Multimodal magnetic resonance imaging in relation to cognitive impairment in neuromyelitis optica spectrum disorder

Cognitive impairment (CI) is reported in 29–57% of patients with neuromyelitis optica spectrum disorder (NMOSD). However, the pathophysiology underlying CI in NMOSD is poorly understood. The present study aims to investigate the predictive values of various conventional and quantitative MRI parameters for cognitive performance in patients with NMOSD. Neurological assessment and conventional, diffusion tensor, and volumetric MRI sequences were collected form 73 patients with NMOSD and 44 healthy controls (HCs). Patients with ≥3 failed tests were considered to have CI. Brain lesion load, gray matter (GM) and white matter (WM) atrophy, deep GM (DGM) atrophy, cortical thickness, and diffuse microstructural WM damage were assessed. Twenty-three (32%) patients with NMOSD had CI. Compared to cognitively preserved (CP) individuals, patients with CI had atrophy in the WM, thalamus, and caudate, decreased fractional anisotropy (FA) and increased mean diffusivity in their WM. A multivariate model indicated that mean FA values in the WM and volume in the nucleus accumbens (NAc) were associated with overall cognition (p = 0.002 and p = 0.008, respectively). Diffuse microstructural damage in the WM and DGM atrophy in the NAc are the strongest predictors of cognitive impairment in patients with NMOSD.

Hippocampal atrophy was associated with poor performance in immediate recall of visual memory (p = 0.006). SDMT scores positively associated with mean FA values (p = 0.004), BPF (p = 0.044), and volumes of thalamus (0.045), putamen (0.015), NAc (p = 0.01), and caudate (p = 0.005). Finally, PASAT scores were positively associated with mean FA values (p = 0.027) and NAc volume (p = 0.01). We found no associations between the hyperintense lesion volume on FLAIR images and cortical thickness or cognitive performance. Table 5 summarizes the results of the multivariate regression analysis performed to identify MRI variables significantly associated with overall global cognitive z-scores and performance in specific cognitive domains. Based on comparisons between CP patients and those with CI and univariate correlate analyses, different candidate predictors for multivariate analyses were used for each cognitive function such as WMF, DGM volume (in the thalamus, putamen, NAc, hippocampus, and caudate), FA, and MD in NAWM. A multivariate model indicated that FA of NAWM and accumbens atrophy were independently associated with global cognitive z-scores (p = 0.002 and p = 0.008, respectively). NAc atrophy was associated with performance on the PASAT (p = 0.007) and immediate recall of visual memory (p = 0.028). FA abnormalities were associated with worse performance in immediate recall (p ≤ 0.001) and delayed recall (p < 0.001) of visuospatial memory and SDMT (p = 0.004). A weak associated was also observed between caudate atrophy and performance on the Stroop interference test (p = 0.018).

Discussion
By combining voxel-wise analysis methods and a multiparametric structural MRI approach, which included measures of both irreversible tissue loss and microstructural tissue damage, we show that CI in NMOSD is the result of a complex interplay between NAWM and GM damage. We found that 32% of patients with NMOSD have CI, with core deficits in attention/information processing speed and visuospatial episodic memory despite preserved verbal episodic memory and verbal fluency. In patients with NMOSD, diffuse white matter microstructural damage (mean FA value in NAWM) and DGM (especially NAc) atrophy are the strongest predictors for CI. Conversely, brain lesion load, total brain volume, and cortical thickness were not associated with CI. This suggests that the role of conventional imaging measures in assessing cognitive impairment is limited.
