Study 1.

Ten grapheme-color synesthetes and twenty-five controls participated in this study. For each subject, we acquired structural images on a Bruker 3T Medspec S300 whole body scanner, equipped with a one channel emission/reception head coil, using a T1-weighted 3D MP-RAGE image consisting in 176 sagittal partitions in two segments with an image matrix of 256x112 (read x phase). Further imaging sequence parameters were: TR/TE/TI: 16/4.96/903 ms, excitation pulse angle: 8°, acquisition matrix: 176x224x256 (x,y,z), fast phase encoding in anterio-posterior direction (112 steps per RAGE train, 2 segments), slow phase encoding in left-right direction, isotropic nominal resolution: 1 mm, BW = 130 Hz/Px, readout in caudo-cranial direction, number of averages: 1 and total measurement time: 14 min 40 s.

Study 2.

Twenty-two grapheme-color synesthetes and twenty-five controls participated in this study (a different volunteers’ population from Study 1). For each subject, we acquired a high-resolution structural MP-RAGE image on a 3T Philips Intera Achieva, using a 32 channels coil, 180 sagittal slices of 256x240 (read x phase). Further imaging sequence parameters were: TR/TE/TI: 25/3.7/800 ms, excitation pulse angle: 15°, isotropic nominal resolution: 1 mm, BW = 191 Hz/Px, readout in anterio-posterior direction, number of averages: 1, sense factor anterio-posterior: 2.2, right-left: 2 and total measurement time: 9 min 41 s. We also acquired diffusion-weighted images (DTI) using a DWI sequence (120x120 matrix size with 70 contiguous transverse slices, FOV = 240 mm, 2x2x1.75 mm³ spatial resolution, TR = 6845 ms, TE = 667 ms, FA = 90°, Sense Factor = 2 to improve the signal to noise ratio, total acquisition time: 18 min). The encoding protocol included 60 different non-collinear directions (gradient factor b = 1000 s/mm²), and one image without diffusion weighting used as reference volume.

Useful characteristics about the synesthetes and controls populations for the two studies are provided in S1–S3 Tables.

Voxel Based Morphometry (VBM).

We analyzed the structural images using the VBM approach [5] where region-wise volumetric comparison among groups of subjects is performed. It requires all individual images to be registered in the same space and segmented in different tissue classes. Then, registration parameters and a mixture of Gaussian distributions for brain tissue modeling have to be estimated. A generative approach has been proposed for such a model parameter estimation that alternates among classification, bias correction and registration [21]. We used the data processing pipeline VBM8 (http://www.neuro.uni-jena.de/vbm/) that implements this approach as a toolbox extension of SPM8 (http://www.fil.ion.ucl.ac.uk/spm/software/spm8/) running with Matlab language. Compared to the previous one, this new version mainly improves the segmentation step in removing noise [22], estimating partial volume effects [23], taking into account local intensity variations and ensuring local coherency of tissue labels [24]. A joint segmentation-registration approach is used to segment and warp the individual tissue probability maps into a common study-specific reference space. Then an affine registration may be applied for transformation into the Montreal Neurological Institute (MNI) referential space. Each structural image is segmented by attributing to each voxel a probability of being in white matter (WM), grey matter (GM), and cerebrospinal fluid (CSF). This procedure uses a maximum a posteriori estimation to take into account local variations of intensity and estimates a mixture model composed of several Gaussian distributions notably for pure tissue (3 gaussians) and partial volume effects (2 gaussians). To impose local coherence, a Markovian approach introduces spatial prior into the model estimation. The high-dimensional DARTEL [21] registration algorithm iteratively computes deformation fields for warping each individual image to the common space. To counterbalance local deformations, expansion, or contraction, induced by highly non-linear registration and affine transformation, the tissues’ probability values computed were scaled by the Jacobian determinants of the non-linear deformations (‘‘modulation step” [25]). Finally, following the recommendations of [26] for decreasing the false positive rate, we smoothed these ‘‘modulated” tissues probability maps using a 12-mm full-width at half-maximum Gaussian kernel (same pattern of results with an 8-mm kernel). A visual control of the sample homogeneity as implemented in VBM8 was realized based on the covariance between the images. No outlier was detected.

Diffusion Tensor Imaging.

DTI images were first denoised [27] and preprocessed using FSL software (http://www.fmrib.ox.ac.uk/fsl/). The images were corrected for geometric distortions caused by Eddy currents and head movements (FSL “eddy” tool). The diffusion tensor was estimated, and the local diffusion FA and mean diffusivity (MD) parameters were calculated for the entire brain in each participant. These parameters were computed from the three estimated eigenvalues describing the water diffusion in three orthogonal directions. FA represents the coefficient of variation of these eigenvalues interpreted as the directionality of water diffusivity into fibers (coherence). MD is the mean of the eigenvalues and blurs the directionality. In general FA is noisier than MD. FA was considered for WM tracks only, MD for the entire brain. These parameters are myelination markers, serving as measures of tissue density for MD and fiber organization for FA.

