|Home | About | Journals | Submit | Contact Us | Français|
The image contrast in magnetic resonance imaging (MRI) is highly sensitive to several mechanisms that are modulated by the properties of the tissue environment. The degree and type of contrast weighting may be viewed as image filters that accentuate specific tissue properties. Maps of quantitative measures of these mechanisms, akin to microstructural/environmental-specific tissue stains, may be generated to characterize the MRI and physiological properties of biological tissues. In this article, three quantitative MRI (qMRI) methods for characterizing white matter (WM) microstructural properties are reviewed. All of these measures measure complementary aspects of how water interacts with the tissue environment. Diffusion MRI, including diffusion tensor imaging, characterizes the diffusion of water in the tissues and is sensitive to the microstructural density, spacing, and orientational organization of tissue membranes, including myelin. Magnetization transfer imaging characterizes the amount and degree of magnetization exchange between free water and macromolecules like proteins found in the myelin bilayers. Relaxometry measures the MRI relaxation constants T1 and T2, which in WM have a component associated with the water trapped in the myelin bilayers. The conduction of signals between distant brain regions occurs primarily through myelinated WM tracts; thus, these methods are potential indicators of pathology and structural connectivity in the brain. This article provides an overview of the qMRI stain mechanisms, acquisition and analysis strategies, and applications for these qMRI stains.
Brain function requires efficient and effective communications between different brain regions and between the brain and body. Recent developments in magnetic resonance imaging (MRI), electro-encephalography (EEG), and magneto-encephalography (MEG) methods have enabled researchers to study the brain as a collection of networks rather than isolated regions. The applications of functional connectivity mapping using resting blood-oxygen-level-dependent (BOLD) functional MRI (fMRI) for mapping functional brain networks are the most rapidly growing neuroimaging methods. These functional networks may be modulated by the white matter (WM) substrates that conduct the electrical signal between different regions of the central nervous system (CNS). In this article we will review several methods that are being used to characterize WM properties.
WM primarily consists of densely bundled nerve fibers, each comprising an axon extending from the neuronal cell body with a long, narrow, cylindrical geometry and surrounded by a myelin sheath. The main role of axons is to conduct electrical signals from the cell body to other neurons. This signal conduction is enhanced by the myelin sheath, which consists of concentric layers of lipids and proteins that insulate the axon while also providing mechanical and biochemical support. The axon at its distal end interfaces with other cells (primarily other neurons) and transmits signals by way of chemical neurotransmitters across synaptic junctions. Other cell types in the WM include the oligodendrocytes that synthesize and maintain myelin as well as astrocytes (biochemical support) and other support glia. Injured or inflamed CNS WM may also contain microglia and macrophages.
Noninvasive imaging techniques that can characterize WM tissue properties and reveal changes induced by healthy development and aging, genetics, disease, and injury are potentially valuable. Ideally, quantitative imaging measures that are indicators of specific changes—for example, myelination, axonal changes, gliosis, inflammation—would be useful for a broad range of clinical and research applications and would improve our understanding of how WM properties influence the function of brain networks. A major challenge is that the spatial resolution of quantitative MRI (qMRI) is currently on the order of millimeters, whereas axonal diameters range from 1 to 20μm and the thickness of the healthy myelin sheath is on the order of 1μm. Consequently, MRI can provide only a macroscopic picture of contrast weighting mechanisms that are sensitive and ideally specific to the WM microstructural properties of interest.
In this article, we will review three qMRI techniques—diffusion MRI (including diffusion tensor imaging [DTI]), magnetization transfer (MT), and multicomponent relaxometry—for characterizing WM properties with an emphasis on myelin. These qMRI methods generate image maps with individual signal contrast or stains, which are potential biomarkers of myelin- and axon-related changes in WM. Each of these stains relies on unique mechanisms influenced by how water in tissue interacts with its environment at the molecular, cellular, and/or tissue microstructural level. For each stain, we will review the underlying mechanisms, methods for measurement, relative merits, and key drawbacks. We will also briefly discuss other potential MRI methods for characterizing WM. Finally, we will attempt to summarize the current state of the field and make some recommendations regarding potential directions of research and development in this area.
Currently, the most widely used imaging method to study and characterize WM in a broad range of diseases and disorders is DTI, which is highly sensitive to variations in the tissue microstructure. Diffusion MRI is sensitive to the random motion of water molecules in a medium (Le Bihan et al., 1991a). The image signal in a diffusion-weighted imaging (DWI) study is modulated by the presence, orientation, and density of membranes and other barriers in the tissue (Fig. 1). As the density of the barriers increases or the spacing between barriers decreases, the water diffusion will be more hindered (Beaulieu and Allen, 1994; Norris et al., 1994). Diffusion measurements that are modulated by the barriers in tissue are often referred to as apparent diffusion coefficients. In fibrous tissues like WM and skeletal muscle, the density of the barriers is much higher in the direction perpendicular to the fibers; thus, the apparent diffusivity is much higher in the direction parallel to the fiber bundles than in the perpendicular orientation (Chenevert et al., 1990; Moseley et al., 1991). Diffusion MRI appears to be highly sensitive to the effects of tissue cellularity, cellular swelling, axonal injury and loss, myelination, edema, necrosis, and inflammation [see Alexander et al. (2007) for review]. DWI and DTI pulse sequences and simple analysis tools are now available on all modern clinical scanner platforms.
The basis for nearly all DWI and DTI clinical and research studies is a pulsed-gradient, spin-echo, echo-planar imaging (EPI) sequence (Le Bihan et al., 1991b), where the diffusion-encoding gradients bracket the refocusing (RF) pulse(s). The basic signal model for DWI is
where So is the signal without diffusion-weighting gradients (b=0), D is the apparent diffusion coefficient, γ is the gyromagnetic ratio, G is the diffusion gradient amplitude, δ is the width of the diffusion gradient pulses, and Δ is the time between diffusion gradients. Since the maximum diffusion gradient on clinical MRI scanners is limited (currently 4–8G/cm), the width of the gradients has to be long to achieve high diffusion-weighting.
An elegant MRI method for characterizing the anisotropic diffusion of water in biological tissues is DTI, which was originally described by Peter Basser and colleagues at NIH in 1994 (Basser et al., 1994). The primary assumption of DTI is that the probability density of displacements from diffusion is a three-dimensional (3D) multivariate Gaussian distribution, where the diffusion tensor is the covariance matrix of diffusion displacements
The Gaussian diffusion distribution described by the diffusion tensor may be represented by a 3D ellipsoid with the lengths and orientations of the major, medium, and minor axes corresponding to the eigenvalues and eigenvectors of the diffusion tensor.
There are many factors associated with the image acquisition of DTI data that will influence the measurements, including the pulse sequence design and parameters (Le Bihan et al., 1991b) (i.e., single vs. dual echo, parallel imaging, repetition time [TR], and echo time [TE]), DTI encoding parameters (diffusion-weighting and encoding directions), factors that influence the signal-to-noise ratio (SNR) such as the B0 magnetic field strength, gradient performance and coil sensitivity, and artifacts including pulsatile (e.g., cardiac) signal fluctuations, ghosting, and distortions from B0 field inhomogeneities and eddy currents; see Alexander et al. (2007) and Tournier et al. (2011) for reviews. For most research applications in the human brain, it is desirable to obtain isotropic spatial resolution. At 1.5T and 3.0T, target resolutions of 2.5mm and 2.0mm, respectively, are commonly used and achievable. A DTI protocol with 30–40 encoding directions at b=1000sec/mm2 and 4–6 averages of b=0 at a resolution of 2mm at 3.0T will achieve good quality DTI maps in many cases though more averaging will improve the measurement accuracy and variance.
The processing of DWI data requires careful inspection of the DW images, correction of artifacts, and tools for obtaining regionally specific measures. If rotation or shear is detected in a DW image volume, ideally the gradient direction should also be adjusted to account for differences in the encoding frame. It has been reported that corrections to the encoding directions after motion and eddy current distortion correction can lead to significant changes in the estimated diffusion tensor measures (Leemans and Jones, 2009). For DTI studies, the diffusion tensor is often estimated using a linear least squares or weighted least squares approach, though a nonlinear least squares approach has been shown to yield the most robust fit in noisy data (Jones and Basser, 2004; Koay et al., 2006). In the case where artifactual noise may be present, robust estimators may be used to identify and either remove or minimize the impact of outlier measurements (RESTORE) (Chang et al., 2005).
One of the greatest confounds in DWI is partial volume averaging between tissue components. The relatively large voxel sizes used in most DWI studies can lead to multiple tissue types (e.g., cerebral spinal fluid [CSF], WM, and gray matter [GM]) within a voxel (Alexander et al., 2001). Further, disease processes like edema and infiltrating tumor can make it difficult to accurately characterize the underlying WM fiber properties. Also, regions of crossing or highly divergent WM fibers will affect the diffusion tensor eigenvalues and diffusion anisotropy measures. In fact in regions of crossing WM tracts, the simple diffusion tensor model is not adequate for characterizing the tissue microstructure (Alexander et al., 2001).
