenuation of signal amplitude, and residual phase errors can result
in quantification errors in diffusion estimates. The extra collection of phase navigators
was suggested to account for the residual phase during the reconstruction of multi-shot
FSE data demonstrated improved image quality in diffusion imaging of the lumbosacral
plexus.[51]
Steady-state sequences, such as reverse fast imaging with steady-state free precession
(PSIF) or double-echo steady state (DESS), can be configured to develop diffusion
contrast[52]
[53] and run ADC mapping.[54] With steady-state sequences, it is rather difficult to implement strong diffusion-weighting
corresponding to high b-values (> 800 s/mm2) in DW-EPI. However, steady-state sequence images are almost artifact-free in high
spatial resolution, making them an attractive alternative to DW-EPI in running diffusion
imaging of musculoskeletal tissue. This is because short T2/T2* relaxation times of
musculoskeletal tissues cause severe blurring due to fast signal decay during EPI
readout.
A high-resolution ADC mapping of human sciatic nerve fascicles in patients with Charcot-Marie-Tooth
type 1A was recently presented using DESS at 7-T MRI.[55] The in-plane resolution was 0.15 mm × 0.15 mm with a slice thickness of 2 mm, significantly
higher than the typical spatial resolution adopted in DW-EPI of peripheral nerves
(> 1 mm for in-plane, ∼ 3 mm for slice thickness). The diffusion sensitivity of the
steady-state sequence also develops high susceptibility to motion artifacts. Therefore,
further research on motion artifact suppression techniques for diffusion-weighted
steady-state sequences, such as averaging phase incoherence by oversampling the central
k-space,[56] should be followed for enabling robust diffusion mapping.
