MRI-based myelin water imaging: A technical review



Multiexponential T2 relaxation time measurement in the central nervous system shows a component that originates from water trapped between the lipid bilayers of myelin. This myelin water component is of significant interest as it provides a myelin-specific MRI signal of value in assessing myelin changes in cerebral white matter in vivo. In this article, the various acquisition and analysis strategies proposed to date for myelin water imaging are reviewed and research conducted into their validity and clinical applicability is presented. Comparisons between the imaging methods are made with a discussion regarding potential difficulties and model limitations. Magn Reson Med 73:70–81, 2015. © 2014 Wiley Periodicals, Inc.


Magnetic resonance imaging (MRI) is very sensitive to white matter (WM) diseases such as multiple sclerosis (MS), and has proven to be a valuable technique for the detection of lesions and treatment monitoring in MS. This is due to the excellent sensitivity to inflammation seen in T2-weighted images, which results in MS lesions appearing hyperintense. With the advent of new therapies, early diagnosis and treatment in MS are of great importance; however, this can be hindered by the fact that many other inflammatory and noninflammatory diseases can mimic MS in neuroimaging [1]. Furthermore, T2 hyperintensities within MS are not themselves specific to the underlying pathology. There are a broad range of processes that can lead to similar changes in signal intensity, such as inflammation, demyelination, gliosis, edema, and axonal loss [2]. For this reason, alternative quantitative MRI techniques for imaging myelin are being investigated as a means for achieving greater pathological specificity.

MRI signals primarily originate from hydrogen protons found in water molecules, which have high mobility, leading them to possess moderate to long T2 times (> 10 ms). Hydrogen protons can also be part of macromolecules, such as the proteins and lipids of myelin. These protons are much less mobile and, thus, have much shorter T2 times (10 µs < T2 < 1 ms), resulting in signal decay that occurs too quickly for direct detection using conventional imaging methods. Efforts to image this semisolid component of tissue directly include ultrashort echo time (TE) MRI. Ultrashort TE imaging has been around for some time but has, thus, far garnered relatively little attention in the brain. However, in a recent study, Horch et al. proposed that semisolid signals in nerves arise predominantly from methylene 1H, originating from phospholipid membranes and various intracellular and extracellular (IE) proteins [3]. This, and results from a more recent study by Wilhelm et al. [4], suggests that ultrashort TE imaging might be useful as a method for direct myelin imaging. Research into indirect approaches for myelin imaging has been focused on: diffusion tensor imaging, magnetization transfer (MT) imaging, and multiexponential T2 (MET2) imaging. The latter technique (MET2) has been used to measure water trapped within the myelin layers, which is generally considered an indirect measure of myelin. However, it is argued by some that myelin water imaging is a direct measure of myelin as water is a significant and critical component of the total myelin composition.

As opposed to T2-weighted images, which are acquired at a single TE, quantitative T2 relaxation studies require images to be collected at multiple TEs, resulting in a T2 decay curve. In 1978, Vasilescu et al. [5] observed three distinct components on analyzing T2 relaxation curves in frog sciatic nerve using nuclear magnetic resonance spectroscopy. In order of ascending T2 time, the three components were attributed to water associated with proteins and phospholipds, axoplasmic water, and extracellular water. In 1991, Menon and Allen [6] found four T2 components in the WM of excised cat brain tissue. They subsequently conducted a study in which they identified four T2 components in crayfish nerve cord [7]. In conjunction with optical and electron microscopy of the morphology, the four components were assigned to extra-axonal water protons (T2 = 600 ± 200 ms), axonal water protons (T2 = 200 ± 30 ms), intramyelinic water protons (T2 = 50 ± 20 ms), and lipid protons (T2 = 7 ± 4 ms).

In 1994, in a landmark publication by MacKay et al. [8], analysis of T2 relaxation data obtained in the central nervous system in vivo lead to the identification of two sizeable T2 components (T2 of 10–40 ms and T2 of 70–100 ms). The shorter T2 component was attributed to water trapped between the myelin bilayers and the longer component to the combination of intra/extra-axonal water [9] (otherwise referred to as IE water). A third component with T2 > 1 s was also observed and attributed to cerebral-spinal fluid (CSF) [this component has since been suggested to be an artifact of fitting a long T2 time to non-Gaussian noise associated with magnitude images [10]]. The group also defined the term myelin water fraction (MWF) as the ratio of myelin water to total water. Strong quantitative histopathologic correlations have been observed between the MWF and myelin content in healthy and injured (crushed/cut) rat sciatic nerve [11] and postmortem MS brain samples [12, 13]. These findings have helped validate the use of myelin water imaging as a biomarker for myelin content in tissue.

A review on how MET2-decay curve analysis, based on the method pioneered by MacKay et al. (henceforth referred to as the “reference method”), can provide specific information on brain anatomy and pathology has been published by MacKay et al. [14], with comparisons to other techniques (such as diffusion and MT imaging). Additionally, Laule et al. [15] have published a review article which summarizes five myelin imaging techniques: conventional T1 and T2-weighted imaging, diffusion, MT imaging, spectroscopy, and T2 relaxation (based on the reference method). With this literature in mind, the authors of this article feel that a technical review of the various myelin water imaging methods proposed to date, their strengths and weaknesses, and comparisons with the reference method is warranted. The remainder of this article is devoted to addressing these points.


Reference Method

MacKay et al. [8] first performed MET2 measurements for MWF imaging using a single-slice 32-echo spin-echo (SE) MRI pulse sequence. The sequence (Fig. 1) consisted of a slice selective 90° pulse, followed by a series of 180° pulses for spin refocusing (with 10-ms interecho spacing). To minimize the effects of radio-frequency (RF) field inhomogeneities, nonselective composite 180° pulses, which are less sensitive to changes in math formula and B0 [16] were used. math formula field inhomogeneities cause refocussing pulses to differ from 180°, in turn generating secondary and stimulated echoes [17], which reportedly lead to large errors in T2 estimation [18]. To dephase flow effects and eliminate signal and contributions from stimulated echoes outside the selected slice, slice-select crusher gradient pulses of alternating sign and decreasing amplitude were applied on either side of each 180° pulse [16].

