Gray and white matter brain atrophy and neuropsychological impairment in multiple sclerosis

Subjects.

All subjects gave written consent to participate in the study, which was approved by the local Institutional Review Board. Patients with MS (n = 40) underwent brain MRI and neuropsychological testing. Whole brain volumes from the MS group were compared with a sample of age-/sex-matched normal control (NC) subjects (n = 15) who received the identical MRI protocol (NC-MR group). The neuropsychological testing performance of the MS group was compared to that of a separate, larger (n = 83) age-/sex-matched NC sample (NC-NP group), as only 7 NC-MR subjects received neuropsychological testing and a larger NC-NP sample better represented the cognitive capacities of healthy controls. While the NC-MR and NC-NP groups were matched on age, sex, and education, we again could not compare their cognitive functioning due to the small number of NC-MR subjects who had undergone neuropsychological testing (n = 7). Patients with MS were consecutive outpatient cases seen at a tertiary care university-affiliated comprehensive MS research and treatment center. Different aspects of some of the subjects were reported previously,^5,12,22,23 including the relationship between gray and white matter atrophy and neurologic disability.¹² Exclusion criteria were age less than 21 or greater than 65 years, pregnancy, other major neurologic or medical illness, past or current substance abuse, past or current psychiatric disorder other than the emergence of psychopathology following the onset of MS, or corticosteroid use or acute relapse in the previous 4 weeks. Patients with MS were permitted to remain on psychoactive drugs, such as antidepressants and anxiolytics, as part of their routine MS medical care.

The MS group (29 women and 11 men; 38 white and 2 African American; 34 relapsing-remitting [RR] MS and 6 secondary progressive [SP] MS) had an age (mean ± SD) of 42.4 ± 8.6 years (range = 23 to 61), an educational level of 15.3 ± 2.3 years (range = 12 to 20), a disease duration of 11.2 ± 8.8 years (range = 2 to 43), mild-to-moderate physical disability (Expanded Disability Status Scale [EDSS]²⁴ score = 3.3 ± 1.9, range = 1.0 to 7.5; Timed 25-Foot Walk²⁵ performance = 13.9 ± 30.7 seconds, range = 3.6 to 170 seconds), and mild-to-moderate lesion burden (T1 hypointense black hole lesion volume = 4.5 ± 5.4 mL; and hyperintense fluid-attenuated inversion-recovery [FLAIR] lesion volume = 15.7 ± 18.6 mL). This MS sample had demographic characteristics similar to those of a large population-based epidemiologic MS study.²⁶

The NC-MR group (10 women and 5 men; 13 white and 2 other) had a mean age of 39.1 ± 6.7 years (range = 29 to 53) and an average educational level of 15.8 ± 1.9 years (range = 14 to 20). The MS and NC-MR groups were not different in age (t_[53] = 1.35, p = 0.18), education (t _[51] = 0.72, p = 0.47), or sex (χ²_{[1,N = 55]} = 0.18, p = 0.67), but were different in terms of race (χ²_{[2,N = 55]} = 6.17, p = 0.04).

The NC-NP group (59 women and 24 men; 77 white, 5 African American, and 1 other) had a mean age of 43.3 ± 9.2 years (range = 21 to 59) and an average educational level of 15.0 ± 2.4 years (range = 86 to 124). No MS vs NC-NP differences existed with regard to age (t_[121] = 0.52, p = 0.61), education (t_[121] = 0.65, p = 0.52), sex (χ²_{[1,N = 123]} = 0.03, p = 0.87), or race (χ²_{[2,N = 123]} = 0.55, p = 0.76).

MRI acquisition.

Brain MRI was obtained on a Philips Gyroscan ACS-NT 1.5-T scanner (Best, The Netherlands). For whole brain gray and white matter segmentation, T1-weighted images of the brain were acquired in the coronal plane with a three-dimensional (gradient echo) technique (repetition time [TR] = 24 msec, echo time [TE] = 7 msec, flip angle = 30°, 256 × 256 acquisition matrix, 70 slices, slice thickness = 2.5 mm, no slice gap, field of view = 25 × 25 cm, and 1 signal average); this resulted in a voxel size of approximately 1.0 mm x 1.0 mm x 2.5 mm.