Diagnonalization of the tensor gives us the eigenvalues (λ1, λ2, and λ3 in decreasing order of magnitude) and the corresponding eigenvectors (e1, e2, and e3). Many different stains may be derived from the eigenvalues, which are invariant to the tensor orientation [e.g., Alexander et al. (2000); Basser (1997); Ennis and Kindlmann (2006); Westin et al. (2002)]. How the eigenvalues are combined and weighted can be viewed as different filters for accentuating specific features of the water diffusion. Commonly used DTI measures include the mean diffusivity (MD) and the fractional anisotropy (FA) (Basser and Pierpaoli, 1996; Pierpaoli and Basser, 1996) (also see Fig. 2):
MD and FA describe complementary information about the diffusion of water. The specific interpretation of all diffusion measures needs to be kept within the context of the diffusion mechanism, which is the modulation of the diffusion coefficient of free water by tissue membranes. FA is a normalized standard deviation of the eigenvalues and is commonly referred to as a summary measure of microstructural integrity. While FA is highly sensitive to microstructural changes, it is not very specific to the type of change and it is highly advisable to also include other DTI measures in any analysis. At a minimum, studies should include the MD, which is the directionally averaged, inverse measure of the membrane density and fluid viscosity and is very similar for both GM and WM, particularly at b~1000sec/mm2. MD is sensitive to cellularity, edema, and necrosis. It is important to understand that noise in the measured signals can lead to overestimates of the diffusion anisotropy, particularly in more isotropic regions (Pierpaoli and Basser, 1996). Thus, SNR is an important factor to consider in the interpretation of noisy DTI data.
The apparent diffusivities in the directions parallel and perpendicular to the WM tracts are the axial and radial diffusivities, DA and DR, respectively, which provide more direct measures of the microstructural dimensions (Fig. 2):
It has been shown in animal models that DR increases in WM with de- or dys-myelination (Song et al., 2002; Wu et al., 2011a). This observation appears to be consistent with many WM pathologies. Changes in the axonal diameters or density may also influence DR. The axial diffusivity, DA, tends to be more variable in a broad range of WM changes and pathology. In axonal injury DA decreases possibly due to the increased debris from the disrupted membrane barriers (Sun et al., 2006). The DAs of WM tracts have been reported to increase with brain maturation as well (Ashtari et al., 2007; Bava et al., 2010; Gao et al., 2009).
It is generally assumed that the major eigenvector, e1, direction is parallel to the orientation of the WM fiber bundles. A common representation of this directional information is to map the x, y, and z portions of the major eigenvector into red, green, and blue color channels, respectively, weighted by FA, which is referred to as the directionally encoded color (DEC) map (Pajevic and Pierpaoli, 1999) (Fig. 2). The DEC map representation is useful for identifying and mapping WM tracts relative to brain lesions before intraoperative mapping (Jellison et al., 2004; Witwer et al., 2002).
It is important to note that the interpretation of all DTI stains (other than MD) is particularly challenging in regions with significant partial volume averaging, particularly in areas of crossing WM tracts (Alexander et al., 2001; Wheeler-Kingshott and Cercignani, 2009). In these areas, the diffusion tensor does not truly characterize the distribution from multiple fiber populations and the diffusivities in the parallel and perpendicular directions. One simple strategy for identifying and characterizing fiber crossings is using the tensor shape, which may be broken down into prolate (linear) (CL), oblate (planar) (CP), and spherical (CS) components (Alexander et al., 2000; Westin et al., 2002) (Fig. 2):
These shape stains, though not widely used, do provide a good assessment of where the diffusion tensor is most valid for describing WM properties—for example, where CL is highest (most prolate). Conversely, WM regions where either CP or CS is high will have problematic interpretations. Other strategies for describing shape properties of DTI include the skewness of the deviatoric portion of the diffusion tensor (Ennis and Kindlmann, 2006; Lange et al., 2010).
A major limitation of DTI is that it can resolve only a single fiber orientation within a voxel and fails in voxels with orientational heterogeneity (e.g., crossing fibers) (Alexander et al., 2001). This shortcoming stems from the tensor model's inherent assumption of Gaussian diffusion. The Gaussian function has only a single directional maximum, while voxels with multiple fiber orientations have multiple maxima, and hence cannot be described by a single Gaussian function. Consequently, such voxels will become partial volume averaged, with artificially reduced FA values, potentially precluding them from being as assigned as WM. Such failure poses a huge obstacle to WM tractography and interpretation of diffusion anisotropy (Pierpaoli et al., 2001).
The applications of DTI in the brain and spinal cord are rapidly expanding. It is currently the most widely used method for assessing WM changes in a broad range of brain diseases, injuries disorders, and changes with brain maturation, aging, and plasticity. A current PubMed search of DTI and brain or spine reveals more than 10,000 publications. Many of these applications and the interpretation of DTI measures are summarized in a recent review article (Alexander et al., 2007) and a recent book on diffusion MRI (Jones, 2011).
The diffusion-weighted signal behavior at low levels of diffusion weighting (e.g., b<1500sec/mm2) is fairly consistent with the diffusion tensor model. However, at higher levels of diffusion weighting (e.g., b>2000sec/mm2), the signal decay is no longer observed to be mono-exponential (Clark and Le Bihan, 2000; Niendorf et al., 1996). Several studies (Mulkern et al., 1999) attributed this deviation from mono-Gaussian diffusion to apparent fast and slow diffusing components of the apparent diffusion coefficient, and measured the decay of the diffusion signal over a range of b-values to estimate the apparent fast and slow components. However, there is controversy over the assignment of these components and whether the bi-exponential model should take into account exchange between compartments (Mulkern et al., 1999).
Diffusion kurtosis imaging (DKI) (Jensen et al., 2005) is another technique that investigates the non-mono-Gaussian properties of water diffusion by measuring the kurtosis of the diffusion propagator, which could reveal brain microstructure information hidden to DTI. The scalar apparent kurtosis coefficient may be used to quantify the extent to which water diffusion in brain tissue is non-mono-Gaussian. More recently, the diffusion kurtosis (DK) tensor was developed to take into account the diffusion anisotropy (Lu et al., 2006). Similar to the diffusion tensor, the 3D kurtosis properties can be completely described by a tensor, in this case the DK tensor being a symmetric 3×3×3×3 matrix with 15 independent elements (compared to the 6 of the diffusion tensor). DKI differs from the bi-exponential model in that DKI does not make an assumption on the number or even existence of biophysical compartments. Analogous to the diffusion tensor, rotationally invariant scalar measures or stains can be obtained from the DK tensor, such as mean kurtosis (MK), axial kurtosis, and radial kurtosis. Recently, a growing body of literature has looked at how these measures relate to aging and pathology. In a study on the effects of aging on the human prefrontal cortex (Falangola et al., 2008), the authors reported increased GM MK with age, when moving from adolescence to adulthood. In a rodent brain maturation study (Cheung et al., 2009), the radial and axial kurtosis measures provided better detection and characterization of the developmental changes in various WM and GM structures than their DTI counterparts. Both studies reveal DKI's great potential to better characterize GM microstructure change, which can be difficult for DTI.
The clinical value of DKI has been validated in several studies that show its ability to detect tissue microstructure abnormality such as in human head and neck squamous cell carcinoma (Jansen et al., 2010), cerebral glioma (Raab et al., 2010), and lung dysfunction (Trampel et al., 2006). In a recently published study looking at DKI in a rat model of traumatic brain injury (Zhuo et al., 2012), the MK significantly increased at the sub-acute stages of injury in all ipsilateral and contralateral regions, while standard DTI parameters gave inconsistent results. The authors associated the MK elevation with increased reactive astrogliosis, suggesting that DKI is sensitive to these microstructural changes, whereas DTI parameters alone may miss them. Although DKI is promising, it has not been studied or developed nearly as widely as DTI.
To resolve multiple fiber orientations, high angular resolution is needed, which can be achieved by increasing the b-value. In DTI, the b-value is typically ~1000sec/mm2, at which the angular resolution is poor for mapping crossing fibers. A growing number of high angular resolution strategies have been developed, based on increasing the number of encoding directions and in some cases multiple levels of diffusion weighting. In general, increasing the maximum diffusion weighting (b>2000sec/mm2) up to b ~ 17,000sec/mm2 or more increases the ability to resolve fiber distributions with better angular resolution, to better characterize diffusion in complex tissue.