Figure 1.

Multiecho spin-echo sequence: A slice-selective 90° excitation pulse is followed by 180° nonselective refocusing pulses, with a Poon–Henkelman gradient scheme [16].

The signal obtained was described in terms of a MET2 decay [19]:

display math(1)

N is the total number of data points, yi, that are measured at times ti and sj is the unknown amplitude of the spectral component at relaxation time math formula. M is the number of logarithmically spaced T2 times within an appropriately selected range.

Equation (1) has the form of a Laplace transform and solving for sj involves computation of an inverse Laplace transform, a numerically ill-conditioned problem [20]. This means that the result of the inverse Laplace transform is highly sensitive to noise [20]. An established method for carrying out the inverse Laplace transform of a signal exhibiting ME decay is nonnegative least squares (NNLS). NNLS is well suited for relaxation time analysis because it is a stable method that requires no iterative solution and no starting model [19]. The NNLS algorithm finds the set of positive sj which minimize a math formula misfit:

display math(2)

where Aij are kernels for exponential decay, math formula.

Typically, in in vivo experiments, the system is underdetermined ( math formula), further confounding estimation of the T2 distribution. This limitation, as well as the ill-posed nature of the inverse Laplace transform can be partially addressed by using a Tikhonov regularization term [20]. In addition, many investigators [21] believe that T2 decay curves are best described by a smooth T2 distribution, achieved through regularization. In this case, the NNLS algorithm is adapted by minimizing math formula as well as the regularizer, with iterative adjustment of the weight (µ) of the regularization [22]:

display math(3)

Hkj is a matrix representing k additional constraints and fk is the corresponding vector of right-hand side values. The matrix H can be chosen so that it minimizes the energy in the spectrum, the energy in the spectrum derivative, or the curvature. In the presence of noise, regularization results in more robust fits and enforces smooth T2 distributions, better representing those which are found in tissue microstructure [14].

As magnetic resonance (MR) is sensitive to all the water in the central nervous system, a summation over all T2 components provides a relative measure of total water content in a voxel [14]. The MWF can be calculated (Eq. (4)) by dividing the portion of the signal with 15 ms < T2 <50 ms (at 1.5 T) or 15 ms < T2 <40 ms (at 3 T) by total signal in the distribution (the total water content) [23]. The range of the short T2 portion is dependent on field strength due to the apparent shortening of T2 values in brain at higher field strength [23].

display math(4)

Observations of MET2 in healthy controls [24] using the reference method reported MWFs in WM ranging from 8.4 to 15% with an average of 11.3% (Table 1). Evidence supporting the assignment of short T2 components to myelin water includes [8]: in vitro T2 relaxation studies of human WM [36], the identification of a similar component with T2 values between 10 and 20 ms in fresh tissue samples of cat brain [6], crayfish nerve cord [7], and guinea pig spinal cord and brain [37]. MacKay et al. [8] also evaluated MWFs in the WM of MS patients and found them to be significantly reduced compared to normal controls, supporting the hypothesis that demyelination could be measured in vivo using MET2 relaxation mapping in pathology.

Table 1. Average Cerebral WM MWF Values Obtained in Healthy Subjects Using the Various Myelin Water Imaging Techniques
 1.5 T3 T
  1. All studies were performed in vivo unless otherwise indicated by an asterisk. Ranges used to define the short T2 component are indicated by the subscripts:

  2. a

    T2 < 50 ms.

  3. b

    T2 < 40 ms.

  4. c

    T2 < 55 ms.

  5. d

    Data obtained by digitizing plots.

Reference method9 ± 3% [23]a, d12 ± 2% [23]b, d
15.6 ± 8.1% [8]c
11 ± 1% [25]a, d
11.3% [24]a
3D GRASE12.2 ± 1.5% [26]a12.5 ± 4.4% [27]b, d
T2 Prep7.7 ± 0.8% [28]a11.1 ± 1.2% [28]a
12.2 ± 1.4% [26]a11.3 ± 0.2% [29]a
13.4 ± 0.9% [30]a
MGRE6.9% [31]11% [32]*
10.2 ± 2.3% [33]
mcDESPOT29.5 ± 5.3% [34]a 
Linear combination11.8 ± 1.5% [25]a, d 
11 ± 2% [35]

Additional metrics that can be extracted from MET2 relaxation mapping include the geometric mean T2 ( math formula) and the width of the IE water peak (which makes up more than 80% of the water in the normal brain). The math formula is the mean T2 time on a logarithmic scale and is the reciprocal of the geometric mean R2 [ math formula] [38]. This does not hold true for the arithmetic means of T2 and R2, making the geometric mean an appropriate choice for distributions of exponential components 24,38). In [24], math formula across WM structures was found to range from 70.3 ± 1.2 to 82.5 ± 1.2 ms, with the spread in values postulated to be due to variations in water diffusion, cellular morphology, and the composition of the cytoplasmic and extracellular spaces. The IE water peak width was also found to vary across WM structures [24] and differences in peak width were suggested to be indicative of brain structure heterogeneity.

The reproducibility of the technique was assessed by Vavasour et al. [39] at 1.5 T. The authors found myelin water content to have high reliability coefficients [calculated using Cronbach's Alpha [40], a measure designed to produce an overall reliability estimate over all time points] suggesting that, although there was variability across subjects, measurements are consistent over time. Levesque et al. [41] conversely, found moderate reproducibility at 1.5 T by measuring within-subject coefficients of variation (CoVs) during same-day scan/rescan experiments and longitudinal CoVs. More recently, using a three-dimensional (3D) multiecho SE acquisition, Meyers et al. [42] observed low intrasite and intersite CoVs for MWFs at 3 T, indicating that the MWF may be a more robust parameter at 3 T than at 1.5 T. The reader is referred to reference [42] for a table summarizing the various studies evaluating MWF reproducibility.