In addition, 2D fast FLAIR axial images (TR = 8,000, TE = 120, inversion time [TI] = 2,200, 192 × 256 matrix, 2 signal averages, 24 slices, 5 mm slices, no slice gap) were used to calculate T2 hyperintense lesion volume and to confirm the effects of MS-related lesions affecting the segmentation of gradient echo scans (see Segmentation Misclassification). A 2D T1-weighted conventional spin echo axial sequence (TR = 400, TE = 10, 192 × 256 matrix, 2 signal averages, 24 slices, 5 mm slices, no slice gap) was used to measure the volume of conventional T1 hypointense (black hole) lesions in the MS sample.

Image analysis.

Scans were transferred to a network of Sun workstations (Sun Microsystems, Inc., Santa Clara, CA). SPM99 (Wellcome Department of Cognitive Neurology, Institute of Neurology, Queen Square, London)²⁷ was used to align T1-weighted gradient echo scans in the same three-dimensional orientation, correct for magnetic field inhomogeneity, and segment brain tissue into gray, white, and CSF compartments, as detailed previously.^12,22,23 We chose SPM99 for the determination of whole brain volume, because SPM99 is a well-tested automated method,^7,8,11 is freely available from the Internet, and, in our research group, has been validated against a semiautomated method for measuring whole brain atrophy in MS.²³ The Jim software package (version 1.0, Xinapse Systems, Ltd., Northants, UK, http://www.xinapse.com) was used for measurement of hyperintense lesions on FLAIR images and hypointense lesions on T1-weighted spin echo images. One trained technician who was unaware of clinical information performed the image analysis for each of the brain measures. Third ventricle width and its relationship with neuropsychological findings has already been reported in this cohort⁵; thus, these MRI data were not used for the present study. Furthermore, our main hypothesis was to compare whole brain gray and white matter atrophy to whole brain lesion measures in terms of their association with neuropsychological findings.

Segmentation.

In SPM99, T1-weighted gradient echo MR scans were aligned by placing the anterior commissure (AC) at the SPM99 origin and centering the x-axis of the origin to pass through the posterior commissure (PC) in the mid-sagittal plane. Brain scans were aligned by matching the interhemispheric fissure to the origin’s x- and y-axis in the anterior-posterior and superior-inferior planes. Next, in one step, realigned scans were stripped of extracranial tissue (e.g., skull, orbit, outer meningeal tissue) and segmented into separate gray matter, white matter, and CSF images (figure 1) using inhomogeneity correction (maximum level). Each segmented image then was masked to eliminate extracranial artifact stemming from the segmentation procedure (see figure 1). Final whole brain volume measurements for gray matter, white matter, and CSF were based on the segmented, masked images from native three-dimensional brain scans, after corrections were made to account for tissue compartment misclassification due to MS lesions (see Segmentation Misclassification).

• Download figure
• Open in new tab
• Download powerpoint

Figure 1. T1-weighted gradient echo MR scans and obtained gray, white, and CSF tissue compartments after automated skull-stripping and segmentation by SPM99 from a normal control. Images from upper left, upper right, lower left, and lower right are as follows: 1) raw T1-weighted image; 2) white matter only (bright areas); 3) gray matter only (bright areas); 4) CSF only (bright areas) with background brain parenchyma (inner dark areas).

For each tissue compartment scan generated by SPM99, the raw tissue compartment volume was automatically calculated and used as an initial uncorrected estimate (see Segmentation Misclassification) of white matter volume (uWMV), gray matter volume (uGMV), CSF volume (uCSFV), intracranial volume (ICV) (ICV = uWMV + uGMV + uCSFV), and brain parenchymal volume (uBPV) (uBPV = uWMV + uGMV), and was used to compute uncorrected normalized white matter fraction (uWMF) (uWMF = [uWMV/ICV]), gray matter fraction (uGMF) (uGMF = [uGMV/ICV]), brain parenchymal fraction (uBPF) (uBPF = [(uWMV + uGMV)/ICV]), and CSF fraction (uCFSF) (uCSFF = [uCSFV/ICV]). These volume estimates were corrected subsequently for tissue misclassification related to MS lesions (see Segmentation Misclassification).

