Study data set
The brain imaging genetics (BIG) study was initiated in 2007 and comprises healthy volunteer subjects, including many university students, who participate in diverse imaging studies at the Donders Center for Cognitive Neuroimaging (DCCN), Nijmegen, The Netherlands (Franke et al. 2010). At the time of this study, the BIG subject-pool consisted of 2709 healthy adult volunteers (1435 females) who had undergone anatomical (T1-weighted) MRI scans, usually as part of their involvement in diverse small-scale studies at the DCCN, and who had given their consent to participate in BIG.
Handedness of the participants was assessed by an item in their enrolment form. This consisted of subjects selecting an answer from the two options “left-handed/right-handed” (in Dutch). Only those subjects who clearly indicated one or the other state were included in our analysis. This resulted in a sample of 2307 right-handed subjects and 119 left-handed subjects, with a mean age of 25.70 years and a standard deviation of 10.56 years. Note that the BIG study was not recruited to specifically study handedness, and therefore, only a simple binary measure was available. Nonetheless, simple self-assessments show close agreement with dichotomous scoring of handedness as derived from multi-item inventories (see “Discussion”). The proportion of left-handers was lower than in the general population; this was due to left-handedness being used as an exclusion criterion for some of the imaging studies that were pooled into the overall BIG dataset. Nonetheless, handedness was not associated with any particular acquisition protocol in the overall dataset (see below).
A subset of 381 subjects (345 right and 8 left-handed) had undergone a brain MRI scan twice, with at least 1-day separation between scans. The median period between scans was 184 days with a range of 1–2650 days. At the time of the first scan, the median age of this group was 22 years. Twice-scanning of these subjects allowed us to perform scan–rescan correlation analysis to assess the stability of individual differences in the brain anatomy measures described below. In principle, if the first and second scans for given individuals had tended to be performed with the same acquisition protocol (see below), there was potential for scan–rescan correlations to be inflated: however, there were no systematic relations of scans for twice-scanned subjects with respect to heterogeneity of image acquisition.
MRI data were acquired with either a 1.5-Tesla Siemens Sonata or Avanto scanner or a 3 Tesla Siemens Trio, TimTrio or Skyra scanner (Siemens Medical Systems, Erlangen, Germany). Given that images were acquired during several smaller scale studies, the parameters used were slight variations of a standard T1-weighted three-dimensional magnetization prepared rapid gradient echo sequence (MPRAGE; 1.0 × 1.0 × 1.0 mm voxel size). The most common variations in the TR/TI/TE/sagittal-slices parameters were the following: 2300/1100/3.03/192, 2730/1000/2.95/176, 2250/850/2.95/176, 2250/850/3.93/176, 2250/850/3.68/176, 2300/1100/3.03/192, 2300/1100/2.92/192, 2300/1100/2.96/192, 2300/1100/2.99/192, 1940/1100/3.93/176 and 1960/1100/4.58/176. To account for magnetic field strength effects, an inhomogeneity correction was applied. There was also variation in the head coils used. The following arrays were employed (with their frequencies) in the right-handed participants: 32-channel (24 %), 12-channel (4 %), 8-channel (38 %) arrays, and single head coil (33 %). In the left-handed participants, this distribution was 32-channel (27 %), 12-channel (0 %), 8-channel (33 %) arrays, and single head coil (40 %).
T1 images were processed using the VBM8 tool and its default settings (http://www.neuro.uni-jena.de/vbm/), implemented in SPM8 (Wellcome Department of Imaging Neuroscience Group, London, UK; http://www.fil.ion.ucl.ac.uk/spm). This procedure segments T1 images into gray matter (GM), white matter (WM), and cerebrospinal fluid (CSF). It then generates the corresponding tissue maps spatially normalized to MNI space (Ashburner 2007) and modulated by the non-linear component of their spatial transformation. The resulting GM images contained information on local volume differences, independent of overall differences in brain size (http://www.neuro.uni-jena.de/vbm/segmentation/modulation/).
In addition, T1 images were independently processed using FreeSurfer’s (v5.3) default “recon-all” pipeline, which performs automated segmentation of non-cortical tissues, as well as automated parcellation of the cerebral cortex (Fischl et al. 2002, 2004).
Measurement of regional volumes
Our analyses focused on the cerebellum, and cortical areas corresponding with the classically defined perisylvian language network, i.e., regions of the inferior frontal gyrus and superior temporal gyrus, as well as the post- and precentral gyri due to their involvement in motor cognition and handedness (see “Introduction”). Volumetric estimates of these regions of interest were derived from the processed T1 images in two ways.
First, regional volumes were extracted from the spatially normalized GM images according to probabilistic atlas definitions. In other words, for a given probabilistic region of interest, we performed a voxel-wise sum of gray matter volumes, weighted by the probability of each voxel belonging to that specific region. Cerebellar estimates were based on the Diedrichsen atlas (Diedrichsen et al. 2009), which contains probabilistic definitions for 28 cerebellar regions in standard space (Fig. 1), 10 of which have left–right counterparts. Only those voxels were included for which the probability weight of belonging to the cerebellum was at least 50 %, to prevent the unintended inclusion of cerebral cortical GM voxels into cerebellar regions. This threshold also meant that cerebellar regions did not generally overlap with each other (see Fig. 1). Cerebral cortical volumes were estimated by the probabilistic Harvard–Oxford (HO) cortical structural atlas that defines 48 bilateral cortical regions in standard space (Goldstein et al. 1999, 2007). Of the 48 bilateral regions, the following was selected and splits at the center of the left–right axis: pars opercularis, pars triangularis, superior temporal gyrus (anterior), superior temporal gyrus (posterior), planum temporale, Heschl’s gyrus, postcentral gyrus, and precentral gyrus (see Fig. 2). Given that there was no overlap between these cortical regions of interest and GM cerebellar voxels, no further manipulation of the HO atlas or of its probabilistic regions was applied. The Diedrichsen and HO atlases were distributed with the FSL software package (http://www.cma.mgh.harvard.edu/fsl_atlas.html).
