White matter damage on diffusion tensor imaging correlates with age-related cognitive decline

Subjects.

A population sample of 106 healthy adults (55 men, 51 women), aged between 50 and 90 years (mean age 69 years) were recruited via a local general practitioner surgery by random sampling. Subjects were recruited by mail, with respondents then screened by a telephone interview for suitability. Inclusion criteria were English as the first language, white ethnicity (due to the cultural and language demands of the neuropsychological assessment), no prior psychiatric or neurologic disorder including stroke, and no contraindications to MRI.

Of the 663 contacted individuals, 408 responded, of whom 158 agreed to participate. Fifty-two individuals were excluded according to the above criteria, leaving 106 individuals enrolled in the study. One declined MRI after cognitive testing, one was unable to be positioned within the head coil (due to kurtosis of the spine), and three did not complete the DTI protocol (two due to anxiety and one due to a technical problem). All images were checked for quality, and two additional subjects were excluded due to a poor-quality scan resulting from artifacts in the image (from movement or metallic bridges on teeth). Ninety-nine scans were deemed of good quality for analysis. Of the 99 subjects for whom good-quality DTI sequences were available, 85 of these also had good-quality spectroscopy data (14 scans being unavailable for analysis due to acquisition problems or postacquisition difficulties).

MRI scanning.

All MRI scanning was performed on a General Electric 1.5 T (22 mT/m) Signa scanner. Whole-brain DTI was acquired (acquisition matrix = 96 × 96, field of view = 24; TE = minimum; TR = 7 s, maximum b value 1000 s mm⁻²) in two interleaved series of four repeats, each containing twenty-five 2.8-mm slices, with a gap of 2.8 mm, providing contiguous whole-brain coverage. Chemical shift imaging (CSI) was conducted during the same scanning session. Placement of the volume of interest (VOI) for CSI spectroscopy was conducted using T2*-weighted axial images acquired for this purpose. The VOI was placed over the white matter in the centrum semiovale, superior to the ventricles. CSI was acquired using the standard GE PROBE CSI sequence with PRESS localization, 15-mm slice thickness, 22-cm square field of view, 16 × 16 voxels, NEX = 1. The repetition time (TR) was 2 seconds, and data were acquired both at long (TE = 136 msec) and short (TE = 30 milliseconds) echo times. The VOI was a square approximately 70 × 70 mm.

DTI analysis.

Data were analyzed on an independent workstation (Sun Blade 100; Sun Microsystems, Mountain View, CA). Images were realigned to remove linear eddy current distortions using the automated image registration (AIR) software.¹⁷ Diffusion tensor elements were computed at each voxel as described by Basser et al.¹⁸ and diagonalized to determine eigenvalues and eigenvectors, from which FA and MD values were generated.

Images were segmented into gray matter, white matter, and CSF concentration maps, incorporating a correction for image-intensity nonuniformity.¹⁹ For each subject, hard segmentations of each tissue type (gray, white, CSF) were generated to include all voxels whose probability was greater than the combined probability of the remaining tissue types, i.e., the white matter segmentation included all voxels whose probability of white matter was greater than the combined probability of gray matter and CSF.

Two methods were used to determine DTI white matter. To measure whole-brain white matter, a histogram analysis was used, and for regional information, a region of interest (ROI) analysis was performed.

Histograms.

Histograms were calculated for both MD and FA maps of the white matter. MD histograms bin width was set to 4 × 10⁻⁵ between 0.0 and 0.004 (i.e., 4 × 10⁻³). FA histogram bin width was 0.01, between 0.0 and 1.0. To correct for individual differences in brain volume, each histogram was normalized over the number of voxels in the segmented images. Previous analyses of whole-brain histograms have revealed nongaussian profiles in MD data.²⁰ As the profile is nongaussian, mean and SD measures are less relevant than other descriptors of the histogram, such as peak height frequency and peak height intensity, which were used in this analysis.

ROI analysis.

Ten consecutive axial slices were selected that passed through the white matter of the centrum semiovale and the periventricular region (figure 1A). Using the white matter segmentation described above, the edge of the white matter was traced with the contour function in dispunc (David Plummer, University College London, UK). Using the same technique, the edge of the ventricles was traced in the T2*-weighted EPI image so that the large white matter ROI (large ROI) generated excluded the ventricles (see figure 1B). Mean FA and MD values within the ROI were calculated.

• Download figure
• Open in new tab
• Download powerpoint

Figure 1. Large ROIs: slices selected (A); white matter selected (B); division of white matter into anterior, middle, and posterior regions (C).

To investigate MD and FA changes across the brain, large ROIs were divided into anterior, middle, and posterior regions, using MRIcro (Chris Rorden, University of Nottingham, UK). The anterior region was defined as anterior to the genu and the posterior region as posterior to the splenium of the corpus callosum (see figure 1C). Mean FA and MD values were extracted for these ROIs. The middle and anterior ROIs primarily include the frontal lobes, the anterior portion being predominantly prefrontal cortex.

Spectroscopic analysis.

Spectroscopy data were transferred to an independent workstation (Silicon Graphics O2, Silicon Graphics, Mountain View, CA) for analysis. All data were transformed and regridded in Spectroscopy Analysis/General Electric (SAGE) to match the brain's midline. Using the LCModel,²¹ data were automatically processed. All spectra were checked and poorly fitted ones were rerun manually. Chemical shift artifact errors in the data were automatically corrected by a locally written program. Metabolite ratios were calculated for NAA/creatine (NAA/tCr), choline/creatine (tCho/tCr), NAA/choline (NAA/tCho), and myoinositol/creatine (mI/tCr). Only voxels containing pure white matter were selected for statistical analysis.