For each participant, the non-diffusion weighted image (T2-weighted) was realigned to the corresponding structural image. The computed realignment parameters were applied to the corresponding MD volumes to be aligned to the structural image. Then they were warped using the deformation field previously computed for this image for warping to the common space, scaled by the Jacobian determinants of the deformations and smoothed (12 mm FWHM). Subsequently, MD volumes were analyzed on a voxel-by-voxel basis similarly to VBM-GM/WM volumes analysis [28]. Because such a voxel-based analysis does not ensure a proper alignment of individual fiber tracts, a Track-Based Spatial Statistics (TBSS) was also processed [29]. Individual FA data were nonlinearly realigned and a mean FA image was computed and used to define a mean FA skeleton on which all individuals FA data were projected. No spatial smoothing was applied to FA maps.

Sulci extraction and morphometry.

“Sulcus-based morphometry” provides measures of the cortical fissures of the brain, which have been found to be associated with brain maturation [30, 31], brain alteration with age or pathology [32, 33], cognitive abilities [34–36] or correlated in a population of twins [37]. Because grapheme-color synesthesia seems to take place during the learning of reading [38] it was of interest to search whether some sulcal anatomical differences could be the cortical support of synesthesia experiences. We investigated sulci morphometry in controls and synesthetes based on shape descriptors (width, length, mean depth, and total surface area) analyzed for each sulcus. We used Freesurfer (https://surfer.nmr.mgh.harvard.edu) to classify grey and white matter tissues and Morphologist 2013, an image-processing pipeline included in BrainVISA (http://brainvisa.info/web/index.html), to quantify the sulcal descriptors. Briefly, the Morphologist 2013 segmentation pipeline computes a brain mask, imports brain tissues (gray, white matter and CSF) classified by Freesurfer, performs gray/white surface identification and spherical triangulation of the external cortical surface of both hemispheres. Sulci were then automatically labeled according to a predefined anatomical nomenclature [39]. The number of voxels on the junction between the sulcal mesh and the brain hull gave a measure of sulcal length. The mean depth was defined as the mean of geodesic distances computed for all voxel belonging to the sulcal fundus, from the bottom of the sulcus to the brain hull, along the cortical mesh. The surface area was the total area of the sulcal mesh. The sulcal width was obtained by dividing the enclosed CSF volume by the sulcal surface area; see [40] for details.

The pipeline failed for two control subjects, leaving 32 synesthetes and 48/50 controls. In addition, each sulcus could not be measured in every subject. We considered that a sulcus could not be measured for a given subject if the four measures (length, width, surface and depth) were equal to zero (that was the case for most zero values). Those cases appeared when individual variability was too high compared to the reference population. The pattern recognition algorithm then failed to sulcus identification, meaning that the corresponding sulcus was absent for this subject or not measurable. When a sulcus was not measured for more than 11% of subjects either in the control group or in the synesthete group (that was, in either 5/48 subjects in the control group or 3/32 subjects in the synesthete group), it was removed from the analysis for the two hemispheres (it does not make sense to compare different sulci if tested on too many different subjects). When a sulcus was missing for a subject, we attributed the median value of all other subjects (controls and synesthetes, left and right sides, independently) for that sulcus. The interpolated value was used only for the sum and the subsequent normalization. We normalized each measure by dividing the within subject sum of the sulci independently for length, depth, surface and width. We computed Pearson’s correlation between left and right sulci values to consider pooling the right and left values or not.

Structural images.

We compared the regional tissue probability maps (modulated and smoothed as described above) of controls and synesthetes by performing a voxel-wise univariate analysis using the general linear model (GLM) as implemented in SPM8. Because the global brain size can vary across subjects, we included brain volume as a factor of noninterest in our statistical tests. In order to calculate the global brain volume, we used the modulated images by summing together the GM and WM probabilities of all voxels. To avoid possible edge effects between different tissue types, we applied an absolute intensity threshold mask of 0.1 on each tissue probability.

For Study 1, average age slightly differed between the 2 groups (29.8 vs. 36.4 years, p = 0.13), and our synesthete group had more women (7/10 vs. 10/25 in our control group). Both factors may generate local differences not related to synesthesia, so we also included sex and age as factors of noninterest. In Study 2, the two groups were not different in term of age (28.2 vs. 27.7 years) and sex (18/22 vs. 18/25 women). However, in order to compare the two studies, we adopted the same model including sex and age as cofactors at the expense of the number of degrees of freedom.

Diffusion Tensor Imaging.

For MD data, we performed (Study 2) a voxel-wise univariate analysis using a GLM model similar to that used for tissue analysis. We considered MD, by definition an average value in each voxel, as a MR biomarker in the same way as GM or WM measures [41, 42].