A general method for estimating the probability distribution of displacements or diffusion propagator is the q-space formalism (Callaghan, 1996; Cory and Garroway, 1990). The wave-vector q=γGδ is analogous to the wave-vector k used in k-space sampling for MR image acquisitions. The diffusion signal in q-space and diffusion propagator are Fourier transform pairs:
where P(R, Δ) is the diffusion propagator, which describes the displacement distribution of the water molecules within a diffusion time Δ. In DTI, the propagator is modeled as a 3D Gaussian distribution. Variations of q-space imaging have been implemented on clinical MRI scanners, despite violating the narrow gradient pulse requirement for q-space. Even with this violation, the Fourier relationship in Eq. (11) is a reasonable approximation (Bar-Shir et al., 2008; Mair et al., 2002; Wedeen et al., 2005), and the diffusion displacements are similar in shape to reality, but may be underestimated (Bar-Shir et al., 2008; Callaghan, 1996; Lori et al., 2003; Mair et al., 2002; Wedeen et al., 2005).
Diffusion propagator details may be estimated using a large number (Ne>40 up to several hundred) of diffusion encoding directions with high angular resolution diffusion imaging (HARDI) strategies (Alexander et al., 2002; Frank, 2002; Tuch, 2004). HARDI acquisitions are on the order of 10–20min or more at resolutions comparable to DTI studies. The most common application of HARDI studies is to estimate the orientation distribution function (ODF) of the diffusion propagator, which is a directional representation of the propagator (Tuch, 2004; Wedeen et al., 2005). Often these methods use a spherical harmonic model of the ODF. Spherical deconvolution of the propagator ODF with the expected single fiber group ODF may be used to estimate the fiber ODF, which is an estimate of the fiber orientations in a voxel (Tournier et al., 2004).
Strategies for combining high angular resolution and multiple levels of diffusion weighting include diffusion spectrum imaging (DSI) (Wedeen et al., 2005), hybrid diffusion imaging (HYDI) (Wu and Alexander, 2007), and combined hindered and restricted model of diffusion (CHARMED) (Assaf and Basser, 2005). The q-space sampling for DSI is on a Cartesian grid, while HYDI and CHARMED used measurements in spherical coordinates (shells of q-space). DSI and HYDI estimate the propagator by taking the Fourier Transform of the q-space signals as described in Eq. (11). These approaches require many measurements (>100 and up to 500 or more) at different levels of diffusion-weighting (up to 20,000sec/mm2) and encoding angles. The acquisition time for whole-brain coverage with these approaches is on the order of 15min to more than an hour. Recently, analytic model solutions for estimating the propagator based upon q-space measurements have been proposed, including diffusion propagator imaging (Descoteaux et al., 2011), spherical polar Fourier expansion (Assemlal et al., 2009), and Bessel Fourier orientation reconstruction (Hosseinbor et al., 2011). These analytic models may enable sparser sampling schemes of q-space to be used.
The HARDI and more general q-space sampling methods have resulted in other quantitative measures that may be used to characterize WM microstructure. A measure of the directional variation or anisotropy of the ODF is the generalized fractional anisotropy (gFA), which is analogous to the FA measure for DTI, but should maintain high anisotropy in areas of fiber crossing (Tuch, 2004). The gFA appears to be reduced in anterior thalamic radiations and cingulum tracts of patients with obsessive compulsive disorder (Chiu et al., 2011), reduced in the corpus callosum with increased alcohol use (Liu et al., 2010), and abnormally asymmetric in association WM tracts in autism (Lo et al., 2011).
A commonly used measure from DSI and HYDI is the zero-displacement probability, Po, which is the integral of the entire q-space signal, normalized by the q=0 signal (Assaf and Cohen, 2000; Wu and Alexander, 2007). Po is the probability density of water molecules that minimally diffuse within the diffusion time and, hence, a measure of restricted diffusion. In a healthy adult brain, Po is greater in WM than GM because WM has more restricting barriers, including multi-layer myelin sheaths, axonal membranes, and microtubules. It also appears to be insensitive to WM fiber crossings, which cause reductions in the FA of DTI. Another measure of the propagator is the mean squared displacement (MSD), which is a displacement variance measure of the diffusion propagator; however, it is very sensitive to the measurement noise (Assaf and Cohen, 2000; Wu and Alexander, 2007). The MSD may be used to estimate the apparent diffusion coefficient by MSD/Δ. Another measure of diffusivity from DSI and HYDI is the q-space inverse variance (QIV), which is the inverse of variance of the q-space signal distribution (Wu et al., 2008). Note that for Gaussian diffusion, the QIV is the MSD, but is only an approximation in the case of non-Gaussian diffusion. Example quantitative q-space stain maps are shown in Figure 3.
Several studies have looked at age-related changes in Po of WM. In healthy children, Po was found to increase with brain maturation, but then plateau during adolescence (Ben Bashat et al., 2005). This age-related plateau was also observed in an in vivo HYDI canine study (Wu et al., 2011a) of brain maturation, where Po of global WM was computed in control dogs within the age range 3–16 months (similar to the period of early childhood through adolescence in humans). In another HYDI study (Wu et al., 2011b) on age-related changes in cerebral diffusion properties in healthy adult human brains, Po was found to be relatively constant across the age range (18–72 years). Several studies have shown Po to be sensitive to brain pathology. In a high b-value study of multiple sclerosis (MS) (Assaf et al., 2002), Po was reduced in both lesions and normal appearing white matter (NAWM). A recent in vivo HYDI study (Wu et al., 2011a) in a canine model of dysmyelination showed a significant reduction in the Po of WM in sick dogs with respect to controls, which is consistent with previous high b-value measurements in fixed, post-mortem spinal cord and brain specimens from myelin deficient rats (Bar-Shir et al., 2009; Biton et al., 2007). These studies suggest that changes in myelin are a significant mechanism for the differences in Po, though the axonal density and diameter may also play a role in modulation of Po.
In DTI, it is generally assumed that the major eigenvector is parallel to the local WM fiber orientation. Newer methods with many encoding directions (e.g., HARDI as discussed above) can estimate the directions of multiple fiber bundles. Tractography is a method for reconstructing the trajectories of major WM tract pathways using the orientation information from DWI. These methods can create 3D depictions of WM tracts (e.g., see Fig. 4).
At the foundation for many tractography methods is the streamline algorithm (Basser et al., 2000; Conturo et al., 1999; Mori et al., 1999), which estimates the WM trajectories in a propagation vector field. The most commonly used propagation vector is the major eigenvector from DTI, but one or multiple propagation vectors for each vector may be derived using the ODF from either HARDI or DSI. The simplest version of this is to step through the vector field in the direction of the local vector for small, finite distances. Smoother reconstructions may be obtained using higher order Runge-Kutta spatial integration methods.
The streamline algorithm is the basis for both deterministic and probabilistic tractography approaches. In deterministic tractography, a single reconstruction is produced for a given DTI data set. Probabilistic tractography methods attempt to characterize the uncertainty in the tract reconstruction, by performing a Monte Carlo tractography experiment by repeating the streamline reconstruction multiple times and perturbing the propagation vector field for every iteration (Lazar and Alexander, 2005; Parker et al., 2003). Strategies for this perturbation include adding noise to the tensor field or bootstrap resampling of the DWI data. In general, the dispersion of the reconstructed tracts increases with the distance from the starting location and decreases with increased diffusion anisotropy (Anderson, 2001; Lazar and Alexander, 2003; Lori et al., 2002).
An alternative strategy for reconstructing WM tracts is global tractography that attempts to find a reconstruction solution that is most consistent with the underlying diffusion MRI data. Global tractography methods include (1) constraining the connection endpoints (Cheng et al., 2006) and (2) optimizing a synthesized diffusion tensor field based upon tractography reconstruction (Fillard et al., 2009; Kreher et al., 2008). The latter approaches are very computationally demanding.
Recently, an interesting strategy for tractography reconstruction is to reconstruct the pathways on much smaller voxels, which reveals tract details that are much finer than the original resolution. Reconstructing the tracts on voxel grids with dimensions an order of magnitude smaller than the original acquisition appears to have super-resolution properties that may provide a unique image contrast mechanism (Calamante et al., 2010). The concentration or density of tracts is potentially a unique contrast mechanism (Calamante et al., 2010, 2012; Roberts et al., 2005), though caution should be used for interpreting as a quantitative measure. Superrresolution tractography is promising for mapping very fine WM detail below the original spatial resolution of the DTI data (e.g., the thalamus in Fig. 5). The extended coherence of the tractography pathway reconstruction can significantly constrain the WM detail at subvoxel levels. This superresolution property is also manifest by differences in the reconstructed pathways for multiple subvoxel seed points within a voxel. Pure upsampling of the DTI data merely provides a smoother image and does not yield similar hyperfine detail.