T2 hyperintense lesions.

FLAIR scans were used for analysis of hyperintense lesions on T2-weighted images. A threshold technique was used to determine the total brain lesion load. Extracranial tissue was first removed using a masking tool which involves an automated contour-tracing tool designed to trace the brain contour, given a user-specified point along the cortical surface. A second part of the masking procedure involves removing non-lesion extra-axial hyperintensities from within the brain surface contour, primarily the result of FLAIR artifacts in the ventricles and subarachnoid space. A threshold then was applied to separate hyperintense lesions from non-lesion tissue. Through development of this analysis technique, the optimal threshold was determined from a 24.00 mm² region-of-interest (ROI) placed in the intensity sampling of the frontal white matter, as seen on the superior-most axial slice containing the central portion of the lateral ventricles. The value of the threshold was defined as the mean + 1.5 SD, as this value proved most effective at parsing out lesions. Hyperintensities of an axial cross-slice area less than 16.0 mm² were excluded. The Jim software package then automatically calculated total brain FLAIR hyperintense lesion volume (FLLV) by multiplying lesion area by slice thickness.

Hypointense lesions.

Analysis of hypointense lesions on T1-weighted spin echo images was performed using a semi-automated edge finding and local thresholding technique. Hypointensities on T1-weighted images were defined as lesions in the brain parenchyma that were of lower signal than white matter and were also at least partially hyperintense on FLAIR images. Thus, normal cystic structures (e.g., perivascular spaces) or part of CSF spaces (e.g., sulci) were excluded. The lesion had to be measurably darker than surrounding brain parenchyma, as defined by the ability of the lesion contour to be detected by the edge-finding tool inherent in the software package. To trace the lesion, the operator clicked on the edge of hypointense area, and the program examined a region 5 × 5 pixels around the mouse click and computed the maximum intensity gradient within the region. The pixel with the highest intensity gradient then was used as a starting point for contour following, thus outlining the region where the intensity was locally lower than at the starting pixel. Total T1 hypointense lesion volume (T1LV) then was calculated automatically by the Jim software program as the sum of the area of all lesions multiplied by the slice thickness.

Reliability/variability.

While SPM99 uses an automated algorithm for scan segmentation, manual input is needed to realign scans and to specify the SPM99 AC origin and the AC-PC line (see Image Analysis). Therefore, we examined the effect of operator input with various indices of reliability and variability using randomly selected MS and NC-MR cases. As shown by coefficients of variation (CV = [SD/mean] x100), intrarater reliability was high for uWMV (CV = 0.03%), uGMV (CV = 0.03%), uCSFV (CV = 0.59%), ICV (CV = 0.06%), uBPV (CV = 0.02%), uWMF (CV = 0.06%), uGMF (CV = 0.06%), uCSFF (CV = 0.54%), and uBPF (CV = 0.06%) using the same rater on six repeated cases (three patients with MS and three controls). In addition, inter-rater reliability was high for uWMV (CV = 0.11%), uGMV (CV = 0.04%), uCSFV (CV = 1.02%), ICV (CV = 0.13%), uBPV (CV = 0.05%), uWMF (CV = 0.18%), uGMF (CV = 0.10%), uCSFF (CV = 0.91%), and uBPF (CV = 0.11%) using two different raters on seven cases (four patients with MS and three controls). Scan-rescan variability (stability) based on repeated MRI acquisitions (1 week apart) for two normal controls (a 26-year-old man and a 36-year-old woman) showed excellent test-retest (scan-rescan) reproducibility for uWMV (CV = 0.83%), uGMV (CV = 0.83%), uCSFV (CV = 0.99%), ICV (CV = 0.34%), uBPV (CV = 0.27%), uWMF (CV = 1.18%), uGMF (CV = 0.48%), uCSFF (CV = 0.65%), and uBPF (CV = 0.07%). For the lesion measurements, CV (n = 10 reanalyzed cases) for intrarater and inter-rater variability was 1.2% and 3.1% for FLLV, and 1.7% and 4.5% for T1LV.⁵

Segmentation misclassification.