Overall, one of the most oft cited drawbacks of the reference method is the single-slice imaging capability and long acquisition time ( math formula26 min, see Table 2: protocol A). Levesque et al. [41] proposed an alternative protocol that results in a reduced acquisition time ( math formula4 min, Table 2: protocol B), while maintaining a signal to noise ratio (SNR) that satisfies the minimum criteria suggested for MET2 analysis [the standard deviation of the noise at the shortest TE must not exceed math formula1% of the signal's strength [14, 22]]. MWFs across various WM structures using the second protocol were consistent with those reported by MacKay's group.

Table 2. Protocol Parameters Used for the Various Myelin Water Imaging Sequences
MethodScan time (min)TR (s)ES (ms)Matrix/FOV resolutionSlices/thickness (mm)Other
  1. The original articles which cite the protocol parameters are indicated by the superscripts.

Protocol A ( [8]26315256 × 128/220 (mm) 0.68 × 1.72 (mm2)5 or 102–8 averages
Protocol B ( [30]4210128 × 96/256 × 192 (mm) 2 × 2 (mm2)7no averaging
Protocol C ( [43]25.4***256 × 12854 averages
*TR varied linearly across k-space (TR = 2120–3800 ms)
**ES = first 32 echoes at 10 ms, last 16 echoes at 50 ms
3D GRASE ( [27]14.4110232 × 19220/5slice oversampling factor = 1.3 SENSE = 2 receiver BW = 188 kHz
1 × 1 (mm2)40 slices reconstructed at 2.5 mm slice thickness
T2 Prep ( [28]162.5 128 × 128/240 (mm) 2 × 2 (mm2)10TE = 6, 17, 28, 38, 49, 60, 70, 92, 124, 177, 220, 294 (ms)
MGRE ( [33]8.52 256 × 256/240 (mm)8/4TE1 = 2.1 (ms)
echo spacing = 1.1 (ms)
mcDESPOT ( [34]30  256 × 160 × 118 SPGR: TE/TR 3.1 ms/6.5 (ms)
220 (mm2) × 160 (mm)2° to 18° in 2° increments BW 22.73 (kHz)
1 × 1.4 × 1.4 (mm3)bSSFP: TE/TR 2.3 ms/4.6 ms 6° to 70° in 8° increments BW 62.5 (kHz)
Linear Combination ( [35] * 256 × 128/240 (mm) *TR: 698/725.2/800 (ms)
TE: 8/35.2/110 (ms)
filter weights: 2.3/−3.8/1.5

3D Gradient and Spin-Echo

In 1991, Oshio and Feinberg [44] proposed a combined gradient and SE sequence to accelerate clinical MR acquisitions, known as gradient and spin-echo (GRASE; Fig. 2). Rather than acquire one line of k-space per excitation (as is the case in the reference method), three or more gradient echoes are created per refocusing pulse on either side of the spin echo, allowing one to acquire multiple lines of k-space per excitation. This results in pure T2 weighting in the center of k-space with additional math formula weighting in the periphery of k-space.

Figure 2.

GRASE MRI sequence: A 90° slab-selective excitation pulse is followed by 180° slab-selective refocusing pulses with three gradient echoes created per refocusing pulse.

In 2000, Does and Gore [45] demonstrated the ability to measure MET2 using a GRASE sequence in rat brain. Since then, Prasloski et al. [27] have introduced myelin water imaging based on 3D GRASE, permitting whole cerebrum imaging in less than 15 min scan time at 3 T (Table 2). The 3D GRASE sequence consists of phase encoding lobes along Gz to allow for 3D imaging, a 90° slab-selective excitation pulse, and slab-selective refocusing pulses. Large alternating-descending crusher gradients are applied along Gz before and after each refocusing pulse to remove unwanted signal from outside the selected slab. Data were analyzed using regularized NNLS with concurrent correction for stimulated echo contamination [46]. The latter was achieved through an algorithm proposed by Prasloski et al. that uses Henning's extended phase graph method [47] to account for echo magnitudes generated from different coherence pathways. The extended phase graph method allows one to track the phase coherence of primary, stimulated, and indirect echoes over the course of a multipulse experiment. Previous work by Jones et al. [48] has shown that the inverse problem can be solved: given a decay curve, T2 and refocussing flip angle can be calculated. Building on the work of Jones et al. [48], Lebel and Wilman [49] extended the method to account for imperfect RF slice profiles (a significant source of stimulated echoes), allowing for efficient multislice imaging and improved accuracy in single component T2 analysis. Most recently, Prasloski et al. [46] introduced the use of stimulated echo correction for MET2 analysis and showed greater accuracy of extracted T2 parameters in voxels with poor math formula homogeneity. Finally, in [27], T2 distributions derived from 3D GRASE data were used to generate MWF maps that were found to be in good agreement with those obtained using the reference method [24] (Table 1). k-Space undersampling (using SENSE) and overlapping slices in the GRASE method were not found to cause a significant loss in SNR. However, slight smoothing of the images was identified and attributed to the decreased high-frequency k-space components resulting from the reduced SNR for the gradient echoes. One of the possible limitations of the technique discussed by the authors is the reduced repetition time (1000 ms), which may result in poor quantification of long T2 and T1 components.

T2 Preparation Imaging Methods

Multislice and whole brain MWF imaging has been proposed using T2 preparation (prep) block-based sequences [26, 28, 29]. A T2-prep block begins with a 90° excitation pulse, is followed by a series of 180° refocusing pulses, and ends with a “flip-back” pulse. This leaves the longitudinal magnetization “prepared” with a certain degree of T2 weighting, dictated by the number and spacing of the refocusing pulses in the T2-prep block. Following another excitation pulse, the magnetization can then be sampled. For each repetition, the number of refocusing pulses (during the T2-prep block) is increased. This allows one to sample the decay of the transverse magnetization at different time points. The ensuing measurements can be fit with a MET2 decay model to obtain MWF maps.