Second, regional cortical volumes were derived from FreeSurfer’s cortical anatomical parcellations, according to the Desikan atlas (Desikan et al. 2006). The selected regions were the pars opercularis, pars triangularis, superior temporal, transverse temporal, precentral, and postcentral cortex (See Fig. 3). FreeSurfer estimates of cerebellar volumes were also derived from its segmentation of the cerebellum into gray and white matters, and further into the left and right structures, but these data were not used further after visual quality control (see below).
We visually inspected the spatially normalized GM maps of all study participants, with respect to two main features: the overall quality of the normalized image, and the correct application of the cerebellar probabilistic atlas with regard to non-cerebellar tissue. The spatially normalized GM images were visualized alone and also overlaid with the cerebellar probabilistic atlas, from 35 internal slices of coronal and sagittal views per participant. Images that had not normalized correctly to the standard brain appeared as distorted or incomplete, and were excluded from further analysis. Detailed inspection showed that these problems resulted from overall low image quality, head-motion artifacts, or unusual anatomy. In addition, images were excluded when we detected an overlap between probabilistic cerebellar definitions and wrongfully segmented dura or sinuses. After applying all of these exclusion criteria, the remaining sample size was 2226 (103 left-handers).
Inspection of FreeSurfer’s cortical parcellations was performed independently of the above, again for the entire data set, and followed the protocol developed by the ENIGMA consortium (Thompson et al. 2014) (http://enigma.ini.usc.edu/protocols/imaging-protocols/). Specifically, it consisted of visually checking individual parcellations, plotted from both internal (axial and coronal) as well as external (lateral and medial) views. Individual measurements derived from erroneous parcellations, and in some cases, whole images were excluded from analysis. Erroneous parcellations were identified from internal views when cortical regions were missing, left–right homologous labels were not grossly comparable in position, or cerebral cortical labels had been mapped to non-cortical tissue (e.g., the cerebellum or dura mater). From external views, global errors could be visualized as a rough/spiky brain surface or highly fragmented and interspersed cortical labels. External views also revealed poor anatomical labeling, specifically when the ‘banks of the superior temporal sulcus’ label mapped extensively onto the externally visible brain surface and affected surrounding regions, and when the ‘supramarginal gyrus’ label extended into the superior temporal gyrus. After excluding the data that did not pass these quality filters, all regional measures except for the superior temporal gyrus had a sample size of 2003 (97 left-handers), while for the superior temporal gyrus, the sample size was 1676 (87 left-handers). The overlap of this sample with the quality checked, spatially normalized GM data, was 1875 participants (88 left-handers), for all regions apart from the superior temporal gyrus. For the superior temporal gyrus, the overlap was 1572 participants (79 left-handers).
After the visual quality control, the number of twice-scanned participants with data available for scan–rescan correlation analysis was 329 for the cerebellum and HO cortical data, 277 with Freesurfer data for all cortical regions apart from the superior temporal gyrus, and 226 with Freesurfer data for the superior temporal gyrus.
Freesurfer’s cerebellar segmentations were also visually inspected by plotting them against participants’ scans in a set of axial and coronal views. Focus was placed on detecting segmentation errors with its surrounding dura mater or dural sinuses, as these are complex structures whose intensities on T1 images are particularly similar to those of cerebellar gray matter (Hwang et al. 2011). An initial inspection of 50 random subjects revealed that these problems, although subtle, occurred frequently (>30 % of the visualized subjects). Freesurfer cerebellar measures were subsequently excluded from our analyses.
In addition, each cerebellar and cerebral cortical measure was approximately normally distributed (not shown), and we excluded outlier values beyond plus or minus 3.5 standard deviations (SD) from the mean. Stability of individual difference measurement was assessed by correlating the values for the twice-scanned subjects from the first scan to the second scan, by Pearson’s correlation.
For each structure and participant, asymmetry was measured by an asymmetry index (AI) using the formula (L − R)/(L + R) where L stands for left-side volume and R stands for right-side volume. Outlier removal and scan–rescan correlations for AIs were performed as described above (“Quality checks”). Whether the mean AIs differed significantly from zero was tested by t tests. All AIs were then adjusted by linear regression (iteratively reweighted least squares) for the potential covariate effects of age, estimated intracranial volume (ICV), sex, field strength, scanner type, and their two-way interactions (with the exception of field strength*scanner type). In addition, we included quadratic terms for age and ICV. All further analyses were conducted using the residuals from these regressions. Not all terms were significant for all AIs, but the inclusion of non-significant terms had negligible effects on the residuals. This uniform approach had the advantage that results could be compared across structures, rather than making them contingent on individual models for each cerebellar lobular AI and cerebral cortical AI.
Associations with handedness and cerebral cortical asymmetries
Welch’s two sample t tests were conducted to assess potential associations between cerebellar AIs and handedness (Welch 1947). This test avoids assumptions of balanced group sizes and equal variances. Pearson’s correlation coefficients were used for assessing the correlations between cerebellar AIs and the AIs of the cerebral cortical regions. Bonferroni correction was applied separately for the correlation analyses of cerebellar AIs with HO-derived cerebral cortical AIs (80 tests) and Freesurfer-derived cerebral cortical AIs (60 tests).