Quantitative T2 mapping of white matter: applications for ageing and cognitive decline

This study included 37 HC participants (22 female, mean age 67 years, age range from 49 to 87 years), 12 participants with MCI (4 female, mean age 74 years, age range 61–87 years) and 9 with mild AD (7 female, mean age 74 years, age range 56–91 years). MCI diagnosis was based on the criteria given by Petersen et al (1997). Participants also underwent the paired associative learning (PAL) task of the CANTAB toolbox (Soares et al 2014). This task requires that the participant recall the locations of various patterns, increasing in complexity as the task progresses, serving as a test of episodic memory. Cognitive testing was within 6 weeks of MRI. All participants gave informed consent and ethical approval was granted by the National Health Service Research Ethics Committee of North Bristol-Frenchay.

All imaging was performed using a Siemens Magnetom Skyra 3T system equipped with a parallel transmit body coil and 32-channel head receiver array coil. The imaging protocol comprised a 3D T1-weighted MPRAGE, 2D multi-contrast spin-echo and 2D DTI with the following parameters: MPRAGE: coronal, TR 2200 ms, TE 2.42 ms, TI 900 ms, flip angle 9°, acquired resolution 0.68  ×  0.68  ×  1.60 mm³, acquired matrix size 152  ×  320  ×  144, reconstructed resolution 0.34  ×  0.34  ×  1.60 mm³ (after 2-fold interpolation in-plane by zero-filling in k-space), reconstructed matrix size 540  ×  640  ×  144, GRAPPA factor 2 (Griswold et al 2002), time 5:25. Multi-contrast spin-echo: TR 4500 ms, TE 12 ms, number of echoes 10, echo spacing 12 ms, acquired resolution 0.68  ×  0.68  ×  1.7 mm³ inclusive of 15% slice gap, acquired matrix size 152  ×  320, 34 slices, reconstructed resolution 0.34  ×  0.34  ×  1.7 mm³ (after 2-fold interpolation in-plane by zero-filling in k-space, and inclusive of 15% slice gap), reconstructed matrix size 540  ×  640, 34 slices, GRAPPA factor 2, time 11:07. DTI: axial, TR 3800 ms, TE 85.2 ms, Bval 1000 s mm⁻², number of gradient directions 60 (full-sphere), acquired and reconstructed resolution 2  ×  2  ×  2 mm³ inclusive of 10% slice gap, matrix size 122  ×  122, 60 slices, GRAPPA factor 2, multi-band factor 2 (Feinberg et al 2010), time 4:30. Note that the T2 mapping did not have full-brain coverage, its anterior-posterior coverage only extending around 1 cm anterior and posterior to the head and tail of the hippocampus respectively, with the acquisition tilted such that the hippocampal body created an axis normal to the slice acquisition plane.

Diffusion tensors were fitted using dtifit after eddy current correction using the eddy_correct program of fsl (Zhang et al 2001). T2 maps were computed by a voxel-wise fit of a mono-exponential function in a logarithmic space, excluding the first echo. This was done since the pulse sequence allows the passage of both spin and stimulated echoes due to the use of identical crusher gradients astride each refocusing pulse, though the first echo contains only spin echo contributions. Its exclusion therefore means that the time points to which the exponential function was fitted are the sum of spin and stimulated echoes. The exclusion of the first echo has previously been examined and applied by other authors (Maier et al 2003), who also demonstrated a marked improvement in quality of fit. This also has the effect of filtering out fast-decaying components of the signal. The final echo time of 120 ms meant that coherence in parenchyma decayed to ~25% of its initial value, a compromise between maintaining sufficient signal for a meaningful measurement but sampling enough of the decay for a precise T2 estimate. T1-weighted images were brain-extracted using the vbm8 toolbox (Ashburner 2012). Total WM volumes were calculated by segmentation of the T1-weighted MPRAGE scans in native space using fsl fast segmentation after brain-extraction by the vbm8 toolbox within spm.