The method proposed by Oh et al. [28, 29] involves a T2-prep block followed by a multislice spiral readout. The sequence diagram is shown in Figure 3a. The T2-prep portion of the sequence consists of a nonselective 90° tip-down pulse, followed by composite 180° hard refocusing pulses (6-ms interpulse spacing) with a Malcom Levitt (MLEV) (50) RF cycling pattern and a hard −90° tip-up pulse. After the T2-prep block, large gradient spoilers are used to dephase any remaining transverse magnetization before applying a spectral-spatial pulse and spiral readout. An RF cycling scheme [51] was used to mitigate the effects of T1 recovery between the T2-prep block and the multislice readout. Following a delay time after the multislice spiral readout, chemical shift selective pulses [52] were used to reset all the remaining longitudinal magnetization to zero. Lastly, a recovery time ( math formula) was inserted and held constant regardless of the duration of the T2-prep block. This was done by modifying the duration of the delay time between the multislice spiral readout and the chemical shift selective pulses while maintaining repetition time constant. The TE of the images is the duration of the T2-prep block corrected by the period of T1-weighted signal decay during each composite pulse [53]. The performance of this sequence was evaluated at both 1.5 and 3 T using an eight-channel phased array coil and volume head coil in phantoms and normal volunteers [28]. To keep the scan time within clinically acceptable limits, 12 nonlinearly sampled echoes were chosen (Table 2) and the resulting multiecho data were fit with a distribution of T2 values using regularized NNLS. As the multiple slices are acquired after the T2-prep, the last acquired slice is expected to experience more signal loss due to T1 recovery than the first acquired slice. However, the authors demonstrate that the difference between the signals acquired with and without an inversion pulse in the RF cycling scheme [51] isolates this T1 contamination. To test the validity of using nonlinearly sampled echoes, T2 distributions obtained using the nonlinearly sampled 12-echo data were compared for four WM pixels from four slices with those obtained from data collected using 32 linearly sampled echoes. Both datasets were acquired with the same T2-prep spiral imaging sequence. The mean MWF of the four region of interest (ROIs) was found to be comparable for both sampling methods (8.2 ± 2.1 for the linearly sampled data and 8.5 ± 2.0 for the nonlinearly sampled data). T2 times found for a commercial T2 phantom were within 5 ms of the known T2 values, suggesting that the refocusing trains did not introduce spurious magnetization from stimulated echoes. The phantom data were also used to show that the estimated T2 relaxation times were both accurate and reproducible. T2 distributions obtained from scans in normal volunteers yielded two to three T2 components mainly in WM. MWFs in WM were found to be in the range of 7–12%, in agreement with results obtained in normal controls (7–16%) using the reference method (54; Table 1).

Figure 3.

a: T2-prep multislice spiral sequence with a T2-prep block, followed by a spectral-spatial excitation, 2D spiral readout block, delay time ( math formula),chemical shift selective saturation pulses and a fixed recovery time ( math formula). b: T2-prep 3D spiral sequence with a T2-prep block, 3D spiral readout block, delay time ( math formula), globally selective saturation pulse and spoiler gradients (SAT block), and a fixed recovery time ( math formula).

Nguyen et al. have recently [26] proposed a similar technique: 3D T2-prep spiral gradient echo imaging. As opposed to the T2-prep multislice spiral readout sequence discussed above, the sequence is able to provide whole brain coverage and does not suffer from the decreased SNR associated with 2D acquisitions. The sequence (Fig. 3b) consists of an initial globally selective T2-prep block: a 90° tip-down pulse, a series of hard refocusing pulses and a −90° tip-up pulse. This is immediately followed by a stack-of-spirals (3D SPIRAL) data acquisition consisting of variable density spiral sampling in the kx-ky plane and conventional Fourier sampling in the kz direction [55]. In each data acquisition cycle, a segmented kz-centric view order was used to ensure accurate T2 weighting and minimize math formula induced signal loss. The 3D SPIRAL readout is followed by a variable time delay ( math formula), and a block with a globally selective saturation pulse and spoiler gradients (SAT block). After the SAT block, a fixed-time delay ( math formula) is introduced to ensure that the longitudinal magnetization recovers independently of the duration of the T2-prep module. For a fixed repetition time, this is done by varying the length of math formula. To evaluate the performance of the sequence, human volunteers were scanned at 1.5 T with both 3D GRASE [27] and 3D SPIRAL. Data were analyzed using regularized NNLS and MWF maps were obtained for both sequences with ROI analyses performed for various brain structures. Differences in the MWFs and the IE water math formula of the ROIs between the two sequences were found to be nonsignificant. MWFs and IE water math formulas obtained were generally comparable to those reported previously [54, 56] using the reference method at 1.5 T in healthy brains. SNR was measured in the splenium of the corpus callosum from the image with the shortest TE (5 ms). When comparing both sequences, 3D GRASE was found to have higher SNR. However, the average acquisition time for 3D SPIRAL was shorter than that of 3D GRASE, resulting in similar SNR efficiencies on a per slice basis. Spatial averaging was used to improve data fidelity [57] and was cited to be the reason for which the MWF and IE water T2 values were similar in both techniques, despite the lower SNR for 3D SPIRAL. Intrasubject CoVs for seven subjects were calculated for the chosen ROIs to compare the variability of the MWF measurements due to noise. The CoVs for 3D GRASE maps were found to vary slightly less ( math formula15% lower mean CoVs), which may render 3D SPIRAL a less attractive choice of sequence. The authors presented the following scenario: assuming a healthy corpus callosum MWF of 12% and CoVs of 0.5 and 0.6 for 3D GRASE and 3D SPIRAL, the minimum number of subjects required to detect a 5% change in MWF for a two-tailed t-test with a significance level of 0.05 and a power of 0.90 is 18 for 3D GRASE and 24 for 3D SPIRAL. The authors suggested acquiring data at higher field strength to overcome the SNR deficiency. At high field strength, this sequence is appealing because of lower specific absorption rate compared to 3D GRASE [26]. 3D SPIRAL has been implemented at 3 T [30], however, the results were not compared with alternative MWF imaging methods in the same subject group.

