Keywords

1 Introduction

Multiple Sclerosis (MS) is a chronic, autoimmune, progressive, inflammatory neurological disease which disturbs the central nervous system (CNS), mainly in young adult, affecting 2.3 million people in the world [1, 2]. MS is characterized by myelin degradation of axons’ sheath. Demyelinated regions, known as white matter (WM) lesions, cause motor, sensory and cognitive degradation in as many as 65% of patients [1, 2].

The detection of WM lesions based on magnetic resonance imaging (MRI) is a key feature to proper MS diagnosis according to McDonald’s criteria and the MAGNIMS consensus guidelines [2, 3]. However, MS misdiagnosis is still a concern [4]. There are other structural changes such as regional WM and grey matter (GM) atrophy in the brain, associated with cognitive deficit and disease severity. Moreover, its cause, how early it begins, and which regions are primarily affected remains unclear [5, 6]. Thus, it is important to identify biomarkers to provide more accurate diagnostics and cues about the pathophysiology of MS disease.

Structure of the brain can be accessed in a non-invasive way through MRI, which is a good starting point to carry out methodical analysis to identify structural biomarkers [7, 8]. Univariate analysis methods such as Voxel-Based-Morphometry (VBM) [9] and neuroimaging pattern recognition using multivariate pattern analysis (MVPA) techniques [8] such as automatic classification have been used to identify morphological alterations in MS [10,11,12]. VBM is an automated statistical test performed in each voxel individually, which evaluates whether there is a difference between groups of participants in the tissue volume. This type of analysis is usually applied based on GM distribution, but it can be also applied to WM (lower sensitivity). Since VBM is an hypothesis-driven univariate test, it will only detect a predicted effect (typically, a mean difference between two groups) in individual voxels (massive multiple comparisons). In contrast, multivariate methods are able to identify unpredicted data patterns reflecting group (many times individually subtle) differences, in the absence of a priori hypotheses, spread across different voxels [8, 13].

Additionally, Surface-Based-Morphometry (SBM) [14] can be an useful and innovative approach to identify new MS structural biomarkers, based in surface variations. Like VBM, SBM is a voxel-wise univariate method to test the hypothesis of mean differences between groups in a particular surface measure, such as the cortical thickness and the gyrification index [15]. It is known that there is an increase in gyrification during early stages of brain development; after childhood, there is stabilization and with aging, there is a decrease of cortical thickness and gyrification [15, 16]. Investigation of how these measures behave age-independently, in the context of MS, can provide valuable information for diagnosis improvement and comprehension of disease.

In order to better understand the brain structural patterns involved in MS, we applied VBM in GM and WM volumes, SBM in cortical thickness and gyrification index, and multivariate classification of GM and WM volumes with a Support Vector Machine (SVM). Despite univariate analysis and MVPA methods cannot be directly comparable [13], as classification is performed using volumetric measures, it allows us to compare the regions highlighted by the weight maps of SVM classification with regions identified by VBM analysis.

2 Methods

Participants.

The study included 59 MS patients (39F; mean age: 37,00 ± 7,24 years) and 64 controls (42F; mean age: 36,84 ± 9,79 years). Patients were recruited at the Coimbra Hospital and Universitary Centre (CHUC) and met the criteria for MS diagnosis according to McDonald Criteria [2]. All participants gave written informed consent. Local ethics committee approved the study.

MRI Data Acquisition and Processing.

Data were collected with a Siemens Magnetom TIM Trio 3 T scanner (Siemens, Munich, Germany) with a phased array 12-channel birdcage head coil. We acquired a 3D anatomical T1-weighted MRI pulse sequence (TR 2530 ms; TE 3.42 ms; TI = 1100 ms; flip angle 7°; 176 slices with no gap, with isotropic voxel size of 1 mm; FOV 256 mm) of all participants. Data were pre-processed using SPM12 software [17] and CAT12 [18] toolbox in MATLAB (version R2019a) [19]. Orientation of structural scans was manually corrected by aligning each image onto the axis of the anterior and posterior commissures. CAT12 was used to perform high-dimensional spatial normalization to a template stereotaxic space, segmentation of each scan in GM and WM tissue images, and modulation to compensate for the effect of spatial normalization. This last step involves scaling, by the amount of expansion or contraction, so that the total amount of GM/WM in the modulated images remains the same, as it would be in the original images [20].

Voxel Based Morphometry (VBM).