The main applications for tractography are (1) stains for visualization of specific WM tracts, (2) defining regions-of-interest (ROIs) for quantitative analyses, and (3), more recently, characterizing connectivity properties between two or more brain regions. The visualization of tractography maps may be used to generate virtual dissections of brain anatomy (e.g., Fig. 4). A clinical application of tractography is to visualize the location of WM tracts relative to a lesion or within a region of planned surgical intervention (Witwer et al., 2002). The application of tractography for quantitative image analysis is described further in the section on analysis methods.
Finally, an exciting application of this technology is to generate connectomes of structural connectivity in the brain (Hagmann et al., 2007, 2008). Structural connectomes may be generated using whole-brain tractography to assess the presence, absence or strength of connections between two or more brain regions. This approach permits the comparison of structural connectivity against functional connectivity defined by fMRI (Honey et al., 2007) or electrophysiology (Sporns et al., 2005). A connectome is a descriptive mathematical construct (e.g., an association matrix) where “edges” describe the connectivity or interactions between nodes of the connectivity graph. The most common way for generating a structural connectomes graph is to define the nodes by parcellation of the cortex and subcortical areas. Two nodes may be considered connected if tractography yields a reconstructed connection between the two nodes. A connectivity graph or associate matrix may be generated using the end-point or cortical regions as nodes and the tractography counts as the edges (Hagmann et al., 2007). Edge weights of the graphs may be defined as the tract connection count, or a binary threshold (Hagmann et al., 2007). Another possibility for edge weights is to integrate the FA or some other qMRI measure along the tract pathways. Analyses of the association matrices may yield measures and properties of graph connectivity including small-worldness, efficiency, hubs, distance, and clustering coefficients (Bassett et al., 2011; Gong et al., 2009; Hagmann et al., 2007). Several excellent articles discuss methods and measures for characterizing and representing brain connectivity properties based upon structural connectomes (Hagmann et al., 2010; Rubinov and Sporns, 2010). DSI appears to yield more accurate and robust connectivity properties than DTI (Bassett et al., 2011). Further, the connectivity properties appear to be relatively stable within subject yet are sensitive to differences between individuals (Bassett et al., 2011). Structural connectome properties have been investigated in several population based studies. One study investigated connectivity properties as a function of age and gender in adults (ages 19–85 years) and found reduced connectivity properties with age and in men (Gong et al., 2009). Conversely, the connectivity properties increase with age in children (Hagmann et al., 2010). Decreased structural connectivity has been observed in patients with Alzheimer's disease (Lo et al., 2010).
One of the key challenges in defining structural connectomes is the definition of node regions in the brain. Recently, a data-driven method, known as epsilon radial connectomes, was proposed by Adluru et al. (2012) and Chung et al. (2011b). A sample illustration of the technique using DTI data is shown in Figure 6. The main idea is to define node regions using clustering tracts in spatially normalized DTI data. In the particular framework (Adluru et al., 2012; Chung et al., 2011b), the authors identify nodes by clustering tract end points into spherical volumes of a particular radius (typically the amount of smoothing used in voxel-based analyses, e.g., 8mm). One can extend the framework to identify nodes by clustering the tracts based on their shape instead of just using end points.
There are several significant caveats and limitations with tractography. A major one is that the tract reconstructions are highly sensitive to errors anywhere along the pathway (Lazar and Alexander, 2003). An artifact in a single plane can lead to highly aberrant pathways (Pierpaoli et al., 2001). In general, tractography algorithms are poor at resolving crossing fibers (Barrick and Clark, 2004). Even tractography reconstructions that include HARDI measurements of crossing fibers may not necessarily reflect the actual connection strengths through these regions. Tractography methods are prone to both false positives (erroneous tracts) and false negatives (missing tracts), which can make the interpretation of tractography measurements challenging (Jones, 2011; Pierpaoli et al., 2001). In particular, the application of tractography to neurosurgical planning should be handled with care (Kinoshita et al., 2005). It should be noted that the interpretation of tractography can be challenging as even erroneous tract reconstructions often appear realistic. Expanded descriptions of tractography methods may be found in recent review articles (Chung et al., 2011a; Lazar, 2010) and book chapters 22–24 in Jones (2011).
Despite the recent proliferation of WM tractography studies and applications, it is difficult to assess whether a reconstruction is reflecting reality. The reasons for this are elaborated above. Several studies have evaluated tractography algorithm performance using anisotropic diffusion phantoms (Moussavi-Biugui et al., 2011; Perrin et al., 2005; Poupon et al., 2008; Watanabe et al., 2006) and synthetic phantom data (Close et al., 2009; Fillard et al., 2011; Lazar and Alexander, 2003). While these tools are good for investigating the properties of tractography algorithms, they may not adequately and realistically mimic the errors associated with neuroimaging and the complexity of WM in the brain. The most standard approach for validation is to compare the connected regions to the tract definitions from classic axon tracer and dissection studies (Caspers et al., 2011; Dauguet et al., 2007; Kier et al., 2004a, 2004b; Lawes et al., 2008; Schmahmann et al., 2007). Another promising approach for tracing WM pathways in vivo is using manganese, which is taken up by the calcium channels of the axons (Lin et al., 2001; Yamada et al., 2008). Using manganese, it is possible to trace transynaptic connections across distant brain regions. Manganese is paramagnetic which reduces the T1 of the tissue, leading to focal tissue enhancement of the specific pathways. In high enough doses, manganese is toxic to neurons though many research studies have applied it in living animal systems. To date, it has been used to map out the trajectories of the optic nerves, projections from the putamen and caudate in nonhuman primates, and from cortical regions (Murayama et al., 2006; Saleem et al., 2002). A recent study compared manganese tracer results with DWI tractography in minipigs, which showed fairly good (but not identical) correspondence (Dyrby et al., 2007).
It would be ideal to be able to directly image myelin in WM directly. Unfortunately, the 1H protons of myelin are essentially invisible using traditional MRI. The protons bound in myelin proteins and lipid bilayers have extremely short transverse relaxation times (T2s), which are in the microsecond range compared to the millisecond range of free water. However, dipolar coupling and chemical exchange facilitates an exchange of magnetization between two pools of protons—a free pool (e.g., water) and a bound pool (e.g., myelin macromolecules) (Fig. 7). This mechanism is known as MT (Wolff and Balaban, 1989), which may be used to sensitize MRI-visible water signal to the myelin macromolecular content. One way to induce the MT effect is to apply a strong off-resonance RF pulse at a frequency far from the free water resonance frequency (>1000Hz). The MT saturation pulse selectively saturates the magnetization of macromolecule-bound protons, which have a very broad frequency spectrum (the bandwidth is inversely proportional to T2), while leaving the free pool (long T2 with a narrow spectrum) relatively unaffected. Subsequently, the fast exchange of magnetization between the pools will partially saturate MRI-visible free water protons causing a decrease of the observed MRI signal intensity (Fig. 8). The MT attenuation of the free water signal is a complex function of the MT pulse properties (amplitude, rate, and frequency offset), the concentration of macromolecules, and the exchange rate of the magnetization between the free water and bound macromolecular pools. The MT effect is modulated by the offset frequency and amplitude of the RF saturation pulse. Maps of MT stains or measures are generated through weighted combinations (filters) of the saturation-weighted images at different frequencies and pulse amplitudes, discussed below.
The most common stain for characterizing the MT effect is the MT ratio (MTR), calculated as the relative change in intensity of images acquired without (So) and with (SMT) off-resonance MT pulses:
The example source images and corresponding MTR map are shown in Figure 9.
Increased MTR values are most often associated with increased macromolecular concentrations in the tissue. The higher MTR in WM is believed to be associated with the proteins and lipids associated with myelinated axons (Stanisz et al., 1999). Consequently, the MTR in WM is reduced in demyelinating diseases such as MS although the MTR can also be influenced by overall water content and other macromolecules in processes such as neuroinflammation (Stanisz et al., 2004). The reported ranges of MTR values in healthy WM and GM are roughly 0.4–0.55 and 0.25–0.3, respectively. This wide range in MTR values reflects the variability of MTR measurements across scanners, transceiver coils, and scanned objects (Berry et al., 1999; Filippi et al., 2000; Silver et al., 1999). One source of variability of MTR in the literature is a lack of standardization of pulse sequence protocols. In each particular implementation, the exact MTR measurement will depend upon the pulse sequence parameters (e.g., TR, TE, and excitation flip angle), the magnetic field strength, as well as the shape, amplitude, and frequency offset of the saturation pulses. Consequently, within a single MTR study, the imaging parameters should be fixed to maximize consistency. Efforts have been also made to solve the stability issues (Tofts et al., 2006) and standardize MTR studies across multiple sites. Another major source of MTR variability (especially at magnetic field strengths of 3T and higher) is system-specific and object-specific inhomogeneity of the B1 excitation field and the main B0 field. B0 inhomogeneities are caused by incomplete shimming and spatial variations in the magnetic susceptibilities in soft tissue, bone, and air, which lead to shifts (errors) in the saturation frequency offsets. At high magnetic fields (B0>1.5T), inhomogeneities in the B1 field may reach up to ±30% within the imaged object and will affect the saturation pulse amplitude and consequently alter the level of MT saturation. The B1 field may be measured and used to retrospectively correct MTR measurements (Ropele et al., 2005; Samson et al., 2006; Yarnykh, 2009). Figure 10 demonstrates the effect of B1 field correction (Yarnykh, 2009). Unlike correction of receiver coil inhomogeneity, correction of B1 transmit field effects on MTR is not trivial due to its effect on MT power deposition and effect on the excitation flip angle. Full correction of B1 effects is only possible using the more complete description of MT effect (such as provided by two-pool model in the next section). B1 correction of MTR relies on some approximations to the model. Other considerations for MTR measurements are discussed in two excellent review articles (Henkelman et al., 2001; Horsfield et al., 2003).