We examined whether MS-related hypointense lesions on the T1-weighted gradient echo images would cause errors in the SPM99 segmentation process. Prior studies with gradient echo images^22,23 have shown that the majority of such hypointensities are misclassified by SPM99 as gray matter. In other words, the reduced signal intensity associated with MS-related hypointensities may result in misclassification of voxels by SPM99 to a given tissue compartment, and therefore introduce error in the initial (automated) SPM99 estimates of gray matter, white matter, CSF, and total parenchymal volume (uWMV, uGMV, uCSFV, and uBPV) and the corresponding brain fraction measures (uWMF, uGMF, uCSFF, and uBPF). As a result, we performed a segmentation misclassification analysis¹² (see appendix E-1 for method details [available on the Neurology Web site at www.neurology.org]) that produced lesion-corrected volumes and fractions for gray matter (cGMV and cGMF), white matter (cWMV and cWMF), total parenchyma (cBPV and cBPF), and CSF (cCSFV and cCSFF). As the results of the segmentation misclassification analysis were highly significant (supplementary appendix E-1), only corrected whole brain volumes and fractions were used in the statistical analyses.

Neuropsychological and neuropsychiatric assessment.

Neuropsychological testing was conducted blind to MRI results (R. Benedict). The median time interval between cognitive testing and the MR scan was 34 days (range 0 to 358 days) for patients with MS. The neuropsychological tests conformed to consensus panel recommendations² regarding the assessment of processing speed and memory. Rao adaptations²⁸ of the Paced Auditory Serial Addition Test (PASAT)²⁹ and Symbol Digit Modalities Test (SDMT)³⁰ were used to evaluate mental processing speed and working memory. The PASAT involved rapid additions of successive single digits presented aurally at 2 and 3 seconds interstimulus intervals, and the total correct for both trials was used in the analyses. With the SDMT, patients were asked to voice a number matching an associated symbol presented in a key at the top of an 8½ × 11 inch sheet. For each test, the dependent measures were the total number of correct responses. Auditory/verbal learning and memory were assessed with the second edition of the California Verbal Learning Test (CVLT-II).³¹ The CVLT-II required the learning of a 16-item word list, presented aurally, five times. Patients were asked to recall as many words as possible after each presentation. CVLT-II Total Learning was the sum of all words recalled over the five learning trials. The CVLT-II Delayed Recall trial, 25 minutes later, required recalling the list without cues or repeated presentation. The Brief Visuospatial Memory Test-Revised (BVMT-R)³² measured visual/spatial learning in a similar format. Patients viewed a matrix of six abstract designs, presented for 10 seconds. After the display was removed from view, participants rendered the designs as accurately as possible and in correct locations, using paper and pencil. BVMT-R Total Learning, the number of points earned over the three immediate learning trials, was followed by the Delayed Recall trial 25 minutes later. Research has shown that these tests are reliable^31,32 and sensitive to the effects of MS. The Neuropsychiatric Inventory (NPI),³³ an informant-based structured interview, was used evaluate psychopathology potentially associated with MS within the following domains: Anxiety, Agitation, Irritability, Dysphoria, Apathy, Disinhibition, and Euphoria. Two domains (Hallucinations and Delusions) were removed from consideration due to the rarity of the problem in MS, and another (Aberrant Motor Activity) due to potential overlap with motor disorder, which is common in the illness. Behaviors present throughout life but worsening with illness progression were considered, but unchanging chronic behavior patterns were not. Informants were asked about the patient’s behavior over the past 4 weeks. For each domain, informants were asked to judge the frequency (4-point scale) and severity (3-point scale) of the target symptom/behavior. Total composite scores for each domain were calculated in accordance with standardized instructions (frequency x severity). If the symptom was not present, a score of 0 was recorded. Depression level in the MS sample was evaluated with the Beck Depression Inventory (BDI).³⁴

Statistical analyses.

