Approximately 50% of patients with multiple sclerosis (MS) are cognitively impaired and with the possible exception of processing speed, memory is the most frequently involved domain.1 2 Episodic memory, or the conscious recollection of previous experience, can be quantified by psychometric tests requiring the unaided recall of recently presented word lists or visual designs. Deficits on such memory tests are consistently found in patients with MS.3^–6

Memory is also impaired in patients with other neurological disorders affecting either the mesial temporal lobe (MTL) or the deep grey matter (DGM) nuclei. The quality of such impairment has frequently been used to separate patients into two general camps: cortical and subcortical dementia. In Alzheimer’s disease (AD), a prototype cortical dementia, amnesia is caused by disruption of cortical networks involving the frontal and parietal association cortex, and their connections with the hippocampus, entorhinal cortex, perirhinal cortex, parahippocampal complex and amygdala.7 8 Patients with AD are often impaired not only in the ability to learn new information but also in retention over an interval exceeding short term memory storage.9 10 For example, patients with AD frequently show very poor recall of word lists after a 20–25 min delay.11 Different profiles of cognitive impairment are encountered in patients with diseases affecting primarily the DGM. Memory testing in Parkinson’s disease (PD) and Huntington’s disease is most often characterised by deficient attention and new learning,12 13 but relatively preserved retention and recognition memory.14 15 Studies suggest that dysmnesia in PD and Huntington’s disease is related to inefficient encoding or retrieval mechanisms whereas the consolidation process is at least relatively preserved.

MS, once regarded as a prototype subcortical dementia, is an interesting disease in that recent imaging research has revealed both cortical16 and DGM17 atrophy. Recently, hippocampus volume loss was associated with poor memory test performance in patients with MS.18 MTL and thalamic MRI spectroscopic19 and cerebral blood flow20 abnormalities have been found in MS, but their specific roles in the memory disorder of MS are not well understood. Neuropsychological studies show impairment in both new learning and delayed recall,2 yet detailed analysis usually reveals more severe deficiencies in the acquisition of new information or in retrieval mechanisms.21 Moreover, retention normalises after controlling for the depth of encoding.22

In summary, previous research suggests that memory impairment in MS may be related to two separate pathologies—dysfunction of the frontal/subcortical axis may contribute to deficient encoding and MTL dysfunction probably leads to deficient consolidation. In the present study, we used quantitative structural imaging techniques to assess MTL and DGM volumes in patients with MS. Correlation and linear regression analyses were designed to determine which tissue compartment is most strongly related to MS associated memory disorder, as quantified by consensus standard, neuropsychological testing.

METHODS

Subjects

The participants were 50 patients with definite MS diagnosed by their treating, board certified neurologist23 and 20 healthy volunteers matched on demographic characteristics. Exclusion criteria for patients were relapse or steroid pulse treatment within 6 weeks of evaluation, pre-existing medical condition known to be associated with cognitive impairment or brain atrophy (neurodegenerative disorder, cerebrovascular disease, history of alcohol abuse, etc), current major depression, as assessed by structured interview, or history of psychiatric disease predating the onset of MS. Normal controls were free of developmental delay and medical history which could impact on cognitive ability. Mean age was 45.8 (8.5) years for MS and 45.2 (10.1) years for controls. Mean education level was 14.2 (2.1) years for MS subjects and 15.0 (1.7) years for controls. The groups were also well matched on race/ethnicity (90% Caucasian for MS, 95% Caucasian for controls) and gender (88% female for MS, 70% for controls). There were no significant differences between the groups on these variables by ANOVA or χ2 test. Course breakdown for MS was 35 relapsing remitting and 15 secondary progressive by commonly accepted standards.24 Mean disease duration was 11.5 (8.8) years. Median Expanded Disability Status Scale (EDSS)25 was 3.0 (range 0–7.0).

The patients were seen in the neuropsychology clinic for routine clinical monitoring of cognitive function (n = 33) or gave consent to participate in research investigating cognitive function in MS (n = 17). Patients and controls underwent brain MRI for clinical purposes as part of our annual MRI evaluation programme. The study was approved by the SUNY Buffalo Institutional Review Board.