Coregistration of DTI and CSI regions.

To determine the precise location of the spectroscopic ROI in the FA and MD maps, scanner acquisition coordinates and the CSI grid were mapped into DTI space. Mean FA and MD characteristics were calculated across the spectroscopic ROI and within each section of the CSI grid. To ensure CSF contamination of mean FA and MD, maps were minimal and all image voxels with >50% CSF were excluded. This was determined by automatic alignment of each patient's T1-weighted image to the EPI T2*-weighted image without diffusion sensitization, followed by automatic image segmentation in SPM2 (Wellcome Department of Cognitive Neurology, Institute of Neurology, London, UK).

Cognitive testing.

Subjects underwent a comprehensive battery of standardized neuropsychological tests measuring the domains of premorbid intelligence, working memory, executive function, and information-processing speed. The following tests were used for each cognitive domain: premorbid intelligence, National Adult Reading Test; executive function, trails (number-letter switching minus motor speed from the Delis-Kaplan Executive Function System (D-KEFS), towers (total achievement score from the D-KEFS), letter fluency (FAS, total correct), category fluency (animals and boys names, total correct), Stroop (total correct), WI Card Sorting Test (number of categories completed); working memory, Digit Span Backwards (total correct from Wechsler Memory Scale [WMS]-III), Letter-Number Sequencing (total correct from WMS-III); information-processing speed, Adult Memory and Information Processing Battery (AMIPB) information-processing speed (total completed), revised Wechsler Adult Intelligence Scale (WAIS-R) Digit symbol (number completed from WAIS-R), grooved pegboard (time to complete).

Raw scores were transformed into Z scores (with a high score reflecting good performance) and collated to produce a mean domain score for working memory, executive function, and information-processing speed. For each variable, Cronbach's α was computed to assess how well the selected tests measured a latent construct. α Values for each group of tasks were moderate to good (>0.6), indicating that each variable measured a unidimensional latent construct.

Statistical analysis.

Bivariate correlations (Pearson's) were used to assess relationships between DTI parameters and age and cognition. Partial correlations were used to assess the relationship between DTI parameters and cognition, controlling for age and premorbid IQ. Bivariate correlations (Pearson's) were used to assess the relationship between DTI parameters and metabolite ratios from the CSI. Partial correlations were used to assess the relationship between DTI parameters and metabolite ratios, controlling for age.

DTI measures and age.

With increasing age, there was a progressive reduction in FA and increase in MD on both ROI and histogram analysis (figure 2). MD histograms showed a progressive reduction in peak height (r = 0.661, p < 0.001) and an increase in width with increasing decade. This is consistent with increasing MD and a wider spread of values with age. FA histograms showed a progressive increase in peak height with age (r = −0.625, p < 0.001), representing a reduction in FA with age.

• Download figure
• Open in new tab
• Download powerpoint

Figure 2. Whole-brain white matter mean fractional anisotropy (FA) and mean diffusivity (MD) and FA histograms by decade.

Figure 3 shows this result for the ROI analysis; the correlations between mean brain FA and age was r = −0.725, p < 0.001, and mean brain MD and age was r = 0.777, p < 0.001. Similar correlations were found for anterior, middle, and posterior white matter regions, with each region showing a progressive decline in FA and increase in MD (see figure 3, and table 1).

• Download figure
• Open in new tab
• Download powerpoint

Figure 3. Scatterplots of mean diffusivity (MD) and fractional anisotropy (FA) with age.

Cognition and age.

There were no differences in premorbid intelligence across the four decades (analysis of variance: p = 0.353). All the composite cognitive scores correlated with age, with performance declining as age increased: executive function (r = −0.507, p < 0.001); working memory (r= −0.341, p < 0.001); speed (r= −0.593, p < 0.001). The pattern of decline across the different decades is shown in figure 4 (a low Z score reflects low performance) and demonstrates that the rate of decline increased with age.

• Download figure
• Open in new tab
• Download powerpoint

Figure 4. Graph of Z scores for cognitive abilities by decade.

DTI and cognition.

Both MD and FA determined by either whole-brain white matter histogram analysis or ROI analysis correlated highly significantly with all three cognitive domains of executive function, working memory, and speed. Correlation coefficients are shown in table 2.

To determine whether DTI correlated with cognition independently of age, the analyses were repeated with age as a covariate. Premorbid IQ was also entered as a covariate. Following this, only working memory correlated with DTI measures (see table 2) with associations present with MD on ROI analysis in all three regions (anterior: r = −0.237, p = 0.019; middle: r = −0.210, p = 0.039; posterior: r = −0.259, p = 0.011) and with FA in the middle region (r = 0.254, p = 0.013). On the histogram analysis, working memory correlated with both MD normalized peak height frequency (r = −0.232, p = 0.022), and FA peak height intensity (r = −0.341, p = 0.001).

DTI and CSI parameters.

MD and FA values obtained on DTI were correlated with metabolite values obtained from CSI from the same brain regions (table 3). FA correlated with both NAA/tCr (r = 0.461, p < 0.001) and tCho/tCr (r = 0.239, p = 0.032) ratios, and MD correlated negatively with NAA/tCr (r= −0.253, p = 0.023). After covarying for age (see table 3), FA remained correlated with NAA/tCr (r = 0.349, p= 0.002).