χ² and t tests were used to examine group differences with respect to the demographic variables. Given that regression-based head size (ICV) normalization has advantages over brain fraction-based methods,²² we examined group differences in predicted volumes for cGMV, cWMV, cCSFV, and cBPV using a one-factor analysis of covariance (ANCOVA) design with ICV and age as covariates. Group differences in brain parenchymal fraction data also were reported to allow comparison with prior data based on traditional analyses and were tested with a one-factor ANCOVA design with age as a covariate. For the neuropsychological and neuropsychiatric variables, a one-factor ANCOVA design was used to examine group differences in cognitive performance (adjusted for age and education) and symptom expression (adjusted for age).

For exploratory purposes, hierarchical stepwise multiple regression analysis was performed to determine if the candidate MRI variables (cGMV, cWMV, T1LV, and FLLV) predicted any of the neuropsychological (CVLT-II, BVMT-R, PASAT, and SDMT) and neuropsychiatric variables (NPI domains and BDI), after first accounting for covariates by forced entry (i.e., ICV, age, education, and the BDI for the neuropsychological variables; ICV and age for the neuropsychiatric measures). The multiple regression analyses did not include variables assessing whole brain parenchymal volume (cBPV) and whole brain atrophy (cCSF), because the effects of whole brain atrophy, which were reported previously,⁵ potentially may obscure relationships with gray and white matter volume, given the part-to-whole relationship among the brain compartments (also see below). All statistical analyses had a two-tailed alpha level of < 0.05 for defining significance.

Due to the nature of the segmentation method used by SPM99 (individual voxels are assigned probability values for the gray, white, and CSF compartments, and their sum adds up 1), statistical analyses yielded results that were identical but opposite in direction for cBPV (cGMV + cWMV) and cCSFV, because cBPV and cCSFV share identical variance due to their probability arithmetic relationship of cBPV + cCSFV = 1 for each voxel (the same is true for uBPV and uCSFV, as well as for the corresponding brain fraction measures). Therefore, to eliminate redundancy in the results, cCSFV was not used in any of the statistical analyses.

Demographic variable analyses.

Partial correlations (adjusted for ICV) for the MS group revealed that age was negatively related to cGMV (r_[38] = –0.40, p = 0.01) and showed a trend to association with cBPV (r_[38] = –0.28, p = 0.08), but was unrelated to cWMV (r_[38] = 0.03, p = 0.84). For patients with MS, educational level was not correlated with any of the corrected brain volumes after adjusting for ICV (all p > 0.10). In addition, none of the corrected brain volumes was related to sex (all p > 0.10).

MS–NC differences.

The one-way group ANCOVA results (adjusted for ICV and age) and the whole brain volume group means revealed that the patients with MS had lower cBPV (F_[1,51] = 10.58, p = 0.002, 1,090 ± 46 mL, –4.1%, d = –1.00), cGMV (F_[1,51] = 7.47, p = 0.009, 708 ± 33 mL, –3.7%, d = –0.84), and cWMV (F_[1,51] = 5.93, p = 0.018, 382 ± 25 mL, –4.7%, d = –0.75) relative to controls (cBPV: 1,136 ± 47 mL; cGMV: 736 ± 33 mL; cWMV: 401 ± 26 mL). As expected, no group differences were found with regard to ICV (p = 0.46). Thus, the MS group had reductions in whole brain gray and white matter volume after correction for misclassification. Similar MS vs NC results also were found with the brain fraction data (cBPF: 0.830 ± 0.036 vs 0.870 ± 0.037, p = 0.001, –4.6%, d = –1.10; cGMF: 0.540 ± 0.027 vs 0.565 ± 0.027, p = 0.004, –4.4%, d = –0.93; cWMF: 0.290 ± 0.019 vs 0.305 ± 0.020, p = 0.015, –4.9%, d = –0.77), which were reported to allow comparison with prior data based on traditional analyses.

For group differences with respect to neuropsychological test performance (table 1), patients with MS had markedly impaired cognitive performance on tasks assessing verbal memory (CVLT-II), visuospatial memory (BVMT-R), and processing speed/working memory (PASAT; SDMT) when compared to controls (all p < 0.0001). As a subset of the NC-NP group also had NPI (n = 43) and BDI ratings (n = 44), group differences were examined regarding the expression of neuropsychiatric symptoms. Relative to the NC-NP group, the MS group had significantly higher BDI and total composite domain scores from the NPI (Anxiety, Agitation, Irritability, Dysphoria, Apathy, and Euphoria), with the exception of Disinhibition (p = 0.11).