Modulated and normalized GM and WM outputs were smoothed using an 8 mm full-width-at-half-maximum (FWHM) isotropic Gaussian kernel, required to guarantee normality of the data for subsequent statistical analysis. The statistical model was built as an independent-samples t-test (two groups), with 3 covariates - gender, age and total intracranial volume (TIV) - to correct for different individual brain sizes. We applied a 0.1 absolute masking threshold to the GM/WM data and considered differences with p < 0.05 (FWE corrected) as significant. Clusters of contiguous voxels with significant differences were overlaid onto a template T1 brain image for visualization and the corresponding brain region was identified by their coordinates in standard space.

Surface Based Morphometry (SBM).

The image processing was performed as for VBM and allowed the extraction of additional surface parameters: cortical thickness and gyrification index. The extracted cortical thickness and gyrification maps were smoothed using 15 mm and 20 mm FWHM kernel, respectively. The statistical model for analysis was built similarly to VBM, with 2 groups and 2 covariates: gender and age (surface measures do not depend on volume measures so TIV is not needed as covariate in this model). Results of surface differences were overlaid as maps in a template reconstruction of the brain and a threshold of p < 0.05 (FWE corrected) was applied.

Support Vector Machine (SVM) Classification.

We used the Pattern Recognition for Neuroimaging Toolbox (PRONTO) [21] in MATLAB to build a classification model using a SVM: a supervised learning method that requires labelled training data to perform the classification of unseen test data [13]. In our study, the labels represent groups, MS patients and controls, whereas the features are the volume of GM or WM in each voxel. Each brain image is assigned to MS or control group and each voxel is labelled according with the image label. Leave-one-subject-per-group-out cross validation (LOO-CV) was implemented, leaving one image per group in each iteration out for testing. During the training phase, in order to learn the relationship between the features and the labels a mathematical decision function is determined, and a weight is assigned to each feature, representing its importance for discrimination. This decision function will be the criterion to classify the unseen images in the test set, and then obtain performance values (accuracy, sensitivity and specificity). To estimate the significance of the classification, we considered the null hypothesis that there is no information about the labels in the data, i.e., there is no difference throughout the voxels of all images, hence there is no discrimination between groups to be done. To calculate this probability, we applied a permutation test, by repeating 1000 times the same classification procedure with randomized labels each time. After this procedure, a level of statistical significance was obtained and expressed as the p-value. The smaller the p-value, the more significant is the performance of our classifier [13, 22, 23]. This classifier was applied in GM and WM volume data to assess whether MS patients can be accurately differentiated from healthy controls. Model performance was evaluated, and we also extracted weigh maps to explore which brain regions contributed more to the discrimination.

3 Results

3.1 VBM

In both VBM analyses with GM and WM we tested the contrasts controls > MS, which yields several significant clusters (p < 0.05, FWE corrected), showing GM and WM atrophy in MS patients (these are illustrated in Figs. 1 and 3, respectively). The inverse contrast, MS > controls, did not reveal any region where patients have significantly higher tissue volume than controls (data not shown).

Fig. 1.
figure 1

Results of VBM analysis of GM (controls > MS). Results are presented at a voxel-level p-value < 0.05, FWE corrected. Color bar scale represents t-values.

Full size image

Grey Matter VBM Results.

We found a decrease of GM volume in MS patients relatively to controls in clusters with standard space coordinates corresponding to regions in right thalamus and lentiform nucleus, right cingulate gyrus, left medial frontal gyrus, middle frontal gyrus, right rectal gyrus, left lingual gyrus and in right fusiform gyrus (see Fig. 1).

White Matter VBM Results.

We found a decrease of WM volume in MS patients in clusters that include regions in both right and left parahippocampal gyrus and regions in frontal and temporal lobe (see Fig. 3).

3.2 SVM Classification

SVM classification applied to GM volume achieved an overall accuracy of 86.51% (sensitivity 74.58%; specificity 98.44%) and a p-value of 0.001. In what concerns WM, SVM correctly classified 82.34% of cases (sensitivity 67.80%; specificity 96.88%), with a p-value of 0.001. The GM and WM weight maps show voxels with higher positive weights in regions similar to those identified with VBM analysis (see Figs. 2 and 3).

Fig. 2.
figure 2

Whole-brain representation of the discriminative map from SVM classification of GM images. In color bar scale, positively weighted voxels are displayed in orange/red and negatively weighted voxels are displayed in blue/green.

Full size image
Fig. 3.
figure 3

Results of VBM analysis of WM (controls > MS), on the left panel. Results are presented at a voxel-level p-value < 0.05, FWE corrected. Color bar scale represents t-values. The right panel represents the whole-brain discriminative map from SVM classification of WM images. In color bar scale, positively weighted voxels are displayed in orange/red and while negatively weighted voxels are displayed in blue/green.