Brain MRI acquisition and analysis

MRI scans were performed at 1.5 T (General Electric Signa 4×/L×, Milwaukee, Wisconsin, USA) unit. The MRI protocol included dual spin echo (SE) T2 weighted image (WI), three-dimensional spoiled gradient recalled (SPGR) T1-WI and fast attenuated inversion recovery (FLAIR) sequences. The axial dual SE sequence was acquired with echo time (TE) 30/90, relaxation time (TR) 3000, number of excitations (NEX) 1, echo train length 14, field of view (FOV) 24×18, matrix 192×256, 5 mm slice thickness with a total of 28 slices, no gap. Axial three-dimensional SPGR T1-WI scan was acquired with FOV 24×18, matrix 192×256, 2.5 mm slice thickness, 70 slices, no gap, TE 7, TR 24, NEX 1, flip angle (FLIP) 30; axial FLAIR with FOV 24×24, matrix192×256, 28 slices, 5 mm slice thickness, no gap, TE 128, TI 2000, TR 8002, echo train length 22, NEX 1. Axial SE T1-WI was acquired with FOV 24×18, matrix 192×256, 28 slices, 5 mm slice thickness, no gap, TE 9, TR 600, NEX 2. Patients were positioned in the magnet according to accepted guidelines.

The image analysis was blinded to patient clinical characteristics and clinical status. As previously described, T2 lesion volume was calculated using a highly reproducible semiautomated local thresholding technique for lesion segmentation,26 and global normalised brain volume (NBV) was measured using SIENAX, as previously described.27^–29

The volume based subcortical segmentation and surface based cortical reconstruction based on the three-dimensional SPGR T1-WIs were completed using FreeSurfer software (http://surfer.nmr.mgh.harvard.edu/fswiki/FreeSurferWiki).

The volume based stream is an automated process which re-slices three-dimensional SPGR T1-WIs to approximately 1 mm3 resolutions for whole brain tissue segmentation, subcortical parcellation and quantification of specific subcortical region tissue volumes (fig 1). The stream consists of five different stages fully described elsewhere.30 Initially, MRI volumes were registered to the Talairach space and the output images were intensity normalised. At the next stage, the skull was automatically stripped off the three-dimensional anatomical data set by using a hybrid method that uses both watershed algorithms and deformable surface models. At this stage, manual intervention was needed to visualise and edit areas of skull and the areas of cortex or cerebellum that should be corrected. After skull stripping, the output brain mask was labelled using a probabilistic atlas where each voxel in the normalised brain mask volume was assigned one of the following labels: cerebral white matter, cerebral cortex, lateral ventricle, inferior lateral ventricle, cerebellum white matter, cerebellum cortex, thalamus, caudate, putamen, pallidum, hippocampus, amygdala, accumbens, third ventricle, fourth ventricle, brainstem and cerebrospinal fluid. Four variables, amygdala, hippocampus, thalamus and caudate, were selected for further analysis and corrected by intracranial volume. The scan–rescan reproducibility measurement was performed separately for five patients with MS and five healthy volunteers within 1 week and measured five times by a single blinded observer. The coefficient of variation for the total subcortical regional volume (mm3) was 3.4%. The coefficient of variation for the selected DGM and mesial temporal regions are shown in table 1. These were the most reliable MTL and DGM structures measured.

• Download figure
• Open in new tab
• Download powerpoint

Figure 1 Representative three-dimensional spoiled gradient recalled T1 weighted image showing regional brain atrophy in a patient with secondary progressive multiple sclerosis (59-year-old female with 24 years of disease duration, and an Expanded Disability Status Scale (EDSS) of 7.0) (bottom row) compared with a patient with relapsing remitting multiple sclerosis (30-year-old female with 19 years of disease duration, and EDSS of 3.0) (top row). Axial (left), sagittal (central) and coronal (right) views of different regions segmented by FreeSurfer are displayed. Colour legend: pink, grey matter; white, left cerebral white matter; neon green, right cerebral white matter; grey, brainstem; yellow, hippocampus; rose pink, putamen; blue, pallidum; green, thalamus; violet, lateral ventricles; light blue, caudate; brown, cerebellum grey matter; light brown, cerebellum white matter.

Neuropsychological testing

Patients were examined under the supervision of a board certified neuropsychologist using tests recommended by a consensus opinion publication31 and recently validated in a large sample.2 The Rao32 adapted Symbol Digit Modalities Test (SDMT)33 was used to measure information processing speed. The SDMT has shown robust correlation with whole brain atrophy and ventricle enlargement in previous MS neuropsychological research34 35 and was included to generate findings for comparison with the episodic memory tests, as well as to replicate the earlier work with the Freesurfer method.