As discussed above, the MTR measurement is highly dependent upon a broad range of technical factors. Moreover, despite its sensitivity to macromolecular tissue content, the traditional MTR is a nonspecific indicator of underlying pathology that is affected by free water MR parameters (Tofts, 2003), which are modulated by other factors, for example, by inflammation (Stanisz et al., 2004). Fully quantitative MT (qMT) methods are required to improve the sensitivity, specificity, and stability of MT metrics (Gochberg et al., 1999; Gochberg and Gore, 2003; Ropele et al., 2003). Several investigators have adapted a two-pool model of MT for in vivo measurements (Sled and Pike, 2001; Tozer et al., 2003; Yarnykh, 2002, 2004; Yarnykh and Yuan, 2004). The two-pool model is fitted to data acquired with MT pulses over a range of offset frequencies and pulse amplitudes to estimate several underlying physical parameters of the tissues; most important are bound pool fraction (BPF or f) (relative concentration of macromolecules), cross-relaxation rate (k), the T2b and T2f (the relaxation times of the bound and free pools, respectively), and longitudinal relaxation rate of the free pool R1. The advanced qMT imaging (qMTI) protocols also include separate acquisitions of B1 and B0 maps, which are used in the data fit to mitigate the impact of scanning imperfections on the quantitative maps (Yarnykh, 2007). Example quantitative maps estimated using a modified cross-relaxation imaging method (Mossahebi and Samsonov, 2011; Yarnykh, 2004; Yarnykh and Yuan, 2004) are shown in Figure 11. Several studies confirmed that qMT methods are much more sensitive and specific to macromolecular content than the conventional MTR methods (Dula et al., 2010; Schmierer et al., 2007; Tozer et al., 2005). Current implementations of qMTI can achieve spatial resolutions on the order of 1.5–2mm; however, the scan times remain long (30min or more).
MTI is emerging as an advanced MRI method sensitive to various CNS injuries (Filippi and Rocca, 2004). Many of the published MT studies have focused on patients with MS, who show decreased MT in both ROI and whole-brain histogram analyses. There is growing evidence that MT-based MRI may be the most sensitive imaging technique capable of tracking myelin changes in patients with MS (Chen et al., 2005, 2007; Deloire-Grassin et al., 2000; Dousset et al., 1992, 1995; Schmierer et al., 2004; Trapp et al., 1998). MT contrast is a stronger predictor of MS disease course than conventional MRI measures (Pike et al., 2000; Rovaris et al., 2003; Santos et al., 2002). In other diseases, similar results were obtained, indicating that MTR is a viable marker for affected white and GM. MTR has been shown to increase with brain development during the first several years of life (Rademacher et al., 1999; van Buchem et al., 2001) and regional decreases with aging have been found (Armstrong et al., 2004). Differences in MTR were sufficiently large to distinguish patients with mild cognitive impairment from patients with Alzheimer's disease and controls (Kabani et al., 2002a, 2002b). A number of published studies have also used MT methods to compare the brains in patients with schizophrenia against healthy control subjects (Bagary et al., 2003; Foong et al., 2001; Kiefer et al., 2004; Kubicki et al., 2005). Reduced MTR measurements in corpus callosum and occipital WM have also been observed in a small sample of patients with late-life major depressive disorders (Kumar et al., 2004).
Several studies have revealed potential clinical significance of qMT measures. The qMT measures were sensitive to tissue composition manifested as regional variations in WM of the brain (Sled and Pike, 2001). Anatomical correlations of qMT parameters estimated by constrained cross-relaxation imaging (Yarnykh, 2004; Yarnykh and Yuan, 2004) also revealed the increase of BPF in major fiber tracts of the human brain (Yarnykh, 2004; Yarnykh and Yuan, 2004) and rat brain (Underhill et al., 2011), showing the strong association of BPF with the fiber density (Fig. 12). The BPF was able to track myelin levels in MS lesions (Davies et al., 2004).
A few studies have attempted to relate MT measurements to measures reflecting brain function. A serial MTR study in the optic nerves of 29 patients with acute optic neuritis was performed with measurements of visual system functioning using visual evoked potentials (VEP) (Hickman et al., 2004). No significant differences in MTR were observed between patients and controls at the onset of optic neuritis, although the MTR did decrease in patients over a period of 1 year. There did not seem to be any direct relationship between MTR and VEP measurements. Another study of 18 patients with early-stage MS (Au Duong et al., 2005) demonstrated a correlation between functional connectivity between left Brodmann areas 45/46 and 24 using an fMRI working memory task, and the MTR of NAWM and also with brain T2 lesion load. Consequently, the functional connectivity relationship with MTR suggests that changes in the functional working memory network is related to changes in the WM pathophysiology. The MTR in normal appearing brain tissue (NABT) has a stronger correlation with cognitive impairment in MS patients than MS lesion load (Filippi et al., 2000). Whole-brain MTR histograms have correlated with neuropsychological impairment in MS patients (Rovaris et al., 1998; van Buchem et al., 1998). A combined MTR and fMRI study (Filippi et al., 2002) of simple motor function in patients with MS revealed correlations between the MTR histogram features of whole-brain, NABT (both GM and WM), and fMRI signal strengths in ipsilateral sensorimotor cortex and supplementary motor area (bilaterally). The fMRI signal in the contralateral sensorimotor cortex was significantly correlated with MTR histogram features in patients with cervical but not dorsal spinal cord myelitis (Rocca et al., 2006). A recent combined MTR and DTI study of 40 schizophrenia patients and 40 healthy participants showed decreased FA in the left uncinate fasciculus in the patients with longer illness duration and increased mean MTR in the right uncinate fasciculus (Mandl et al., 2010). Finally, an interesting study comparing MTR and FA of fronto-striatal WM pathways in attention deficit/hyperactivity disorder (ADHD) showed reductions in FA, but not in MTR, suggesting that the microstructural features are more altered than myelin in ADHD (de Zeeuw et al., 2011).
The remaining challenges of MTR imaging to serve as a stable and reproducible measure of neural tissue integrity include standardization of MTR protocols and making MTR imaging independent of system- and object-specific factors. Recently, an alternative MT measure, MT saturation, was proposed to minimize variability related to B1 effects (Helms et al., 2008). In this approach, MT measurements are augmented with an additional T1w scan, which is used along with MT-weighted data to yield MT saturation maps. Truly, qMTI has the potential to overcome aforementioned shortcomings of MTR. Unfortunately, no studies to date have attempted to relate qMT measures to brain function. One possible reason is a limited clinical utility of early qMT methods, which incurred a multiple fold increase in scan time (>1h) compared with MTR. Recent developments demonstrated that optimized acquisition is possible within clinically acceptable times (Cercignani and Alexander, 2006; Underhill et al., 2009), which may facilitate application of these sensitive methods in clinical studies.
The signal intensity of different brain tissues in a typical MRI experiment is a function of the fundamental properties of water protons within the tissue. The interaction of the surrounding tissue environment with water protons influences the relaxation times of both the longitudinal and the transverse components of the magnetization. These interactions are influenced by the random motion (e.g., diffusion) and local magnetic field fluctuations within the tissue medium. The spin-lattice relaxation time (T1) is the recovery of longitudinal magnetization back to equilibrium after excitation by a radiofrequency pulse. The spin-spin relaxation time (T2) is the decay time associated with the loss of transverse magnetization signal due to dephasing. Additional dephasing due to reversible field gradients (T2′) leads to a more rapid loss of signal (T2*<T2) when a 180° RF pulse is not utilized. In addition to relaxation, the strength of the observed signal is also a function of the overall proton density (PD), or number of nuclei contributing signal within a given voxel. In general, more restricted, dense or viscous tissue environments will exhibit reduced T1 and T2. The presence of metal ions in the tissue will also influence the relaxation properties. Example T1 and T2 maps are shown in Figure 13.