Multigradient Echo

ME analysis of math formula decay using a multigradient echo (MGRE) sequence was proposed by Du et al. [32, 33] as an alternative to overcome the shortcomings of the reference method (single-slice imaging capability and long scan time). In Du's multislice MGRE implementation (Fig. 4), a slice-selective minimum-phase Shinnar-LeRoux excitation pulse [58] of 3.2 ms duration was used with a train of readout gradients of alternating polarity immediately after phase-encoding. The echoes were collected on both the flat-top and the readout gradient ramps and gradient spoilers were applied on all three axes at the end of the readout train to destroy any residual transverse magnetization. The development of this sequence was motivated by the desire to increase the number of acquired echoes before the decay of the myelin water signal, increase volume coverage (while maintaining clinically acceptable scan times), and reduce the specific absorption rate in the acquisition. With the TE1 and echo spacing values typically used (10 ms for both) in the reference method, myelin water signal with a T2 of 15 ms is reduced to 51% at the first echo, 26% at the second echo, and 13.5% at the third echo [32]. Such a reduced signal could compromise the accuracy in the estimation of the MWF and is the reason why a shortened TE1 and echo spacing was sought. With the TE1 and echo spacing used (2.1 and 1.1 ms), myelin signal with a math formula of 10 ms is 81% at the first echo and 24% at the 12th echo. The MGRE sequence is appealing because it only requires one low specific absorption rate excitation pulse (90°) and has improved SNR efficiency, as compared to the reference method, due to the removal of the refocusing pulses (allowing more time to be dedicated to acquiring data during the echo train and rapid decay of the short math formula components).

Figure 4.

MGRE sequence: A 90° slice-selective excitation pulse is followed by a bipolar MGRE readout and spoiler gradients.

Based on findings by Lancaster et al. [59] and Andrews et al. [60], Du et al. [32] affirmed that the T2 of myelin water is not well defined when using the NNLS algorithm. Consequently, the signal from myelinated axon water could be misclassified into myelin water, leading to an overestimation of the MWF [32]. For this reason, a three-pool model of WM [59] was used for ME analysis of math formula decay data. In this model, WM is assumed to be composed of a myelin (my) water pool, a myelinated axon (ma) water pool, and a mixed (mx) water pool (all other water). The authors claim that the three-pool model offers the possibility of investigating compartments other than myelin water, such as myelinated axon water, providing insight into axonal injury and degenerative processes separate from demyelination. The technique was validated with fixed brains in a preliminary study [32] and has recently [33] been applied in vivo. In the latter study, healthy subjects were scanned using an eight-channel phased-array head coil at 3 T (Table 2).

A number of challenges present themselves when using the MGRE sequence in vivo. The respiratory cycle has been shown to induce fluctuations in susceptibility associated with variations in oxygen concentration [61, 62], which can lead to ghosting and blurring in gradient echo images [63]. Provided that increased SNR results in increased sensitivity to such artifacts, this is of particular concern at higher field strengths. Because the slice thickness is typically greater than the in-plane voxel dimensions, local (B0) field gradients along the slice-select direction will be more prominent and contribute to nonexponential signal decay [64, 65]. Low-frequency resting state signal fluctuations observed in the brain [66] may also contribute to nonwhite noise in math formula signal decay [33]. The authors attempted to circumvent these challenges by applying a couple of postprocessing data analysis techniques. Nonexponential signal decay caused by local (B0) field gradients was corrected for by modeling the additional signal decay as a sinc function with a first-order approximation of the linear field gradient [65] prior to ME math formula analysis with a three-pool model [33]. An anisotropic diffusion filter was then applied to images at each time point to improve the SNR and the robustness of the fitting [67, 68]. Five iterations of the filter were used to provide a reasonable trade-off between increased SNR and over-filtering (which leads to smoothly changing intensities being converted to isointense regions), as suggested in [67]. The corrected math formula decay curve was fit with the three-pool model by minimizing the root-mean-squares error between the fit and the signal using a quasi-Newton algorithm for multivariable optimization. To produce accurate fits, math formula ranges for the three pools were needed. These ranges were selected by running repeated test fittings. The math formula ranges which resulted in a minimal fitting error and minimal variation of the MWF were: 3 ms math formula 25 ms math formula 45 ms math formula. When comparing the fitting error of the ME analysis in the in vivo dataset [33] to that obtained in vitro [32], the fitting error in vivo was found to be 3.0 times higher. This suggests that physiological noise was dominant in the experiment. The SNR was increased 2–2.5 fold using anisotropic diffusion filtering, and the authors argue that sufficiently high SNR was achieved using a single acquisition and an eight-channel phased-array coil for MWF mapping. MWF maps obtained with and without local (B0) field gradient correction were compared. It was found that the signal loss due to field inhomogeneities resulted in an overestimation of the MWF values and led to an overall decease in the contrast between WM and gray matter. Correction for field inhomogeneities can reduce bias in the MWF maps, however this failed in areas of severe inhomogeneities. In addition to the unusually high MWF values observed due to field inhomogeneities, high MWF values were also observed in veins and the globus pallidus. These were attributed to increased amounts of deoxyhemoglobin in venous blood [69] and nonheme iron in certain brain regions [70], which leads to math formula shortening. The use of susceptibility-weighted imaging [69] has been suggested as a possible means for identifying these structures and subsequently excluding these regions from further math formula decay analysis. The mean MWF was measured in a number of WM regions and the values were reported to be in accordance with those found in the same regions by other groups using the reference method [25, 54] at 1.5 T and a multislice T2 prep spiral acquisition [28] at 1.5 and 3 T (Table 1).