To analyse the diffusion tensor and relaxometry data, the TBSS toolbox of fsl was used (Smith et al 2004, 2006). FA images were registered to the FMRIB58_FA standard template and the FA skeleton determined at a threshold of 0.2 after which permutation testing by the program randomize (Winkler et al 2014) was used to test for group differences. The demeaned age was included as a confound variable. 500 permutations were used. Group differences in FA, MD, RD and AxD were first tested for. Group differences in T2 were tested for by reciprocating the T2 maps to create R2 maps (T2  =  1/R2), performing 6-DOF registrations of the R2 maps to the T1-weighted images using the program flirt (Jenkinson and Smith 2001, Jenkinson et al 2002), then non-linearly registering the T1-weighted images to the native DTI space using fnirt (Smith et al 2004, Jenkinson et al 2012) (the reciprocal of the AxD image was used as a registration target) and applying the same transformations to the R2 maps. The R2 maps in native DTI space could then be fed into the TBSS analysis using the tbss_non_FA command and group differences assessed as before. R2 maps were chosen as registration targets due to their (broadly) similar contrast to a T1-weighted image. Likewise, the reciprocal of the L1 eigenvalue image has broadly similar contrast to a T1-weighted image and registration quality was excellent.

To test for differences in rates of WM ageing across tracts within the T2 FOV, the median T2 in each WM tract was calculated for each participant, using the Johns Hopkins University (JHU) WM tractography atlas (Wakana et al 2007, Hua et al 2008) at 95% probability as a primary mask, and superimposing the FA skeleton mask from the TBSS analysis as a secondary mask. Left and right hemispheric data were pooled. The effect of age and tract upon the T2 median in each tract and for each participant as defined above was then assessed by ANCOVA, as implemented in Matlab 2013b. The Bonferroni correction for multiple comparisons was also applied. The median T2 in each tract was used in preference to any other metric of 'average' since it is insensitive to outliers, and the T2 distribution in parenchyma generally deviates from normality. The median is the most robust and straightforward measure. The following 7 tracts were used in the analysis: Anterior thalamic radius (ATR), corticospinal tracts (CST), Inferior fronto-occipital fasciculus (IFOF), Inferior longitudinal fasciculus (ILoF), superior longitudinal fasciculus (SLoF), temporal part of superior longitudinal fasciculus (TSLoF) and forceps minor (FMi). These were the tracts in which differences between groups were detectable. The JHU atlas also contains the following tracts, which were not included in our analyses since the TBSS failed to detect any significant group differences: Cingulum (CG), hippocampal cingulum (CH), forceps major (FMa) and uncinate fasciculus (UF).

To examine the main sources of variances within the data, principal component analysis (PCA) was used, as implemented in Matlab 2013b. The median T2, FA, RD and AxD in the 7 WM tracts for which it was calculated were included, as well as age, total WM volume, and two measures of cognitive function, namely the total accuracy and mean reaction time from the PAL task of the CANTAB toolbox. Therefore, in total, 32 variables were included in the PCA. Variables were discarded if the magnitude of their PCA score vector in the space of the first four principal components was below the mean magnitude. Before analysis, variables were demeaned and normalized by their standard deviation.

An example R2 (=1/T2) map is shown in figure 1, indicating the data quality and field of view analyzable. Using an FA skeletonisation threshold of 0.2 and a cutoff in cluster-corrected p-values of 0.05, the mono-exponential T2 maps were able to detect regional group differences, and these are shown in figure 2. In particular, we were able to localize regions in which the T2 for the AD group was longer than that of the MCI group, and for which the T2 of the AD group was longer than that of the NC group. In no case was the T2 of the NC group longer than that of the MCI or AD group, nor the MCI group T2 longer than that of the AD group; T2 only lengthened with diminishing cognitive function.

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

The main group differences in which T2 was increased in the AD group relative to both NC and MCI were in the FMi, IFOF, ILoF, SLoF and TSLoF. The ATR and CST were also affected, though to a lesser extent. The remaining tracts of the JHU WM tract atlas were negligibly affected, if at all. The number of voxels in the FA skeleton which distinguish groups in each WM tract is plotted in figure 2(j). The main group differences in which T2 was increased in the AD group relative to the NC group but not the MCI group were bilaterally in the temporal lobe (IFOF and ILoF), and uniquely in the left medial temporal lobe superior to the hippocampal head, in the ATR and CST. Many differences were not bilateral, or at least not symmetric. In general, T2 increases relative to controls were more pronounced, affecting larger areas, on the left hemisphere, which was particularly apparent in the ILoF, SLoF, TSLoF and CST. However, T2 increases separating the AD group from the MCI group were very much more symmetric, except in the ILoF. Some additional plots are provided in the supplementary information (stacks.iop.org/PMB/61/5587/mmedia). We did not detect group differences using DTI at the 95% significance level. Whilst slightly higher diffusivities and slightly lower FA were associated with AD and MCI, these were significant only at thresholds below 95% and are not presented in our main results.