By exciting and acquiring MR signal at different time intervals or sequence filters (i.e., TR and TE), one may preferentially stain the intensity and contrast of different tissues, generating T1-weighted (T1w), T2w, or PDw anatomical images. Relaxometry, on the other hand, refers to quantitative methods to map relaxation times within tissues. Accurate techniques seek highly specific measurements of one particular parameter (T1, T2, or PD), and thus remove other confounds to image contrast such as receiver coil sensitivity profiles or PD (in the case of T1 and T2; see Fig. 13). These images show improved contrast between brain structures (Deoni et al., 2005a, 2005b) and may be more useful for the segmentation of brain tissue types than traditional imaging (Alfano et al., 1997). In the context of brain connectivity, these maps may provide more specific delineation of brain structures to serve as priors or landmarks for structural connectivity methods such as tractography.
T1 and T2 are sensitive measures of the local microstructural environment within WM tracts. Although many studies show that neurological diseases affect T1 and T2, these relaxation times depend on a wide range of tissue factors and are thus nonspecific. Both are highly dependent on water content, and tend to increase with bulk water in tissues. T1 may also decrease with decreasing lipid content, as observed in MS plaques (Lacomis and Osbakken, 1986), while T2 decreases as the size of the local water compartment becomes restricted.
Traditional gold-standard methods to measure T1 rely on saturating or inverting the longitudinal magnetization, sampling its recovery at different time points (TI) (Pykett et al., 1983), and then fitting it to a monoexponential model of magnetization recovery. This method is slow, as it requires multiple inversion times and a TR approximately five times longer than T1 to allow complete recovery of longitudinal magnetization. Although the Look–Locker method (Look and Locker, 1970) improves the efficiency of the technique by sampling multiple TI per repetition, all inversion techniques are limited in spatial resolution due to long scan times.
Traditional methods to measure T2 rely on obtaining spin echo measurements at different TEs and fitting them to a monoexponential model of signal decay. T2* measurements are similar to T2, except a 180° RF pulse is not used. Multiple spin echoes can be measured in each repetition with the Carr-Purcell-Meiboom-Gill (CPMG) method (Meiboom and Gill, 1958). However, this sequence is extremely sensitive to imperfect RF pulses, which will generate magnetization that follows a T1 pathway and thus overestimates T2. These can be suppressed with a proper sequence of crusher gradients inserted symmetrically around RF pulses (Look and Locker, 1970).
Faster methods of relaxometry rely on steady-state acquisitions. The variable flip angle method (Christensen et al., 1974), also known as DESPOT1 (Homer and Beevers, 1985), can generate a T1 map from spoiled gradient echo (SPGR/FLASH) images at two or more flip angles, while the DESPOT2 method can generate a T2 map from steady-state free precession (bSSFP/FISP/FIESTA) at two or more flip angles in combination with DESPOT1 T1 maps (Deoni et al., 2003). These techniques may be more attractive in the context of brain connectivity, as whole-brain coverage can be achieved at high resolutions and reasonable scan times (Deoni et al., 2005b). These methods are also extremely easy to optimize given a single T1 time (Deoni et al., 2005b) or a range of expected T1 values (Cheng and Wright, 2006), and can be casted into a linear form for straightforward data fitting. The estimation of T2 from steady state sequences is a bit more complex as the signal also depends upon T1, so that both must be measured in the same experiment.
Although steady-state techniques have clear advantages over traditional relaxometry, they must be implemented carefully to ensure accurate measurements. Both DESPOT1 and DESPOT2 suffer from a strong dependence on excitation flip angle, although several well-matched steady-state calibration techniques have been developed to correct for this (Deoni, 2007; Sacolick et al., 2010; Yarnykh, 2007). DESPOT1 is also highly sensitive to proper SPGR sequence spoiling (Yarnykh, 2010), which may require the use of large gradients and increase the overall time of the technique. DESPOT2 also suffers from a strong dependence on main field inhomogeneity; a technique to correct this has been developed, although it requires twice as many data points to be acquired (Deoni, 2009). Even with careful consideration to these technical issues, further errors may be present due to an incomplete model of the MR signal. In addition to T1 and T2 relaxation, the rapid radiofrequency excitation of steady-state sequences may induce on-resonant MT effects, biasing T1 measurements based on the macromolecular content of tissues (Ou and Gochberg, 2008). A typical DESPOT1 protocol at 3T involves two SPGR scans and a flip angle calibration scan (usually performed at ½ resolution, as this parameter is spatially smooth). A typical 1.5-mm isotropic resolution protocol with AFI flip angle correction (Yarnykh, 2007) takes about 10min. Resolution can be increased at the cost of longer scan time, but is not limited by gradient performance or distortions as in EPI acquisitions typically utilized in diffusion imaging. For example, excellent results at 0.34mm3 isotropic resolution have been achieved through multiple averaging of DESPOT1 datasets (Deoni et al., 2005a).
All of the aforementioned relaxometry methods, both traditional and steady-state, operate under the assumption of a single, well-mixed pool of water within each voxel. This is an inaccurate assumption in the case of neural tissues, as water is known to exist within extracellular, intra-axonal, and myelin spaces separated by diffusion barriers. In particular, the water that is trapped in the myelin bilayers has extremely limited mobility, which significantly reduces the T2 of myelin water (Fig. 14). In the case of CPMG T2 imaging, this results in a nonmonoexponential decay of the signal with TE (MacKay et al., 1994; Menon and Allen, 1991; Whittall et al., 1997). This has led to the development of multicomponent T2 (MET2) relaxometry, which seeks to model multiple T2 values in distinct components of water within a single voxel, such as free water (e.g., edema and CSF, which have long T2>120msec), extracellular water (T2 ~ 60–90msec), and water within the myelin membranes of axons (T2 ~ 10–40msec) (MacKay et al., 1994; Stewart et al., 1993). The ratio of the myelin water component to the overall signal has been coined myelin water fraction (MWF) (Stewart et al., 1993), and may be useful as an indirect measure of myelination within the brain. An example MWF map is shown in Figure 13. Note that the relative fraction of the signal within the myelin water component is relatively small (<20%) and the T2 of this component is relatively short (<30msec) such that short echo spacings (<10msec) are necessary to be able to sense it.
MWF may be measured using a more traditional multi-echo CPMG experiment described above with typically 32 echoes or more. There are several challenges associated with this kind of measurement in addition to issues related to stimulated echoes and RF inhomogeneities present in single component T2 measurements. The short T2 of the myelin water favors CPMG sequences with very short echo spacing to obtain an accurate and precise estimate of the short T2 signal (Dula et al., 2009). The estimation of the multiple T2 components and times is typically performed using a nonnegative least squares algorithm with a multi-component exponential decay function and is highly sensitive to measurement noise and errors (Whittall and Mackay, 1989). Some form of data regularization is also needed, as the model is usually highly underdetermined. This analysis also assumes no exchange of water between the compartments over the time scale of the measurement (typically 128–320msec for 32 echoes and 4–10msec echo spacing). Recent studies suggest that exchange rates may differ substantially with differing myelin thickness, causing a systematic underestimation of MWF for faster exchange rates (i.e., thinner myelin) (Dula et al., 2010). Standard two-dimensional (2D) multiple spin echo sequences often have imperfect RF pulses, which will lead to stimulated echo pathways in the echo train and potentially bias the T2 measurements. To minimize the image noise and bias errors in the echo signals, these measurements are typically performed using a single-slice 2D CPMG sequence with thick slices (4mm or more), hard or composite RF pulses, and variable-amplitude gradient-crusher schemes with a total scan time of 12–20min. Thus, the spatial coverage is limited and the method is fairly inefficient for characterizing the MWF or T2 for multiple slices. A 3D CPMG version of this method has been demonstrated, which improves the coverage and improve the overall sensitivity through increased signal averaging (Mädler and MacKay, 2006).
Multicomponent behavior has also been observed in 3D steady-state SPGR (Deoni et al., 2007) and SSFP (Deoni et al., 2008b) sequences. This has recently led to the development of a multi-component version of DESPOT (mcDESPOT) (Deoni et al., 2008a) that is able to generate whole-brain MWF maps at 1.5–2.0mm resolution with scan times similar to traditional single-slice CPMG MET2 methods (10–20min). The current mcDESPOT signal model includes two (fast and slow) T1 and T2 components, an exchange time constant, and the relative signal fractions of the components. The signal with the short T1 and T2 is assumed to originate from the myelin water in WM. An issue is that this model is mathematically complex. It has issues with nonconvergence due to local minima and requires computationally intensive nonlinear stochastic global optimization methods (Berger and Silverman, 1991) to converge to a reasonable solution. The model will also not properly converge for in tissues with long T2 species like either CSF or edema although convergence may be improved by adding a nonexchanging long T2 compartment at the cost of even further computational complexity (Deoni, 2011). Despite some of these issues, multicomponent relaxometry with steady-state sequences is promising for characterizing myelin water over the entire brain in a time feasible for many research applications.