Auditory/verbal episodic memory was assessed using the California Verbal Learning Test–second edition (CVLT2).36 The CVLT2 required learning of a 16 item word list, presented aurally, five times. Patients were asked to recall as many words as possible after each presentation. Total Learning (CVLT2-TL) was the sum of all words recalled over the five learning trials. The Delayed Recall trial (CVLT2-DR), administered 25 min later, required recalling the list without cues or repeated presentation. The recognition Discrimination Index (CVLT2-DI) was calculated by the recommended algorithm based on the patient’s ability to recognise previously presented target words with a yes response, without misidentifying non-target foils. Higher CVLT2-DI values signified higher true positive as opposed to false positive rate.

Visual/spatial memory was assessed using analogous indices from the Brief Visuospatial Memory Test–Revised (BVMTR).37 38 Patients viewed a matrix of six abstract designs, presented for 10 s. After the display was removed from view, participants rendered the designs as accurately as possible in correct locations, using paper and pencil. Total Learning (BVMTR-TL) was the number of points earned over the three immediate learning trials, and was followed by the Delayed Recall (BVMTR-DR) trial 25 min later. Finally, a yes/no recognition Discrimination Index (BVMTR-DI), similar to that of the CVLT2-DI, was calculated, using the six target and six non-target stimuli.

Research has previously shown that these traditional memory tests are reliable in MS39 and associated with atrophy of the temporal and frontal lobes.40 Z score equivalents for the cognitive tests were calculated using previously acquired normative data.2

The Beck Depression Inventory Fast Screen (BDIFS)41 and the Fatigue Severity Scale (FSS)42 were used to assess the influence of depression and fatigue, respectively. Both are well validated in MS samples.43 44

Statistical analysis

Descriptive statistics (z scores) were derived by comparing the cognitive performance of this MS sample against data derived from prior publications from our group using the same test battery.2 For MRI, we used the demographically matched sample of normal controls available for this investigation. Our primary objective was to test the hypothesis that memory dysfunction in MS is caused by degeneration of MTL compared with DGM structures. In an attempt to reduce the number of predictor variables, a priori, we selected two MTL and DGM structures. The hippocampus and amygdala were chosen because they comprise the largest area within the anterior MTL region, judged to be most important for memory.45 46 The caudate and thalamus were selected because they are the most likely subcortical nuclei to mediate cognitive function, are well studied in MS, and have shown relationships to cognitive performance in MS or other diseases.17 47 48 As shown in table 1, all of these structures were reliably measured with Freesurfer. In order to characterise the performance of the sample, the NP testing data were normalised against an age and education matched control group published previously.2 All of the neuropsychological outcomes conformed to a Gaussian distribution by the Komolgorov–Smirnov test except for BVMTR-DI. We generated linear regression models using a standard forward selection process (p to enter = 0.05 and to leave = 0.10) predicting the NP testing dependent variables on the basis of regional volumes derived from Freesurfer, controlling for the effects of age. BVMTR-DI was treated as a dichotomous variable (normal = 6) and we used ANCOVA techniques for hypothesis testing. As the right and left MTL and DGM structures were strongly correlated (r values from 0.76 for the caudate to 0.86 for the thalamus), the left and right were combined in order to further reduce the predictor variables to the smallest possible set. Regression models controlled for the effects of age, and used the MTL and DGM volumes as predictors of each NP measure. The models were then repeated including T2 lesion volume (T2-LV) and normalised brain volume (NBV) as additional predictor variables.

RESULTS

The cognitive and whole brain MRI z scores are presented in table 2. It can be seen that there is significant cognitive impairment as in previous work with the same tests,2 39 as well as abnormal thalamic48 and whole brain volume.49 Mean T2-LV was 16.9 (16.8) ml. Among the cognitive tests the smallest effects were for CVLT2, where z = −0.8, whereas the most sensitive measure was SDMT, where the mean z score was −1.7.

The mean BDIFS for patients was 3.7 (3.9), which represents borderline or mild depression.41 Mean FSS was 4.9 (1.7), representing a moderate degree of fatigue.42 There were no significant correlations between the MRI and cognitive measures and either BDIFS or FSS, and consequently we did not analyse these variables further.