MGRE has since been extended to 3D imaging by Lenz et al. [31] and used to acquire math formula relaxation data at 1.5 T in healthy subjects and patients with MS. The group used sinc-shaped slice-selective excitation pulses of 700 µs duration, 96 echoes with a first TE of 1.4 ms and an echo spacing of 0.99 ms for a total scan time of 9 min 40 s. math formula distributions were obtained using regularized NNLS. Two isolated smooth peaks were seen in WM voxels (corresponding to the myelin and IE water pools) and the MWF was calculated as the ratio of the signal with math formula < 25 ms to the total signal. Due to the lack of distinction of three isolated peaks, the authors deemed the application of the fixed three-pool model, as suggested by Du et al. [32, 33], to be unfeasible. When comparing healthy subjects to MS patients, no significant difference was seen in MWF values for various WM structures [31]. The authors suggest this may be due to too few healthy subjects (three) or MS patients with a high amount of normal appearing white matter. The average MWF for all subjects across various WM structures was calculated and found to be 6.9%. Although the average MWF appears to be low compared to values obtained using the reference method [8, 24], the authors point out that some studies [23, 28] have found lower MWF values at 1.5 vs. 3 T. It was suggested that the higher SNR at 3 T may improve the detectability of the short components, and thus, account for the higher MWF values observed at higher fields.

Another math formula-based approach for myelin water imaging was recently proposed by Oh et al. [71]. The technique exploits the presumed ME nature of T1, as evidenced in [9, 72, 73], by suppressing long T1 signals with a double inversion RF pair, leaving short T1 signals to be sampled with a gradient echo readout. The T1-prepared image is then normalized by a single-echo gradient echo image. The method is referred to as direct visualization of short transverse relaxation time component. Visualization of short transverse relaxation time component images were shown to be successful in delineating MS lesions, indicating potential clinical usefulness. However, T1-weighting confounds accurate quantification [71].


Another important technique for measuring the MWF, which does not rely on measurement of the T2 decay curve, is that of multicomponent driven-equilibrium single-pulse observation of T1 and T2 (mcDESPOT) [34]. mcDESPOT is based on an extension of the driven-equilibrium single-pulse observation of T1 (DESPOT1) and driven-equilibrium single-pulse observation of T2 (DESPOT2) techniques [74]. DESPOT1 and DESPOT2 are themselves based on spoiled gradient recalled echo (SPGR) and balanced steady-state free precession (bSSFP), respectively. SPGR-based T1 estimation was originally introduced in 1974 [75]. The method involves the combination of SPGR images, each of which establishes their own steady-state, to generate a T1-dependent signal curve as a function of flip angle. The bSSFP pulse sequence has also been around for some time [76]. In this sequence both longitudinal and transverse magnetization are brought into dynamic equilibrium through the application of α pulses and fully refocusing the transverse magnetization prior to each excitation pulse. Collecting bSSFP images over a range of flip angles yields a signal curve that is a function of both T1 and T2 [77]. If one obtains T1 from the SPGR sequence, the bSSFP signal curve can be used to derive T2.

The mcDESPOT technique involves fitting the data obtained from the DESPOT1 and DESPOT2 sequences with a two-pool model of longitudinal and transverse relaxation that includes intercompartemental water exchange (Fig. 5). Two-component signal models have been described previously by a number of groups [34, 78-81]. The model considers brain tissue to be composed of two water compartments, free IE water and water trapped between the lipid bilayers of the myelin sheath, in exchange with one another. The technique offers the advantage of whole brain ME T1 and T2 quantification in 16 to 30 min (Table 2) with high SNR efficiency. Furthermore, inclusion of intercompartmental water exchange between the two pools is unique to mcDESPOT, with regards to alternative ME relaxation methods. This makes the technique particularly appealing for two reasons. First, the mean residence time of myelin-associated water has been found to be a potential measure of myelin thickness [82, 83]. Second, the presence of intercompartmental water exchange may impact MWF quantification [84].

Figure 5.

The two-pool model: fF and fM are the free and myelin water fractions math formula and math formula are the T1 times of the free and myelin water pools, math formula and math formula are the T2 times of the free and myelin water pools, and τ is the mean residence time of the free and myelin water pools.

The average WM MWF derived from the mcDESPOT model was compared to average WM MWFs obtained with alternative techniques at 1.5 T (Table 1) and found to be significantly greater (α = 0.05). Only results that were quoted with a standard deviation in their respective publications were used in the comparison, and results from [35] were omitted as only one subject was used in the study. Zhang et al. [85] compared MWFs obtained with the 3D GRASE approach to those obtained with mcDESPOT, in the same subject group, and also found mcDESPOT-derived values to be significantly greater, by a factor of approximately 3. These results persisted when omitting intercompartmental water exchange in the mcDESPOT model. Another possible explanation for the greater MWFs observed with mcDESPOT is that of MT effects, which are known to contribute to bSSFP [86, 87] and SPGR [88] signals. To assess the role of MT in mcDESPOT results, Zhang et al. [89] obtained mcDESPOT measurements using longer, lower amplitude RF pulses, which should lead to reduced MT effects (compared to the short, conventionally used, RF pulses). The results indicated that MT effects did have an impact on the mcDESPOT-derived MWFs. Lastly, Zhang et al. [90] also attempted to fit data obtained with the mcDESPOT technique with a MET2 relaxation model using NNLS. The resulting T2 distributions revealed three distinct pools and the corresponding MWFs were significantly lower than those obtained using the two-pool model for mcDESPOT analysis.

Recently, the mcDESPOT model was extended to include a third, slow-relaxing, and nonexchanging bulk free water pool (e.g., CSF) [91]. It is thought that in voxels containing tissue and CSF, such as voxels containing tissue and ventricle or meninges, the two-pool model may result in MWF underestimation. The assumption that the third pool is nonexchanging is justified by the fact that ventricular and meningeal water would need to cross the blood-CSF barrier [92]to mix with the other two pools. However, this assumption is not likely to hold in certain pathologies, such as in the case of transient ischemic attacks [93].