By pooling the observations in each tract and calculating the tract-wise median for each imaging parameter (T2, FA, RD, AxD), we were able to delineate age effects in different WM tracts. Figure 3 shows ANCOVA results, comparing tract-wise T2, FA, RD and AxD medians within the entire cohort. Age was the covariate, and the 7 WM tracts analysed were the groups. We sought to determine whether the imaging parameters were different between different tracts, and whether they 'aged' at different rates.

Zoom In Zoom Out Reset image size

The effect of age on T2 could be resolved at the 95% confidence level by both slope and intercept for each tract, whilst for DTI measures FA, RD and AxD we could only separate the intercepts at the 95% level. Therefore a common gradient was used for all tracts in those analyses. ANCOVA tables may be found in the supplementary information. T2 in all tracts increased with age, whilst FA always decreased. RD and AxD also increased with age. CST showed the longest T2 on average, (figure 3(a)), the highest FA (figure 3(b)) (by a smaller relative margin) and lowest RD (figure 3(c)) (again by a smaller relative margin than the T2 effect). However, the AxD were more similar in general. WM ageing rate (measured by the slope of the linear fit to tract-wise median T2 as a function of age in ANCOVA) was significantly higher than average in the ATR (p  =  0.021) and lower than average in the ILoF (p  =  0.026) (inclusive of Bonferroni correction). The ageing rates of the ILoF and ATR were also distinguishable at the 95% confidence level. Ageing rates are shown in figure 3(a).

The results of our PCA, using the median T2, FA, MD, RD and AxD in all 7 major WM tracts as well as age, WM volume, accuracy and reaction time in the PAL task are shown in figures 4 and 5. Only the variables making a substantial contribution (see methods section) to explaining the variance in the data are shown.

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

We found that 4 principal components (PC) were sufficient to explain 83% of the variance of the data (figure 5(a)), with rapidly diminishing returns thereafter. Figure 4(a) shows the contributions of different variables to the PC1–PC3 which are able to explain 78% variance. This figure shows that FA, RD and AxD make distinct contributions to explaining the variance of the data. T2 makes a similar contribution to age, which may be resolved in the PC4 (figure 4(b)). The distinct variable types (FA, RD, AxD, T2) are well resolved from one another in three PCs and the PAL variables, age and WM volume distinguished by the PC4. Although in the PC1, the diffusivity indices RD and AxD are similar, they are fully resolved in the PC2, demonstrating their distinct information content. T2 projects positively and roughly diagonally into the space of the PC1-3.

We also found that WM volume, as well the PAL performance indices, made large contributions mainly in higher PCs. The WM volume dominates the PC4, whilst age makes large contributions to the PC3 and PC4. PAL reaction time makes a small positive projection diagonally across the PC(1,3,4) space, but is largely disassociated from WM volume. Positive T2 coefficients in PC1 and PC3 are associated with negative coefficients for PAL accuracy and positive coefficients for PAL time. This means that longer T2 implies poorer pattern recognition and longer reaction times. This is of particular interest as the TBSS identified regional T2 increases to be associated with AD.

We also examined whether the unsupervised PCA, having reduced the 32-dimensional data to 4 dimensions, would reveal potential differences between participant groups. Figures 5(b)–(f) show scatter plots demonstrating that this is indeed the case, at least to some extent. Note that the three groups are not separated in PC1, but PC2-4 show better separation.

Our TBSS analysis identified regions within WM in which T2 was longer in early AD patients, whose PAL accuracy and reaction times were significantly worse than MCI or HC. Our PCA analysis, using an atlas-based labelling of WM tracts, and not limited to the regions separating participant groups, confirmed that, independent of age or WM volume, high T2 in WM is associated with both poorer PAL accuracy and longer reaction times to identify patterns. This is interesting given recent literature indicating that widespread WM changes may occur concurrently with, or even before, GM tissue loss in cognitive decline and AD (Gold et al 2010, Salat et al 2010, Agosta et al 2011, Heise et al 2011, Selnes et al 2012). Magnetization transfer (MT) saturation has also been shown to correlate negatively with age extensively in WM (Draganski et al 2011, Callaghan et al 2014). The age-related MT saturation decreases reported in the latter studies were interpreted as being due to reduction in myelin content.