Full size image

3.3 SBM

The same contrasts, as defined in VBM, were tested to study differences of cortical thickness and gyrification index between groups with on SBM analysis.

Cortical Thickness.

SBM analysis of cortical thickness yielded several significant clusters (p < 0.05, FWE corrected), shown in Fig. 4. We found lower values of cortical thickness in MS patients, as compared to controls, in the following regions: left postcentral gyrus; right and left precentral gyrus; left paracentral lobule; right and left middle frontal gyrus; right superior frontal gyrus; right and left inferior parietal lobule; right precuneus; right and left superior temporal gyrus; right and left middle temporal gyrus; right cuneus; right middle occipital gyrus; left parahippocampal gyrus; right posterior cingulate; insular lobe; cerebellum. The symmetric contrast (MS > controls) did not reveal regions with significant difference (data not shown).

Fig. 4.
figure 4

Results of SBM analysis of cortical thickness (controls > MS). Results are presented at a voxel-level p-value < 0.05, FWE corrected. Color bar scale represents log p-values.

Full size image

Gyrification Index.

Since this is an exploratory study, we show gyrification maps with thresholds at p < 0.001 (uncorrected), as with p < 0.05 (FWE corrected) there were no significant differences between groups. We found decreased gyrification in MS patients in: precentral gyrus; medial frontal gyrus; postcentral gyrus; superior temporal gyrus; cuneus; left superior occipital gyrus; right cingulate gyrus and in insular cortex (see Fig. 5). We found increased gyrification in MS patients in: inferior frontal gyrus; superior frontal gyrus; rectal gyrus; middle frontal gyrus; subcallosal gyrus; paracentral lobule; postcentral gyrus; precuneus; inferior temporal gyrus; superior temporal gyrus; lingual gyrus; cuneus; parahippocampal gyrus; cingulate gyrus and insular cortex (see Fig. 5).

Fig. 5.
figure 5

Results of SBM analysis of gyrification. Results are presented at a voxel-level p-value < 0.001 (uncorrected). Blue and red indicates decreased (controls > MS) and increased gyrification (MS > controls), respectively. Color bar scale represents p-values.

Full size image

4 Discussion and Conclusions

We performed both univariate and multivariate analyses of structural MRI scans of MS patients compared to control subjects. We aimed to identify different brain structural changes that occur in MS and to demonstrate the potential of these approaches to differentiate effects of this disease. Our VBM results revealed GM and WM atrophy. GM atrophy is in accordance with previous literature [11, 24]. On the other hand, WM atrophy is not commonly studied. As such, our results add to the understanding of the disease by demonstrating that WM is in fact a feature in MS. The more significantly affected region is thalamus, which is highly associated with cognitive deficit, a disease effect [25]. Since there is atrophy in MS patients, it is not expected any increase of volume relatively to controls. The few cases that we found are related to physiognomy differences in brain or hyperintensity of images.

SVM classification results demonstrated reliable detection of MS patients, which reveals this is indeed a useful predictive tool that could be used as biomarker with potential for support MS diagnosis. Weight maps from SVM classification highlighted similar regions to those found with VBM analysis, suggesting that differences in such regions are crucial for the discrimination between patients and control groups.

SBM analysis revealed a substantial decrease of cortical thickness in MS patients, which stresses the need of investigate its relationship with the disease effects. Regarding gyrification, both decreased and increased gyrification indexes were observed in different brain regions of MS patients. Nevertheless, only the increased gyrification cases survived to statistical correction for multiple comparisons. Studies addressing correlations between surface measures and aging have suggested that increased gyrification is linked to brain plasticity mechanisms to facilitate brain connectivity and functional development [16]. On the other hand, it is believed that MS patients have their brain connectivity altered [26]. As such, increased gyrification might be related with this MS effect, reflecting compensatory mechanisms.

In all analyses, clusters with between-group differences in GM/WM/cortical thickness/gyrification index between groups or voxel-wise classification weights are assigned a brain location, which is identified by their anatomical correspondence with coordinates in a standard space. These allow a qualitative description of disease discrimination based on macrostructural measures from MRI. Quantitative analysis of the statistical maps, e.g. number of voxels per cluster, peak localization, is required for further investigation of disease pathophysiology.

This study is, to the best of our knowledge, the first to apply VBM, SBM analysis of less common morphometric measures, and classification analysis of structural MRI scans in the same group of MS patients. Furthermore, these data-driven analyses indicated spatially distributed networks of brain regions with abnormal structure in individuals with MS providing important clues for the pathophysiology of this disorder.