Fitting to the mcDESPOT model is based on a genetic [34] or stochastic region contraction approach [91]. To assess the accuracy and precision of the three-pool model fitting, Deoni et al. [91] ran numerical simulations over a range of experimental conditions. As predicted by the group, the findings demonstrated that when the bulk free water compartment size was increased, the two pool MWF value became increasingly underestimated. Conversely, the three-pool model resulted in no MWF underestimation. In vivo results showed that MWF differences in the two-pool and three-pool results were greatest in the partial volume voxels. Results were comparable for the two models in voxels pertaining to homogeneous WM and gray matter. The solution uniqueness was also investigated [91] by plotting histograms showing the distribution of MWF and free-water fractions for 50,000 simulation realizations for a given parameter set. The resulting histograms were found to be Gaussian distributed, indicating that the algorithm converges to the same solution repeatedly.

Recently, work presented by Lankford et al. [94] has suggested mcDESPOT signals are not able to provide parameter estimates with useful levels of precision. The precision of the two-component mcDESPOT model parameters was determined by calculating the Cramer Rao lower bounds (CRLBs) of their variances. CRLBs were calculated for a number of tissue models using constrained and unconstrained fitting analysis. CRLBs were found to be very high at feasible SNRs when the transverse relaxation and water exchange rates were constrained to their correct model values. However, this increase in precision comes at the cost of increased parameter estimate bias. Monte Carlo simulations were run to validate the CRLB estimates in both the constrained and unconstrained cases. The simulations provided parameter variances that differed by up to 8% with respect to the CRLBs. In the context of these results, the authors concluded that parameter estimates obtained using the mcDESPOT model cannot provide useful levels of precision. This can be improved by constraining certain parameters, such as the transverse relaxation rates and water exchange. However, although the transverse relaxation rates can be assumed to be relatively invariant across tissues in healthy subjects, the authors suggested that water exchange cannot [94].

The work of Lankford et al. appears to be in disagreement with the reproducibility of mcDESPOT results published by Deoni et al. in a number of studies [91, 95-97]. Reproducibility of mcDESPOT-derived MWF maps obtained with the two-pool model was first assessed in [96]. Volunteers were scanned on repeat occasions in one center, and then at two different centers to determine the intrasite and intersite reproducibility. Results showed high reproducibility longitudinally and across different imaging centers. A similar study was repeated in [95], wherein five consecutive scans were acquired in one scan session to assess intrascan reproducibility using the bootstrapping method [67]. In addition, five scans were acquired on successive days to measure interscan reproducibility. CoV maps were generated in both cases and averaged across five ROIs, showing high intrascan and interscan reproducibility. Lastly, and most recently, in a study aimed at assessing the efficacy of mcDESPOT at obtaining high resolution cervical spinal cord MWF images [97], Kolind et al. found that the MWF values calculated were highly reproducible between subjects and over time. Deoni et al. [91] also measured the fitting algorithm stability and interscan and subject reproducibility when using the three-pool model. The stability of the fitting algorithm was evaluated by repeat fitting on a given dataset and subsequently measuring the CoV. In all three cases CoVs were low.

Discrepancies between results obtained by Lankford et al. [94] and those obtained by Deoni et al. [91, 95-97] were suggested [94] to be attributable to inadvertent constraint of the solution to a value other than the global math formula minimum in the mcDESPOT analysis. Alternatively, additional effects, such as MT, which are not accounted for the mcDESPOT model, may result in significantly reduced variance in the estimates of the myelin water-associated signal in the two-pool model [94].

Linear Combination Myelin Imaging

The idea of using coefficients to linearly combine images so that desired tissues are filtered out was first introduced by Macovski in 1982 [98, 99]. The method described involved a matrix inversion technique that can be used to linearly combine multiecho data to select for short T2 components. The T2 filter is defined as math formula, where ci are the weighting coefficients and N is the number of echo images. The technique, however, did not allow for constraints on the coefficients nor on the filter, the lack of which can lead to the misinterpretation of short T2 components [25]. Similarly, in 1991, Whittall et al. [100] used the Backus-Gilbert technique [101] to calculate the set of coefficients needed to linearly combine multiecho data for short T2 component selection. Both of the above mentioned techniques make use of linear algorithms for determining the best set of coefficients.

Jones et al. argue [25] that a nonlinear algorithm would allow for greater flexibility on filter constraints. Such a technique was used in a study [25] which involved the acquisition of 32 SE images at 1.5 T. Two sets of coefficients were calculated, one that selected water with short T2 (10 ms < T2 < 50 ms) and one for water with T2 > 10 ms (uniform T2 selection image). The MWF was defined as the ratio of the short T2 selection image to the uniform T2 selection image. Simulations revealed good agreement between the linear combination and the traditional NNLS method for calculating the MWF as well as with the true values. MWF maps were generated for five healthy volunteers using both methods and the mean MWF was calculated for various WM and gray matter regions. Mean MWFs calculated using the linear combination technique were found to be within 5% of those calculated using NNLS. The authors deemed the linear combination method to possess the advantage of greatly reducing processing time as well as not relying on assumptions about underlying models [25].

More recently, a three-echo linear combination myelin imaging method was proposed by Vidarsson et al. [35] with the benefit of reduced imaging time and providing multislice coverage. In that work, three echoes were acquired for six slices using three SE sequences in 5 min scan time (see Table 2). The TEs were optimized so that a 50-fold suppression of IE water could be achieved, along with a 10-fold suppression of CSF. The filter coefficients used for generating a short T2-weighted image were obtained by means of an algorithm that maximizes the SNR of short T2 components. The algorithm requires predefined T2 bands for myelin water (set to be 10–20 ms), IE water (75–85 ms), and CSF (200 ms–5 s). A uniform T2-profile filter was also computed and, as in [25], the ratio of the short T2-weighted image to the uniform T2-weighted image was used to produce a MWF map. The filter design was validated using phantom scans with known T2 and T1, and a healthy volunteer was scanned at 1.5 T. The mean MWF was calculated for a number of WM and gray matter regions and found to be in accordance with literature values (see Table 1). For comparison, 4, 5, and 32 echo filters were also designed using the same method. The SNR efficiency of the three-echo filter was found to be roughly the same as the 4, 5, and 32 echo filters, with the 32-echo filter achieving at most 5% better SNR efficiency than the four-echo filter. One of the drawbacks of the three-echo linear combination method discussed by the authors is the possibility of variations in the T2 of WM and its potential impact on component suppression and short T2-estimate accuracy. To increase the accuracy of short T2 estimates, it is suggested that the IE water suppression band be widened using a four-echo filter. A wider IE water suppression band has the benefit of better eliminating interfering IE water and providing increased short T2-estimate accuracy, but it comes at the cost of reduced SNR efficiency.