We postulate that the changes in T2 we have observed are also due to microstructural degradation of WM, by which the large induced field gradients which attenuate the signal by randomizing spin phase are lost. Equally then, we follow the existing literature that AD is as much a disease of WM as GM (Sexton et al 2011, Sachdev et al 2013), even if abnormality becomes evident at different times. Our PCA analysis was able to delineate the effect of WM volume loss from T2 increase (as well as from changes in diffusivity and FA). This evidences the postulate that WM structural changes, not WM loss, explain the data. However, with the current MRI approach we cannot separate cerebrovascular alterations in WM from changes in microstructural changes, such as change in myelination.

Generally, one can anticipate more rapid spin phase decoherence as being a consequence of field perturbers, as stated in the introduction. The myelin sheath, amongst other entities, is such a structure, whose different susceptibility relative to its surroundings has received substantial attention (Wharton and Bowtell 2012, Rudko et al 2014). A short T2 in WM may then likely to be indicative of healthy tissue, whose summative field lines conspire to dephase magnetization rather quickly. As such destruction of the ordered myelin sheath may account for our T2 observations.

Capillaries and small vessels tend to lie parallel to axons (Carmeliet and Tessier-Lavigne 2005, Guthrie 2007, Eichmann and Thomas 2013), so there is likely to be an orientation-dependent dephasing related to cerebrovascular state and architecture in WM. Alteration of cerebrovascular architecture, which is known to occur which age and differently in AD and cognitive impairment in general (Gorelick et al 2011, Honjo et al 2012), may be reflected in the increased WM T2 observed here. However, the effect may be small, considering blood volume. It is also possible that reduced cerebral perfusion may lengthen coherence lifetime since molecules in tissues may be less affected by field lines created by flowing paramagnetic deoxyhaemoglobin (Kennan et al 1994, Boxerman et al 1995) (diffusion-mediated non-refocussable dephasing is distinct from the BOLD effect though). Cerebral perfusion is reasonably stable in the age range sampled (Parkes et al 2004, Biagi et al 2007), though is affected by AD (Alsop et al 2010, Wolk and Detre 2012), being reduced generally in parietal areas. The reasonable stability of WM perfusion versus the strong correlation of T2 with age argues against this mechanism. The accumulation of paramagnetic substances in tissue, such as iron (e.g. in the form of ferritin), may also reduce T2 substantially (Mitsumori et al 2007). Conversely, a loss of tissue iron may account for lengthened T2, but there is scant evidence of iron loss with age as a general phenomenon. Rather, the opposite is generally reported (Qin et al 2011, Daugherty and Raz 2015), making non-heme iron an unlikely candidate to explain our observations.

Therefore our working hypothesis is that the T2 increases observed with age, and regional disproportionate increases in AD patients, are due to WM degradation involving breakdown of the ordered myelin sheath and/or alterations to microvascular structure.

The observation of large age-related changes in T2 in the ATR and FMi, but smaller age-related changes in T2 in the ILoF in the age range sampled is evidence of different ageing rates in WM tracts, albeit derived from cross-sectional data. Equally, it is a fine demonstration of the power of quantitative T2 to detect changes which are likely to be microstructural in origin, such as Wallerian degeneration (Alves et al 2015). We interpret the finding as adding to the evidence that not all WM is the same. Different ageing rates may also affect information processing and a variety of memory processes.

The T2 was longest in the CST and shortest in the ATR, irrespective of age and cognitive state. In part, this is likely due to inherent differences in cell biology and biochemistry between the regions, such as g-ratios, axon diameters, water content etc (Wharton and Bowtell 2012). The CST and ATR are also distinguished by DTI, in particular the high FA of the CST and low AxD of the ATR. Myelinated axon diameters are heterogeneous within and across different WM tracts, with generally larger diameters in the SLoF and IFOF, but generally smaller diameters in the corpus callosum (including FMi) (Liewald et al 2014). The CST has a broad distribution of axonal diameters similar to the SLoF. However, the magnetic resonance physics are also likely to be involved. In particular, this observation is likely in part to be a consequence of coherence lifetime anisotropy, meaning that T2 (again, used interchangeably with coherence lifetime) is a function of the angle between the system under observation and the applied magnetic field (Knight et al 2015). This is the case as long as diffusion effects contribute to coherence lifetime, for they depend on magnetic susceptibility differences which create anisotropic field lines (Majumdar and Gore 1988, Borgia et al 1996, Hürlimann 1998, Sen and Axelrod 1999, Audoly et al 2003, Cho et al 2009). It is also the case in any motionally restricted systems.