We have previously reported a lower prevalence of CI and different patterns of CI in patients with NMOSD compared to patients with multiple sclerosis (MS). Specifically, patients with MS have more widespread CI with greater impairment, particularly in verbal episodic memory 4 . CI correlates with lesion burden, brain atrophy, and cortical thickness in patients with MS, while these MRI parameters are not significantly associated with cognitive function in patients with NMOSD 4, 12 . The differences in the frequency and pattern of CI and the different MRI substrates of CI between the two diseases might suggest different pathophysiologies of brain injury in the two diseases. Several studies of patients with MS have attempted to identify relevant MRI substrates using multiparametric structural MRI approaches. Diffuse WM microstructural damage and deep gray matter atrophy, particularly in  Table 2. Comparison of cognitive scores of healthy controls and patients with NMOSD. Abbreviations: BFI = Brief Fatigue Inventory; EDSS = Expanded Disability Status Scale; HCs = healthy controls; NA = not applicable; NMOSD = neuromyelitis optica spectrum disorder; PHQ-9 = Patient Health Questionnaire-9; SD = standard deviation; SE = standard error. All comparisons of cognitive function were corrected for age, educational level, and depression level. * p < 0.05 compared to healthy controls. ** p < 0.01 compared to healthy controls.
the thalamus and hippocampus, are consistently reported as strong predictors of cognitive impairment in MS [13][14][15] . We found that diffuse WM damage also contributes to CI in NMOSD. Nevertheless, observation of diffuse WM damage, even in CP patients, suggests that CI is not evident until a certain threshold level of WM damage occurs. It is worth noting that we found different patterns of subcortical atrophy in patients with NMOSD compared to previous studies of patients with MS. For instance, NAc atrophy was prominently correlated in NMOSD patients with CI. The associations between thalamic, WM or global brain atrophy, and cognition in the univariate analysis were consistent with the results of previous studies of NMOSD 8,9 . However, these associations were not significant in the multivariate analysis of the current study. These findings suggest that the use of a single MRI modality is not sufficient to properly explain cognitive impairment in NMOSD. Recently, Liu and colleagues suggested that hippocampal atrophy is the main MRI predictor of cognition in NMOSD using a multiparametrical MRI approach 7 . However, we found no significant correlations between hippocampal atrophy and CI. No relationships between brain lesion load or cortical thickness and cognition in patients with NMOSD in the current study were consistent with Liu's study 7 . The discrepancies may be explained by differences in the clinical characteristics of the patients enrolled, the methods used for neuropsychological evaluation, and the statistical analysis methods. The discovery of a relationship between NAc atrophy and cognition in patients with NMOSD is novel and warrants further consideration. AQP4 expression has been shown to be higher in the hypothalamic and periventricular areas of the brain 16 . The NAc is a region in the basal forebrain rostral to the preoptic area of the hypothalamus and has important functional relationships with the hypothalamus 17 . The NAc's location may make it more susceptible to NMOSD-related brain injury. NAc plays a central functional role in positive emotional responses. In fact, patients with mood disorders, such as depression, have NAc atrophy 17 . In our cohort, patients with CI had higher depression scores than CP patients. Consistently, a recent study indicates that 28% of patients with NMOSD suffer from unrecognized depression and that the severity of depression is not associated with physical disability or disease duration 18 . Furthermore, the NAc is involved in the generation of motivated behavior and NAc atrophy is suggested to underlie the apathy of patients with Parkinson's disease 19 . Thus it is conceivable that the decrease in desire induced by NAc atrophy might also contribute to the impairment observed in attention/ information processing speed in patients with NMOSD.
Our study has limitations. First of all, this study was a single-center study, which may have introduced selection bias when compared to a population-wide study. We selected a priori a set of imaging modalities (3D T1-weighted and DTI) and thus cannot rule out that the use of other imaging modalities (e.g., myelin water imaging or functional MRI) would have led to different results. In addition, as there is no neuropsychological battery developed specifically for NMOSD, evaluation using the brief battery in the current study may limit the comprehensive evaluation of CI. As our neuropsychological battery has not been validated elsewhere, direct comparison of our results with other studies warrants caution. Finally, NMOSD pathology in this study was not  directly observed; inferences about these pathologies were based on MRI findings. Future research will be necessary to verify or refute our hypothesis of MRI findings corresponding to the underlying pathologies. Additionally, future research is needed to clarify the relationships of longitudinal MRI changes with cognitive decline. The development and validation of useful MRI biomarkers for CI will enable MRI to be utilized as a powerful tool for the evaluation of therapeutic efficacy in NMOSD clinical trials.