Correlations between the regional brain MRI and the neuropsychological measures are presented in table 3. SDMT showed a robust correlation with the DGM, the largest effect being r = 0.62, p<0.001 for the thalamus. Correlations with episodic memory tests (CVLT2, BVMTR) were more modest, ranging from not significant to r = 0.49, p<0.001 for CVLT2-DI/amygdala. In general, the strength of association was similar between MTL and DGM structures, but slightly favouring DGM contribution. All of the significant correlations were in the expected direction with higher volumes being associated with better cognitive ability (table 4).

The first set of regression analyses yielded no additive models. In each case, a single regional MRI variable contributed to the final model after controlling for the effects of age. The direction of the correlation was always in the direction of higher volume predicting better performance. For SDMT, the thalamus was retained, the model yielding an R2 of 0.41, p<0.001.

For CVLT2, the caudate was the single MRI measure retained in the models predicting both CVLT2-TL and CVLT2-DR, with R2 values of 0.18 (p<0.05) and 0.14 (p<0.05), respectively. In contrast, the amygdala (but not the hippocampus), was retained in the model predicting CVLT2-DI, with an R2 of 0.25 (p<0.01).

BVMTR, the visual/spatial memory test, showed similar findings for the TL score in that another DGM structure, the thalamus, was the single retained predictor (R2 = 0.21, p<0.05). However, the amygdala was retained in the model predicting BVMTR-DR. Here the R2 was modest at 0.14 (p<0.05). For the recognition task BVMTR-DI, the sample was separated into normal and abnormal performance caused by the non-Gaussian score distribution. The logistic regression model also controlling for age while predicting subgroup membership retained only the amygdala (Wald = 4.4, p<0.05).

Regression analyses were repeated with the inclusion of T2-LV and NBV in the list of predictor variables. The results were largely unchanged with the following exceptions. For SDMT, T2-LV entered on the last step bringing the total R2 to 0.47 (p<0.001). For BVMTR-TL, only T2-LV was retained in the final model, yielding an R2 of 0.22 (p = 0.009).

DISCUSSION

Historically, MS has been regarded as a disease affecting the cerebral white matter with cognitive deficits resembling a subcortical dementia syndrome. In more recent years, structural imaging studies have emphasised atrophy affecting the cortex as well as deep grey matter nuclei. We endeavoured to compare correlations between neuropsychological testing and both MTL and DGM structures, and determine which region is most predictive of memory impairment in MS. Our findings show that both regions are important, and that they explain somewhat different aspects of MS memory performance, in ways that are reminiscent of the well known differences in the cognitive profiles of patients with AD and PD.50 Whereas the DGM atrophy is the primary predictor of new learning and acquisition impairment, MTL atrophy plays a more critical role in the retention of recently learned information, as measured by 20–25 min delayed recall and recognition tasks. Thus we find that MS has features of both cortical and subcortical dementia. One clinical implication is that these data support a rationale for therapies directed at both the encoding51 and consolidation52 aspects of memory in MS.

As in previous research, the SDMT was the most sensitive cognitive test in the battery2 and showed robust correlation with DGM structures, especially the thalamus.48 The reasons are not entirely clear. We have noted that the SDMT has very high reliability53 and may thus benefit from reduction in error variance. The SDMT is also a visual task requiring not only working memory and processing speed but also visual perception and scanning. The strong correlation between SDMT and third ventricle width34 35 and the thalamus48 may reflect pathology in structures affecting both visual processing and higher cognitive function.54 The thalamus is proximal to the lateral geniculate which mediates visual input from anterior visual processing structures. However, after accounting for the influence of central atrophy, we have also shown significant correlation with right frontal cortex volume.54 More work is needed to understand the relationship between SDMT and the anterior visual system in MS. In the meantime, we are impressed that this simple 90 s task correlates significantly with multiple regional atrophy measures,28 40 54 including the MTL region in this study.

Little is known about the integrity of the hippocampus and amygdala in MS, or the functional significance of atrophy in this region. However, atrophy of MTL structures such as the hippocampus and rapid forgetting are core features, if not the hallmarks, of AD.9 55 In the present study, MTL volume accounted for most variance in predicting delayed visual/spatial memory and delayed recognition in both the auditory/verbal and visual/spatial domains. Delayed recall after a 20–25 min interval is generally regarded as a measure of both consolidation and retrieval. In contrast, recognition places very little demand on retrieval and consequently may be preserved in patients with dementia stemming from diseases affecting the frontal–subcortical axis.15 56 57 In this study, as expected, both recognition memory tests were most strongly associated with MTL volume. Therefore, while the MTL atrophy is not a core feature of MS, the relationship between MTL atrophy and forgetting is significant in this disease.