The main drawbacks of the reference method have been addressed by the alternative myelin water imaging techniques presented in this review. All of the sequences discussed offer the possibility of acquiring images with multislice or whole brain coverage, that are sensitive to myelin water, in 30 min or less. However, with the exception of mcDESPOT, none of the methods model the effects of water exchange between the microanatomical compartments. We feel that this is an important issue to be resolved, and that it merits further discussion here.

Exchange has generally been thought to take place on timescales longer than that of MET2 experiments [24, 102], thus having a negligible effect on MWF measures. Kalantari et al. [103] used an analytic solution to the Bloch equation describing the magnetization evolution of a four-pool model [102, 104] to in vivo human MRI data, and found that the corrections that should be applied to MWFs, for different T1 relaxation scenarios, are less than 15%. However, recent simulations of varying exchange rates' impact on MWFs suggest otherwise. Simulations of MET2 relaxation mapping were performed by Levesque and Pike [84], with pathology-inspired modifications on the water exchange rate, to gauge their effects on MET2 observations. Motivation for performing these simulations included findings in fresh bovine WM [105] that exchange rates are temperature dependent, suggesting that exchange is largely mediated by diffusion. Levesque and Pike proposed that since pathology such as MS leads to changes in diffusion, this may consequently affect water exchange, and subsequently found that MWF estimates, and the mean T2 values for the fast and the slow relaxing components decreased with increasing exchange. It was, therefore, concluded that pathology affecting water exchange can result in variations in the MWF that could be incorrectly interpreted as changes in myelin content. More recently, Harkins et al. [83] performed MET2 experiments in vivo within the rat spinal cord and found a wide variation in the MWFs of spinal cord tracts with similar myelin content. A numerical simulation based on segmented histology images was used to quantitatively account for T2 variations between the tracts. The model indicated that differences in water exchange between the tracts may account for the observed differences in MWFs. Water in smaller compartments (such as smaller axons and thinner myelin) was predicted to display a faster exchange rate. Therefore, if the exchange rate depends on the compartment dimensions, the slow exchange model, on which MET2 imaging is based, may be more suitable in some myelinated tissues than others [82]. In [106], relaxation exchange spectroscopy was used to provide direct observations of intercompartmental water exchange in freshly excised rat optic nerve and frog sciatic nerve. The myelin water residence time (τm) measured in optic nerve samples was 138 ± 15 ms and 2046 ± 40 ms in sciatic nerve samples. The higher τm in sciatic nerve was hypothesized to be due to its thicker myelin sheaths. The authors report that extrapolation of these results to in vivo temperature supports previous claims that MWFs are significantly affected by water exchange. While potentially undermining a simple interpretation of MWF as being a quantitative measure of myelin content, these findings suggest that a comprehensive analysis of water T2 spectra based on MET2 studies could reveal microstructural information related to myelin thickness and axonal diameter, offering the potential for the development of novel MRI methods for myelin measurement and characterization [106]. It should be noted that myelin content in cerebral WM can be expected to be different from that found in the tissues mentioned above, as would be the exchange properties of myelin water. Therefore, future studies should include investigation into the exchange properties of water in cerebral WM.

Another caveat of myelin water imaging, is that myelin debris can be interpreted as intact myelin [11, 107, 108]. Webb et al. [11] performed MET2 relaxation and quantitative histomorphometry on normal and injured rat sciatic nerve, and found that the size of the short T2 component reflected the process of myelin loss and remyelination over the course of 6 weeks following trauma. However, histopathology indicated that the short T2 component did not distinguish between intact myelin and myelin debris. Kozlowski et al. [107] acquired MET2 data and histological data using luxol fast blue staining in control and injured rat spinal cords 3 weeks post injury. The authors found that MWFs and luxol fast blue staining had increased at the site of injury, and attributed these findings to the presence of myelin debris, whose clearance is known to be very slow in the central nervous system [109]. Finally, McCreary et al. [108] detected a significant decrease in mice spinal cord MWFs 14 days post lysolecithin injection and a return to control levels after 28 days, corresponding to clearance of myelin debris and remyelination (as shown by eriochrome cyanine and oil red O staining of histological sections). However, the authors observed that, although the changes in MWFs paralleled histological changes, MWF changes were delayed by 1 week with respect to the latter. This was hypothesized to be due to the presence of degraded myelin remaining in the lesion, as supported by the large amount of red oil O staining 7 days post injury [108].


Myelin water imaging holds promise as a means for assessing changes in myelination. It offers improved specificity as compared to traditional means for detecting WM changes, such as monoexponential T2 contrast, and has been shown to be a reliable marker for myelin. A number of myelin water imaging techniques are presented in this review, ranging from slow single-slice methods on which the field has been established to whole brain imaging in clinically acceptable scan times. However, various questions need to be addressed before these techniques can be reliably applied and interpreted in a clinical setting. Future work should also include assessment of the reproducibility of the techniques. Furthermore, when comparing MWFs, it would be beneficial to perform measurements with both the reference method and the sequence in question, in the same subject group. Care should also be taken when selecting the myelin water range cutoff, as it is not always consistently chosen amongst various groups and past work has clearly shown that it can significantly impact MWF estimates [41]. Lastly, a discrete component model, as was applied by Du et al. [32] in the analysis of MGRE data, warrants further investigation, as does the exchange of water between microanatomical compartments in cerebral WM.