Methods
Patients. Seventy-three patients with NMOSD were recruited prospectively from the Department of Neurology of the National Cancer Center (NCC) in Korea between December 2013 and June 2015. NMOSD was diagnosed according to the 2006 Criteria for NMOSD 20 or one of the cardinal symptoms of NMO with seropositivity for anti-AQP4 antibodies 21 . Eligible individuals were aged 20-50 years, relapse-and steroid-free for at least 6 months, and had undergone both, MRI and neuropsychological tests. Exclusion criteria included co-morbid neurological diseases other than NMOSD, alcohol or other substance abuse, and a history of psychiatric illness, except for stable depressive symptoms. Forty-four age-and sex-matched healthy controls (HCs) without a history of neurological or psychiatric illness or substance abuse were recruited. The following parameters were investigated: current age, gender, disease duration, time interval from disease onset to treatment initiation, anti-AQP4 antibody status, and disability, as measured by the Expanded Disability Status Scale (EDSS). This study was approved by the Institutional Review Board of the NCC. Written informed consent was obtained from all patients. All experiments were performed in accordance with relevant guidelines and regulations.
Neuropsychological assessment. All patients with NMOSD and HCs underwent neuropsychological tests to assess cognitive function. The neuropsychological battery comprised tests for verbal/visual memory, visuospatial function, attention, information processing speed, language, and working memory 4 . Verbal learning and memory were assessed using the Seoul Verbal Learning Test (SVLT) 22 and the Korean version of the Hopkins Verbal Learning Test-revised (HVLT-R) 23 . The visual memory score was composed of scores from the immediate recall and delayed recall portions of the Rey Complex Figure Test (RCFT) 24 . The Controlled Oral Word Association Test (COWAT) 25 was administered to assess associative verbal fluency for phonemic stimuli. The Symbol Digit Modalities Test (SDMT) 26 , the Paced Auditory Serial Addition Test (PASAT) 27 at 3 seconds and Stroop Color and Word tests 28 were used to assess sustained attention and the speed of information processing. In addition, the Digit Span 29 test was administered to assess working memory 30 . Each subject completed a questionnaire regarding depressive symptoms (Patient Health Questionnaire-9, PHQ-9). Because depression itself  Table 4. MRI measures of healthy controls, the cognitively preserved and cognitively impaired patient groups. Abbreviations: BPF = brain parenchymal fraction; FHLV = FLAIR-hyperintense lesion volume; GMF = gray matter fraction; HCs = healthy controls; NA = not applicable; NMOSD = neuromyelitis optica spectrum disorder; SD = standard deviation; SE = standard error; WMF = white matter fraction. All comparisons were corrected for age. Post hoc analysis was performed using Bonferroni correction. * p < 0.05 compared to healthy controls. ** p < 0.01 compared to healthy controls. *** p < 0.001 compared to healthy controls. # p < 0.05 compared to cognitively preserved patients. ## p < 0.01 compared to cognitively preserved patients.
negatively impacts cognitive performance, each patient's raw score was compared to the mean value obtained from HCs matched for PHQ-9 level using z-scores. The z-scores were calculated for each neuropsychological score using the following formula: (patient's score -mean value from HCs matched for PHQ-9/standard deviation from the matched HCs). The z-scores for specific cognitive domains were combined to generate a global cognitive z-score. Patients were considered to have CI if they scored lower than the fifth percentile of the HCs in at least three domains 4 . The other patients were classified as CP.