There are several pitfalls using VBM to analyze DTI data mainly due to alignment issues. The use of the ‘FA skeleton’ and the projection of FA values onto it is an interesting approach that improves individual images alignment; moreover, it reduces the need for smoothing and the number of voxels to be tested; so, we used this TBSS approach as implemented in FSL to compare FA projections onto the mean population skeleton between synesthetes and controls. Note that in this case the analysis is reduced to white matter only.

We used two different statistical approaches, the Random field Theory (RFT) for MD and voxel-wise permutation tests [43] for FA, in order to have the best method adapted to the measures we have. Moreover, using the statistical tools commonly used for FA and MD allows the comparison of our results with the literature. The RFT cannot be applied to the tracks of the skeleton (a track is definitely not a random field). Permutation tests are valid because they deal only with group exchangeability; unfortunately, permutation tests do not allow any meaningful inference since they do not allow building up confidence intervals of effect sizes, because they cannot test any other hypothesis than the null hypothesis [44]. These tests allow only the computation of p-values. The only conclusion based on a permutation test when getting a small p-value is that the observed sample is surprising if we suppose exchangeability (so the two groups may not be exchangeable). This does not provide much information but may yet constitute a starting point: finding locations with small p-values indicates regions where it is worth exploring the data in more details. In order to obtain at least p-values as accurate as possible for the FA measure, we performed cluster-based statistics with the "threshold-free cluster enhancement" (TFCE) approach as implemented in FSL “randomize” function [45].

Sulci.

We compared the distribution of sulcal length, mean depth, surface and width between the two populations across both studies for right and left hemispheres. We computed 99.9% CI for the group difference using a linear model, with age and sex as covariates. We chose quite arbitrarily the 99.9% value to partially account for multiple comparisons (4 measures in 61 sulci in each hemisphere) while avoiding being too conservative, because of possible correlations between measures (a 99.9% CI corresponds to a 95% family-wise CI for 50 independent comparisons). We could not perform multivariate analyses because of the large number of missing values (not all sulci could be identified in every subject).

Data exploration.

In order to compare the results of Study 2 with our results published in 2012, we considered as a starting point the family wise error (FWE correction for multiple comparison) measured at the cluster level. We defined two regions of interest (spheres, radius 6 mm) located at the coordinates of the clusters found in [8] that survived a strict FWE correction, in the right retrosplenial cortex (RSC) and the left superior temporal sulcus (STS). Two additional regions in the fusiform gyrus and in the parietal cortex were considered. For each sphere, we computed in both studies WM estimated using our general linear model for the two groups. Recently Eklund et al. [17] showed that such a multiple comparison correction procedure did not guarantee at all the control of false positives at a cluster level. Besides standard p-values for comparison with published studies, we reported CIs of effect sizes to facilitate appropriate data interpretation in context. Moreover, our statistical maps are displayed without introducing the classical p-value level threshold. We adopt the dual-coding approach proposed in [19] where differences in effect size (beta estimates) are color-coded and associated t-statistics mapped to color transparency.

Study 1.

We compared the local distributions of WM and GM in the brains of 10 synesthetes with the brains of 25 controls. In order to replicate our previous results [8] we first used the same, p-value based, statistical analysis conditions (cluster forming threshold at p<0.0001, cluster extent k>70) but we did not obtain any cluster for WM or GM. At a more lenient threshold (p<0.001, the default cluster-defining threshold in SPM) we found local increases of WM in synesthetes compared to controls in the right retrosplenial cortex (Fig 1A) and the left anterior middle temporal gyrus (Fig 1B), similarly to [8] (see their Fig 6). However, these clusters were found at a statistical threshold non-corrected for multiple comparisons (see Table 1). At this threshold new clusters appeared (see Table 1) in the right inferior parietal lobe (Fig 1C) and in the vicinity of color-sensitive regions of the fusiform gyrus [8, 46–48] (posterior cluster in Fig 1B). These two regions were close to some reported results [9, 10, 12].

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Local increases of WM in synesthetes compared with controls.

Detected changes are projected onto the study-specific structural image transformed in the MNI space. (A) Increase in the right RSC (14, -57, 6). (B) Increase in the left anterior middle temporal gyrus (Blue cross: -56–15–14). Note the second cluster in the posterior part located in the fusiform gyrus (-26–73–12). (C) Increase in the right inferior parietal lobe (58–31 28). All coordinates are in MNI space expressed in mm. Cluster-forming threshold p<0.001 uncorrected, t>3.4, k>70 voxels i.e 70 mm³. T scale is between 3 and 5.5. Neurological convention (right hemisphere displayed on the right side of the figure).

https://doi.org/10.1371/journal.pone.0194422.g001

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Local increase of WM in synesthetes (n = 10) compared with controls (n = 25).

https://doi.org/10.1371/journal.pone.0194422.t001

Fig 2 displays structural differences without the introduction of a threshold.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. The structural differences between the two groups are projected without the introduction of a threshold onto the study-specific structural image transformed in the MNI space.