Magnetic susceptibility is a tensor quantity (Haacke et al 2015), and as such depends on orientation with respect to the applied magnetic field. It can be measured by susceptibility tensor imaging (Liu 2010), and its effects have been studied extensively in gradient echo experiments. By similar reasoning, field lines at microscopic scale created by tensorial magnetic susceptibility also depend on system orientation with respect to B0, and as already alluded to make a substantial contribution to spin phase decoherence. When the long axis of a system's susceptibility tensor is parallel to the B0, coherence lifetime is maximized (Yablonskiy and Haacke 1994). When perpendicular, it is minimized. The long axis of the susceptibility tensor, in WM, is to a good approximation parallel to the tract, as has been demonstrated (Liu et al 2012). The CST, with participants in the supine position, are approximately parallel to B0, whereas the ATR is approximately perpendicular to B0.

This study has weakness both in terms of the size of clinical cohort and imaging techniques used. Firstly, the MCI and AD participant numbers were rather low, and the study was cross-sectional. Since characterizing effects of age across a reasonably broad age range was a study aim, a cross-sectional design is preferable, despite the confounds it introduces. However, we were able to observe extensive T2 effects, which were distinct from diffusion and age effects as shown by PCA. This is a promising sign that quantitative T2 is able to detect AD pathology in WM. The ability to detect the 'spread' or even existence of WM abnormality in AD by prolonged T2 may have future clinical utility. An advantage over DTI is the ease with which T2 measurements can be made with higher resolution, to which arguably some of our findings may be attributable. The drawbacks of TBSS apply (Bach et al 2014), for example that the FA skeleton is unlikely to represent the 'tract centre'. Our cohort was not gender-balanced (22/37 HC female, 4/12 MCI female, 7/9 AD female). It is not impossible that this imbalance contributes to the group differences seen in T2 and/or the failure to observe significant group differences by DTI. However, gender differences seen by DTI, in normal ageing and AD, have generally been small and heterogeneous, if seen at all (Dunst et al 2014) (leaving us probably underpowered to detect them). Only in large studies have consistent effects been detectable (Ingalhalikar et al 2014). The use of T2 to examine gender differences is poorly represented in the literature, though T2 has been reported to be consistent between males and females (Siemonsen et al 2008). Atrophy patterns between males and females are almost identical, so there are grounds to suspect that the underlying parameters are also similar (Fjell et al 2009).

Secondly, we have heuristically described loss of spin phase coherence by a mono-exponential function, from the second (24 ms) echo onwards. There is literature supporting the view that decoherence in WM is better (albeit still heuristically) described by bi-exponential kinetics (MacKay et al 1994, Whittall et al 1997, Alonso-Ortiz et al 2015). There exists a correlation between the population of the short-T2 component and luxol fast blue optical density in formalin-fixed ex-vivo human brains (Laule et al 2008). Motivated by this, if a 2-state slow-exchange model be assumed, the time constants may be interpreted as belonging to 'free water' and 'myelin-associated water' (or the Eigenvalues of a magnetization evolution matrix outside the slow-exchange limit). This has been powerful in addressing and quantifying myelin sheath damage in multiple sclerosis (amongst other applications) (Alonso-Ortiz et al 2015). However, we have an inadequate number of TE data points to extract parameters with sufficient precision for such a model to be useful, in part due to the relatively high resolution we used, and in part the exclusion of the first (spin) echo due to identical crusher gradients throughout the experiment; fast-decaying components of the kinetics would be lost before we observe (or even refocus) them. In principle, the first echo could be retained if a more sophisticated fitting approach such as the extended phase graph method were used (Weigel 2015), but we would be including only a single extra data point which may therefore not markedly increase the information content. Our DTI data processing followed similar principles to the relaxation data modelling, treating each voxel as having a single 'effective' diffusion tensor, rather than multiple slowly or non-exchanging states of water in distinct 'compartments' with distinct diffusion tensors. This is entirely normal in DTI. Spin phase decoherence, as such whatever parameters in any model for it, are somewhat pulse-sequence and magnetic field dependent. DTI suffers this to a lesser extent. Any future clinical application would need to keep this in mind.

Thirdly, our field of view in our T2 experiment only covered part of the brain, missing some of the occipital and frontal lobes (but including most regions in which AD pathology is known). Again, coverage was sacrificed for resolution. Continued improvements in MRI technology and pulse sequences may in future ameliorate this. Most major WM tracts are partially or fully covered, as well as the most of the temporal lobe and entire limbic system, cerebellum and brain stem.