Image acquisition. Imaging was performed using a 3.0-Tesla MR scanner (Philips Achieva; Philips Medical Systems, Best, Netherlands). Brain MRI included structural 3-dimensional T1-weighted scans for volumetric and cortical thickness analysis, axial 2-dimensional T2-weighted and fluid attenuated inversion recovery (FLAIR) scans for lesion detection, and 32-direction diffusion tensor imaging (DTI). The T2, T1, FLAIR, and DT images were acquired using the following sequences: (1)  Volumetrics. The 3D T1 images were corrected for intensity non-uniformity artifacts 31 and normalized to the MNI 152 template using linear transformation 32 . An artificial neural network classifier was used to identify GM, WM, and cerebrospinal fluid. To avoid mistakenly counting lesions as GM volume, we calculated corrected GM volumes and WM volumes as follows: corrected GM volume = uncorrected GM volume -lesion volume in WM, and corrected WM volume = uncorrected WM volume + lesion volume in WM. We divided brain volume by total intracranial volume (ICV) to adjust head size differences. Gray matter fraction (GMF) and white matter fraction (WMF) were calculated as follows: GMF = corrected GM volume/ICV, and WMF = corrected WM volume/ICV, where ICV = brain parenchymal volume + cerebrospinal fluid volume 33 .
After performing an intensity inhomogeneity correction, the images were segmented into 14 DGM structures, including the bilateral nucleus accumbens (NAc), amygdala, caudate, hippocampus, pallidum, putamen, and thalamus using the Oxford Centre for Functional MRI of the Brain (FMRIB) Integrated Registration and Segmentation Tool (FIRST) 34 . The shapes and appearances of the FIRST models were determined based on point distributions through the parameterization of surface meshes built from labels of manually segmented images. Intensity information along each vertex of the surface was also modeled. The models are used to search for optimal modes of shape variation in a novel image by estimating proper Bayesian probabilities. Each voxel in the novel image is assigned an appropriate label. Using the labeled image, the DGM volume corresponding to the label is calculated. The right and left volumes for each DGM structure were averaged and normalized in SIENAX by V-scaling, which consists of calculating the ratio of the subject's head size to a standard head size to reduce the effects of head size differences between subjects 35 . In order to perform V-scaling, the brain tissue of the subject is first extracted from the subject's raw image. The brain and skull images are then used to perform V-scaling between the subject's image and the standard space.

Lesion load.
To analyze the volumes of hyperintense lesions on T2-weighed images, FLAIR images were examined using a previously described thresholding technique 36 . The results of all automated segmentations were visually inspected by a rater experienced in anatomy and volumetry. When necessary, the segmentation was manually corrected using ITK-Snap (www.itksnap.org).  Cortical thickness. The surfaces of the inner and outer cortices were extracted automatically using the Constrained Laplacian-based Automated Segmentation with Proximities algorithm 32,37 . Cortical thickness was defined as the Euclidean distance between the linked vertices of the inner and outer surfaces.
DTI data processing. DTI processing was performed using FMRIB Software Library (FSL) ver. 5.0.2.1 (http://www.fmrib.ox.ac.uk/fsl). Voxel-wise statistical analyses of fractional anisotropy (FA), mean diffusivity (MD), axial diffusivity (AD), and radial diffusivity (RD) were carried out using Tract-Based Spatial Statistics (TBSS) 38 . In order to specifically explore the contribution of microstructural NAWM damage, individual lesion masks after normalization to the FA skeleton space were averaged for each group to obtain lesion probability maps. The analyses included only those WM voxels where the lesion probability was <5%, as the voxels with a lesion probability ≥5% were set to zero 39 . We extracted mean FA, MD, RD, and AD values for global NAWM in each subject.

Statistical analysis.
To compare the clinical characteristics of the different groups, analyses of variance (ANOVAs) was used for continuous variables, and chi-squared tests or Fisher's exact tests was used for categorical variables. Generalized linear regression was used to assess group differences in cognitive function and MRI variables. To screen for correlates of cognitive function, univariate regression analyses were performed between cognition and clinical characteristics and MRI parameters. These analyses were corrected for age, education level, and depression. Multivariate linear regression analyses were performed to examine associations between MRI parameters and cognitive performance, controlling for age, education level, and depression. MRI variables that were significantly different between CP patients and those with CI or those that correlated with cognition in univariate regression analyses were entered into the multivariate regression analysis. Final multivariate model was determined using backward selection method with elimination criterion of p-value > 0.05. Bonferroni corrections were applied where applicable and p-values < 0.05 were considered statistically significant. SAS (ver. 9.3; SAS Institute, Cary, NC, USA) was used for all analyses.