MRI–neuropsychological relationships.

Hierarchical stepwise multiple regression analysis (table 2) was performed to determine if the candidate MRI variables (cGMV, cWMV, T1LV, and FLLV) predicted any of the neuropsychological and neuropsychiatric measures, after first accounting for important covariates. For neuropsychological performance, cGMV (after first adjusting for the covariates of ICV, age, education, and the BDI) was the only candidate MRI variable that best predicted short- and long-term auditory/verbal memory performance, as assessed by CVLT-II Total Learning (partial r = 0.33) and Delayed Recall (partial r = 0.34). On the other hand, cWMV (after first accounting for the above four covariates) was the main candidate MRI variable that best predicted performance on the SDMT (partial r = 0.59), PASAT (partial r = 0.41), and BVMT-R Delayed Recall (partial r = 0.38), although BVMT-R Total Learning was unrelated to whole gray matter, white matter, and lesion volume. Furthermore, in the regression model for the SDMT, FLLV also accounted for additional unique variance (Step 3, partial r = –0.47), separate from the stronger effect of cWMV (Step 2, partial r = 0.59). Otherwise, the lesion load measures (T1LV and FLLV) were not significant candidate MRI variables with any of the neuropsychological test measures. Graphs of these findings indicated that these relationships were uniform and with minimum artifact (see figure 2, A and B, and figures E-1A, E-1B, and E-1C).

• Download figure
• Open in new tab
• Download powerpoint

Figure 2. Selected partial correlation scatterplots of the results obtained from the hierarchical stepwise multiple regression analysis (see table 2), wherein the candidate MRI variables (whole gray matter, whole white matter, T1 hypointense lesion, and FLAIR hyperintense lesion volume) were tested for their ability to predict the neuropsychological and neuropsychiatric measures (all remaining scatterplots are provided in figures E-1A through E-1F). Residuals of the x- and y-axis variables are based on separate regression equations in which the considered variable is regressed with all covariates. Higher x-axis values represent larger brain volumes, and higher y-axis values represent better cognitive performance or greater neuropsychiatric disturbance. (A) The plot of corrected whole brain white matter volume with the Symbol Digit Modalities Test (Step 2, partial r_[26] = 0.59, p < 0.001) after first adjusting for the covariates of intracranial volume (ICV), age, education, and the Beck Depression Inventory (BDI). (B) The plot of corrected whole gray matter volume with the California Verbal Learning Test-II (CVLT-II) Total Learning score (partial r_[33] = 0.33, p = 0.050) after first adjusting for the covariates of ICV, age, education, and the BDI. (C) The plot of corrected whole gray matter volume with the Neuropsychiatric Inventory (NPI) Total Composite Euphoria Scale score (partial r_[26] = −0.55, p = 0.003) after first adjusting for the covariates of ICV and age.

For neuropsychiatric symptoms (adjusted first for the covariates of ICV and age), cGMV was the sole candidate MRI variable that best predicted Euphoria (partial r = –0.55) and Disinhibition (partial r = –0.46). In addition, T1LV was associated with Dysphoria (partial r = 0.54) and Agitation levels (partial r = 0.38) (but see below). Finally, whole gray matter, white matter, and lesion volumes were unrelated to Anxiety, Irritability, Apathy, and BDI levels. While graphs of the cGMV findings revealed acceptable relationships for Euphoria (figure 2C) and Disinhibition (supplementary appendix E-1D), the T1LV results for Dysphoria and Agitation were heavily influenced by outlier observations (supplementary appendix E-1E and E-1F). After performing power transformations on these latter variables ([T1LV + 1]*10^0.35, [Dysphoria + 1]*10^0.03, and [Agitation + 1]*10^0.20) and repeating the regression analysis, T1LV was not related to Dysphoria (partial r = 0.32, p = 0.09) and Agitation (partial r = 0.10, p = 0.61) after adjusting for ICV and age. As a result, these relationships were not considered in the discussion.