Single-component relaxometry stains are sensitive, although nonspecific, to subtle differences in the microstructure of WM tracts. Studies in healthy individuals have shown that T1 times in specific WM tracts differ over a range of 640–836msec at 1.5T (Yarnykh, 2004, 2009; Yarnykh and Yuan, 2004), while tract T2* values between 48.8 and 57.0msec have been observed at 3T (Cherubini et al., 2009). A more direct application to measuring brain connectivity is through the introduction of manganese as a tracer. Manganese is taken up by the calcium channels of the axons and follows transynaptic connections across distant brain regions. It is also a paramagnetic T1 shortening agent. While many studies utilize T1w imaging, T1 mapping can be used to make accurate measurements of in vivo tracer concentration (Kim et al., 2011).
MWF is of particular interest as this stain is most closely associated with myelination and myelin geometry, which have a significant impact on the conduction velocity of WM connections (Smith and Koles, 1970). MWF as measured with MET2 is highly correlated with myelin content in profoundly demyelinating diseases such as MS (Tozer et al., 2005), although much less so among subtle variations in healthy myelinated tracts (Dula et al., 2010). Whole-brain mcDESPOT studies in a canine sh pup model of dysmyelination have also shown substantially diminished MWF consistent with the paucity of myelin in the affected animals (Hurley et al., 2010). Recent human mcDESPOT studies have showed an increase in the WM MWF with brain development in children from birth up to 5 years of age (Deoni et al., 2011). The mcDESPOT MWF also appears to be affected in both lesions and NAWM in patients with MS and these changes appeared to relate to the degree of clinical severity (Kitzler et al., 2011). MWF is a promising stain for characterizing the myelination of WM tracts, although further work needs to be done to investigate its sensitivity to small variations in myelination and how it is affected by different microstructural properties such as myelin thickness, g-ratio, overall myelin content, differentiation between intact versus damaged myelin, and exchange of water between myelin and nonmyelin compartments. These are critical issues for using MWF to assess and track myelin repair therapies.
Although the vast majority of recent MRI studies of WM have focused on diffusion, MT, or relaxometry, there are other techniques that may provide complementary information. One of the oldest methods is MR spectroscopy (MRS), which may be used to characterize specific metabolites in the tissue, including N-acetylaspartate (NAA), creatine, choline, and neurotransmitters like gamma-aminobutyric acid (GABA) and glutamine/glutamate. Each of these metabolites reflects different physiological processes and has unique spectral signatures. Of significant interest in WM is NAA, which is a marker of the presence, density, and health of neurons, including the axonal processes. In fact, NAA may be one of the most specific markers of healthy axons and, as such, it is surprising that it is not used more widely for the investigation of WM in the brain. This may be due in part to the fact that MRS is extremely sensitive to the homogeneity of the magnetic field, which makes it challenging to apply in areas near air or bone interfaces. The concentrations of the metabolites are also in the micromolar range (compare with multiple molar for water); thus, large voxels must be used and the acquisition speed is slow. Therefore, MRS studies are often limited by poor coverage, poor resolution, and long scan times.
The recent push toward ever higher magnetic fields makes qMRI methods more challenging. Imaging distortions in DTI studies increase proportional to the field strength. The RF power deposition (specific absorption rate) increases quadratically with the magnetic field strength, which limits the application of MT pulses and can also limit the flip angles used in steady-state imaging. However, susceptibility weighted imaging is one method that greatly benefits from higher magnetic field strengths. Recent studies have observed interesting contrast in WM tracts as a function of orientation and degree of myelination (Liu et al., 2012). Stunning images of WM tracts have recently been obtained in ex vivo brain specimens (Sati et al., 2012). Techniques for characterizing WM in the human brain are only beginning to be developed.
Other WM cellular components are the glia, which include oligodendrocytes, astrocytes, and microglia. In general, there are no specific markers of changes in either oligodendrocytes or astrocytes. Recent evidence suggests that hypointense WM lesions on T1w imaging may indicate reactive astrocytes (Sibson et al., 2008). Increases in microglia often accompany inflammation, which can be detected using contrast agents, either gadolinium or superparamagnetic iron oxide (SPIO) particles. Recent studies have suggested that SPIO particles are preferentially taken up by macrophages in inflammatory regions. The impact of these contrast agents on other qMRI measures has not (Oweida et al., 2004) been widely studied, and thus multimodal imaging studies must be designed carefully.
Many of the qMRI measures appear to demonstrate sensitivity to myelinated WM; in general, MWF, BPF, and FA all are increased and DR is decreased in myelinated WM relative to GM, whereas demyelinating WM lesions all show reversed trends in these measures. Note that in areas of crossing WM tracts, the relationship of FA and DR with myelination is likely to be less predictable. However, the qMRI methods are sensitive to different mechanisms and therefore are potentially complementary. For example, either the BPF from qMT or the MWF from relaxometry might be able to provide more specific information than either FA or DR from DTI regarding myelin changes in brain areas with crossing WM tracts. While both the MT and T2 relaxometry appear to be more specific to myelin, head-to-head comparisons have showed poor correspondence. A comparison of the MTR and the MWF in both healthy controls and MS patients did not show any correlations in GM or NAWM, but did demonstrate a modest correlation in MS lesions (r=0.5) (Vavasour et al., 1998). This may be due to the differences in the mechanisms—the MWF will be sensitive to the spacing of the myelin membranes, whereas MT effects are more sensitive to the presence of proteins and other macromolecules in the myelin bilayers. Recently, a few studies have applied multiple contrast mechanisms to investigate WM differences. Several studies attempted to investigate the association between qMT measures and DTI parameters (Stikov et al., 2011; Underhill et al., 2009). They observed the weak correlation between BPF and FA in the WM of human brain; however, strong correlations between DTI and qMT parameters were found in GM (Underhill et al., 2009). It was concluded that the lack of correlations between qMT and DTI parameters in WM comes from the differences in the physical principle of these two methods as DTI parameters are dependent on the direction of fibers while qMT parameters are also associated myelination level and with the density of myelinated fibers (Underhill et al., 2009). A comparison of qMT (BPF) and DTI measures in fixed brain tissue from a dysmyelinated, shiverer mouse model showed decreased BPF and increased DR in dysmyelinated WM with a modest correlation (r=0.57) between BPF and DR in WM (Ou et al., 2009). Gender comparisons in regional corpus callosum measurements of both DTI measures and T2 MWF showed weak correlations between MWF and both FA and DR (r=0.39 and 0.35, respectively) and the regional significance by gender was not consistent for DTI and MWF (Liu et al., 2010). A couple of studies have compared MWF and BPF measurements in WM. Tozer and associates (2005) found a negligible correlation between these measures in NAWM of MS patients. Another more recent study reported that the qMTI measures, including BPF, were much more sensitive than MET2 relaxometry to temporal changes associated with the disease in MS lesions, though it has been reported that qMT is also sensitive to inflammatory processes (Levesque et al., 2010). An aging study of several qMRI measures (DTI FA and MD, MTR, and T1, T2, T2* relaxation times) in healthy adults versus age (range 18–85 years) revealed distinct age-related trajectories and spatial patterns for each of the qMRI measures (Draganski et al., 2011).
A number of studies have also combined MRS and qMRI measures. As discussed above, the NAA from MRS is often described as a fairly specific marker of viable neurons and axons in WM. A summary of a few of these studies is included here. A small (N=8) study comparing NAA and MTR in frontal WM of patients with late-onset depression showed a high correlation (r=0.89) (McLean and Barker, 2006). An older study comparing NAA and MTR in (N=13) patients found strong correlations (r=0.73) in WM lesions (Kimura et al., 1996). A DTI study (N=25) of small vessel disease reported modest correlations between NAA and either FA or MD (r ~ 0.5) in the centrum semiovale, a region with significant fiber crossings (Nitkunan et al., 2006). A recent study of 15 MS patients showed no relationship in the splenium of the corpus callosum between NAA and FA (Cader et al., 2007). In that study, they concluded that the changes in FA were from axon loss, which was not detected by differences in NAA. Conversely, an MRS versus DTI FA study of glioma patients found a high correlation between NAA and FA in and around the tumor (r>0.94), but the relationship was much more modest in NAWM (r~0.5) (Goebell et al., 2006). A comparison of high b-value q-space imaging (DSI) and MRS showed modest correlations between the NAA and both the MSD and the zero-displacement probability (r=0.61 and 0.54, respectively) (Assaf et al., 2005). The relationship between NAA and MWF has largely been unexplored.