One of the more surprising findings was that the amygdala accounted for more variance than hippocampal volume in the prediction of memory impairment. While this may represent something unique about MTL degeneration in MS, some research shows that amygdala volume, while not as frequently studied as the hippocampus, does show a significant association with measures of delayed memory and forgetting in AD.46 58 The reasons why the amygdala accounted for more variance in this study are unclear and merits further study using other imaging techniques.

It is widely recognised that memory tests emphasising the unaided recall of recently presented information are more sensitive in MS than tests measuring recognition, a hallmark of so-called “subcortical dementia.” In addition, recent imaging research has highlighted the role of DGM volume loss, especially regarding the caudate59 and thalamus.17 48 Of the four tests requiring retrieval in this study (total learning and delayed recall indices), three were most strongly predicted by DGM volumes. This is the first study showing a significant correlation between caudate atrophy and memory impairment in MS. The correlation between visual/spatial learning and thalamic volume was reported previously.48 In general, our findings highlight the important role of DGM atrophy in memory performance in patients with MS. Moreover, the findings also point to defective encoding and acquisition processes in MS that are most likely caused by deficiency within the so called frontal-striatal-pallidal-thalamo-cortical network.

The relatively small sample size for this study prevents a detailed analysis of the potential contribution of subject characteristics to brain atrophy, such as gender and education. One of the potential weaknesses of the study was the reliance upon an automated calculation of regional volume rather than manual tracing, which would certainly result in better accuracy. Although we did not compare the automated Freesurfer data with manual tracings in the present study, the cross validation of Freesurfer against manual traced training sets was previously performed. Fischl et al30 showed that Freesurfer automated labelling was comparable in accuracy to manual labelling, and was of sufficient sensitivity to robustly detect changes in the volume of structures that predicted onset of probable AD. FreeSurfer is a reliable method60 and has been validated histologically and manually on aged and pathological brains.61 62 Although the four structures analysed in this study were different in their size and shape, the scan–rescan reproducibility was very similar among them (table 1). The acceptable reliability does not necessarily equate with validity, and we did not confirm the accuracy of the Freesurfer volumes with volumetrics derived from manual tracings. Thalamic volume was the highest (approximately 17 500 mm3), followed by hippocampal volume (approximately 9500 mm3), caudate (7500 mm3) and amygdala (approximately 4000 mm3). It is conceivable that despite good reliability, Freesurfer measures of the hippocampus may be less accurate than of the thalamus. Nevertheless, if the accuracy and reproducibility of segmentation would have influenced the correlation results with NP measures in our study, that would have affected more smaller structures such as the amygdala than the hippocampus. The study would also be improved by inclusion of a larger sample, as well as a single control group, with whom to compare both psychometric testing and brain imaging.

Another potential limit of this study relates to the slice thickness of the three-dimensional SPGR images. While the slice thickness of ⩽1.5 mm for Freesurfer cortical thickness segmentation was recommended, slice thickness between 1.5 and 3 mm is sufficient for reliable Freesurfer subcortical segmentation (http://www.mail-archive.com/freesurfer@nmr.mgh.harvard.edu/msg06745.html). We obtained similar subcortical volumes and segmentation reproducibility on the datasets obtained with 1.5 mm slice thickness (data not published). One should also bear in mind that the linear relationships between the regional volumetric measures and cognitive testing were fairly similar, and that it would be incorrect to conclude that the correlation between the amygdala and CVLT2-DI (0.49) is significantly greater than the correlation between the thalamus and CVLT2-DI (0.34), for example. Finally, recent work has shown evidence of demyelinating63 64 lesions in the grey matter in patients with MS and we did not account for the influence of these lesions in our analysis. On the other hand, we did not observe any hyperintense or hypointense lesions in the proximity of the four examined regions.

The above notwithstanding, for the first time, we have compared the relative clinical significance of MTL and DGM atrophy and found evidence supporting significant and distinct contributions of each region to MS associated memory disorder. DGM atrophy is the primary predictor of new learning and free recall whereas MTL atrophy plays a larger role in the recognition of information presented 25 min earlier. These findings may have a bearing on evolving conceptions of MS as exemplary of both cortical and subcortical dementia.