Same slices, convention and cursor positions as in Fig 1. Beta difference scale is between -0.15 and 0.15 and mapped to color hue. t-statistic magnitude (| t |) is mapped to color transparency.

https://doi.org/10.1371/journal.pone.0194422.g002

Fig 3 shows the individual data (denoted by circles for this first study) for the four regions of interest identified in Fig 2, using 6-mm spheres centered on the peak coordinates of Table 1, as well as the 99.99% CI (corresponding to the arbitrary voxel-wise threshold of our statistical map used to identify clusters). By construction, the extent of each CI of the group difference was expected to be above zero, since each value is the average of voxels with p-values < 0.001 for the group comparison. Note however that for small clusters (like in the left STS) this was not necessarily the case since the 6-mm sphere may extend beyond the cluster of small p-value voxels.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. WM in synesthete and control groups, and 99.99% CI for the group difference, for four regions of interest centered in the right RSC, left STS, the right inferior parietal lobe and the left fusiform gyrus (coordinates in Table 1), for the subjects of Study 1, Study 2, and all subjects together.

The WM tissue probability at each voxel was estimated by the linear model that included total brain volume, age and sex as covariates.

https://doi.org/10.1371/journal.pone.0194422.g003

Study 2.

We compared the local distributions of WM and GM in the brains of 22 synesthetes with the brains of 25 controls. Fig 4 shows the structural differences of WM without introducing any arbitrary threshold, in the same slices as for Study 1. Fig 4 (crosses) indicates that structural differences are unlikely (or too small to be relevant and detectable) in the regions of interest defined in the first study (see also the CI of group differences for Study 2 in Fig 3). Using conventional 5% family-wise error risk, we did not find any ground (increase or decrease) for rejecting the Null hypothesis of no difference between the two groups. We obtained the same result for GM analysis. Following the protocol proposed by Kurth et al [49] we searched to assess whether the voxel-wise gray matter asymmetry in the synesthete group was significantly different from the voxel-wise gray matter asymmetry in the control group. No difference was found using a standard p<0.001 for individual voxels.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Study 2.

Structural differences between the two groups are projected without the introduction of a threshold onto the study-specific structural image transformed in the MNI space. Same slices, convention and cursor position as in Fig 1.

https://doi.org/10.1371/journal.pone.0194422.g004

Pooling data from Study 1 and Study 2.

Here we considered pooling data from Study 1 and Study 2 (synesthetes = 32, controls = 50). Because data were acquired on two different scanners, 3T Bruker for Study 1 and 3T Philips for Study 2, we were particularly meticulous with the quality check after the preprocessing including realignment and segmentation steps. We used the module available in VBM8 for checking homogeneity of the segmented volume GM and WM separately using covariance. We added in our general linear model a fourth covariable of non-interest corresponding to the two different conditions of data acquisition. Fig 5 shows all the WM differences without introducing any arbitrary threshold. Fig 3 shows the CIs for the regions considered in Figs 2 and 4. Two clusters emerged at our threshold, none of them allowing us to reject the null hypothesis at the 0.05 FWE level (see Table 2).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. All the structural differences between the two groups are projected without the introduction of a threshold onto the study-specific structural image transformed in the MNI space.

Same slices, convention and cursor position as in Fig 1.

https://doi.org/10.1371/journal.pone.0194422.g005

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Local increase of WM in synesthetes (n = 32) compared with nonsynesthetes (n = 50).

https://doi.org/10.1371/journal.pone.0194422.t002

In order to provide an idea of the power of these “Null” results, one may consider the upper (for increases) and lower (for decreases) limits of the CIs. For the increases in synesthetes relative to control reported in Fig 3, the largest differences of WM probability compatible with our measures in the 4 regions of interest were about 0.05 (based on a 99.99% CI). In the voxels with the maximal statistical differences reported in Fig 5 (max t-value = 4.39 and min t-value = -4.90) the upper and lower limits of the WM probability differences were respectively 0.0555 and -0.0525.

For 29/32 synesthetes, we had measured the strength of associations (“photism strength”), estimated during psychophysics Stroop-like tasks [50] (see S1 Table). We inserted this value as a covariate in our design matrix, in addition to brain volume, sex, age and scanner, and searched for structural data correlations using the VBM approach. The reasoning was that if effects, even weak, were found in the data, correlation with photism strength would have indicated a direct relationship with the synesthetic experience. Correlations were between -3x10^-4 to +3x10^-4, with no brain region exhibiting any reliable correlation (see S1 Fig). Results were similar for GM data.