The inconsistent correlations between these WM measures suggest either that the measures are reflecting different physiological features of the WM or that the measures are fairly noisy. In reality, it is probably a combination of both factors. It should be noted that there is considerable variability in the measurement protocols and analysis methods for each of the qMRI modalities, which greatly influences the accuracy and variance of the measurements.
The analyses of qMRI maps are particularly challenging because many of the measures are extremely heterogeneous across the brain. For example, the FA values in healthy WM can range from roughly 0.2 up to nearly 1.0. However, within a fixed and small region of WM (e.g., the genu of the corpus callosum or the posterior limb of the internal capsule), it is possible to compare values between individuals, though it is critical to use methods with anatomic specificity. There are many strategies for comparing qMRI measurements between subjects with their relative strengths and weaknesses. Table 1 lists different analysis strategies and the relative merits and weaknesses. The simplest approach is to compute a whole brain histogram at the obvious sacrifice of anatomical specificity. The three primary strategies for obtaining region-specific measures are (1) manual segmentation of a ROI, which can be extremely tedious, (2) automated application of an atlas-based template to the DTI data, and (3) tractography-based segmentation of specific WM pathways. Tract-based analyses based upon tractography have appeal for obvious reasons, though the confidence of the qMRI measures is higher in the trunk of the reconstructed tract and less so at branches and the periphery (see Fig. 15). More generic analyses of regional differences may be tested using voxel-based analysis (VBA) methods similar to voxel-based morphometry (Ashburner and Friston, 2000). An advantage of VBA methods is that differences may be detected anywhere without any specific a priori anatomic hypotheses, though the statistical power is much reduced. An issue with VBA is that the image processing steps—spatial normalization and blurring—introduce more partial volume averaging, which causes mixing of different tissue types (e.g., WM and GM or CSF) and makes the analysis sensitive to morphology differences, such as might be present between experimental groups being compared. Recent VBA strategies—T-SPOON (tissue-specific, smoothing-compensated VBA) (Lee et al., 2009) and tract-based spatial statistics (TBSS) (Smith et al., 2006)—help to ameliorate some of these limitations. An important step for several of these analysis methods is spatial normalization, which attempts to co-register the anatomic-specific qMRI measures across subjects. Recent advances in nonlinear, diffeomorphic spatial normalization methods (Zhang et al., 2010) and full diffusion tensor matching and reorientation (Zhang et al., 2006) (see Fig. 16) significantly improve the anatomic correspondence between subjects.
In summary, we have presented several qMRI measures or stains that are promising for characterizing WM in vivo. Diffusion MRI (including DTI), MT, and relaxometry are all sensitive measures of myelination and axons; however, each is based upon different mechanisms. Further, the specificity of these measures to specific WM properties like the degree of myelination is less clear and many questions remain. Diffusion MRI is modulated by the presence and spacing of membranes and any other barriers, which include the myelin membranes as well as any other cellular structures in the WM. MT is sensitive not only to the myelin proteins, but also to any other proteins and any other macromolecules, such as those found in regions of inflammation. The short T2 signal from the water trapped in the myelin bilayers appears to be fairly specific, though the actual quantification in terms of the amount of myelination is more challenging.
An important perspective to maintain in the interpretation of these qMRI stains is that all of these measures are modulations of the water signal. A change in the overall amount of water in a region of tissue will significantly influence the qMRI measure irrespective of the axonal properties. For example, edema will increase the extracellular water fraction, which will impact all of the qMRI measures described here. A decreased MWF or increased DR in a region of WM with edema does not necessarily reflect a decreased level of myelination. The size of the voxels is also very large relative to the cellular structural features that are being characterized. A 1-mm cubic voxel may contain more than a thousand axons (1–20μm in diameter), which may have a broad range of diameters, degree of myelination, and numbers and types of glia; thus, these images are very course and blurred maps of the microstructural detail.
The qMRI stain maps are all derived from multiple contrast-weighted images, which causes these imaging methods to be sensitive to misregistration from head motion, measurement noise, and artifacts in any of the images. Thus, it is critical to carefully review individual image quality and the registration fidelity before computing quantitative maps. The calculation of the quantitative measures often uses highly nonlinear models with multiple local energy minima in the solution space, making them highly sensitive to the measurement noise. These measurement noise effects can subsequently lead to biased estimates with high variance. If the noise is too high to achieve reliable estimates, then either scan time should be increased or spatial resolution decreased. Ideally, the SNR of the original image measurements should be reported in publications to be able to assess the level of image quality. Studies are also needed to determine SNR cutoff thresholds below which the calculations are either biased or unstable. While obtaining multiple qMRI measures in a single study is appealing, the scan time can be considerable—for example, DTI is on the order of 10min and qMT and multicomponent relaxometry are on the order of 20–30min or more each. Thus, if imaging time is limited, it is probably preferable to spend more time on a single qMRI measure or chose simpler measures that can be estimated from smaller data sets (e.g., MD instead of the full diffusion tensor, or T1 with B1 calibration instead of MWF). Current and future improvements to coil sensitivity design, parallel imaging, and constrained reconstruction methods for undersampled multiparametric image data (Velikina et al., 2011) may be used to significantly accelerate acquisition times and/or improve the measurement accuracy.
Note that there is an inherent trade-off between resolution and SNR. In general, imaging can accurately resolve signal from structures that are at least twice as large as the resolution. If the imaging resolution dimension is larger, then the minimum resolvable structure size likewise increases. For smaller structures (e.g., fornix and cingulum bundles), the measurements will have some partial volume averaging, which makes it difficult to disambiguate the microstructural properties from the macrostructure. Another consideration is that as long as the SNR is not too low (>3–4) for any of the images, the SNR can be improved by spatial smoothing so obtaining DW images at the highest possible resolution is a reasonable strategy. The concept of superresolution tractography is particularly exciting and novel and may provide details beyond the inherent image resolution; however, the quantitative measures along those pathways may still have partial volume averaging effects.
To apply these qMRI stains to multicenter clinical trials, it is necessary to develop standardized acquisition protocols that include methods to correct for errors and inhomogeneities in both B0 (static field strength/frequency) and B1 (flip angle). This is currently challenged by differences in pulse sequences on different scanner platforms. Phantom materials with specific qMRI properties may be useful for comparing measurements across scanners and sites. Further, while there are a growing number of software tools for calculating and analyzing DTI images, there are no widely available tools for either qMT or multicomponent relaxometry, which limits their application to more technically advanced research groups.
DTI has clearly been the most widely used method for investigating and describing structural connectivity properties in the brain. As the field moves forward, it is critical to also investigate MT and relaxometry measures along specific WM pathways to obtain complementary and potentially more specific information about the biological properties of these connections. Relevant to this point, there is still a lot that is not known about the mechanisms of these qMRI measures and how they are influenced by subtle variations in CNS pathology. More detailed and specific studies that relate WM histology and pathology to qMRI measures are essential to move this field forward and make the interpretation of these measurements more clear.
The recent work in mapping the global networks or connectomes of structural connectivity using tractography-based approaches is extremely exciting and has generated considerable enthusiasm in the neuroscience community. However, we must remember that these networks represent abstractions of the real structural brain connections through the modulation of water diffusion properties by the WM microstructure. Sophisticated mathematical models are being applied to characterize these networks, which hopefully reflect the structural properties of biological substrates that we are trying to characterize. How to define and/or interpret connectivity based upon structural connectomes is a work in progress. To date, connectome studies have focused on DTI/DWI properties based upon tractography properties; however, future connectome studies may also incorporate other WM measures like the FA, the BPF, or the MWF.
After these methods are standardized, qMRI stain atlases can be generated as a function of age, gender, disease, and or trait measure. Either tractography-based or morphologic-based templates of WM regions or structures can be used to characterize WM properties across populations. Integration of these atlases and tract-based measures may subsequently be compared against functional connectivity measures. This integration of qMRI stains with functional connectivity will provide a more complete picture of brain connectivity properties.
The authors acknowledge support from several relevant grants that supported and inspired some of this work, including NICHD P30 HD003352 (A.L.A.), NIMH P50 MH84051 (A.L.A.), NIMH RO1 MH080826 (A.L.A.), NINDS RO1 NS065034 (A.A.S.), NINDS RO1 NS050466 (A.S.F.), the National MS Society (A.L.A., A.A.S., and A.S.F.), and the Morgridge Institutes for Research (MIR-University of Wisconsin, NA). The authors are also grateful for figure contributions from Fernando Calamante and Vasily Yarnykh. Figure 12 has been reprinted from Yarnykh and Yuan, Copyright (2004), with permission from Elsevier.
No competing financial interests exist.