High resolution data analysis strategies for mesoscale human functional MRI at 7 and 9.4 T

Abstract

The advent of ultra-high field functional magnetic resonance imaging (fMRI) has greatly facilitated submillimeter resolution acquisitions (voxel volume below (1 mm³)), allowing the investigation of cortical columns and cortical depth dependent (i.e. laminar) structures in the human brain. Advanced data analysis techniques are essential to exploit the information in high resolution functional measures. In this article, we use recent, exemplary 9.4 T human functional and anatomical data to review the advantages and disadvantages of (1) pooling high resolution data across regions of interest for cortical depth profile analysis, (2) pooling across cortical depths for mapping patches of cortex while discarding depth-dependent (i.e. columnar) effects, and (3) isotropic sampling without pooling to assess individual voxel’s responses. A set of cortical depth meshes may be a solution to sampling information tangentially while keeping correspondence across depths. For quantitative analysis of the spatial organization in fine-grained structures, a cortical grid approach is advantageous. We further extend this general framework by combining it with a previously introduced cortical layer volume-preserving (equi-volume) approach. This framework can readily accommodate the research questions which allow for spatial smoothing within or across layers. We demonstrate and discuss that equi-volume sampling yields a slight advantage over equidistant sampling given the current limitations of fMRI voxel size, participant motion, coregistration and segmentation. Our 9.4 T human anatomical and functional data indicate the advantage over lower fields including 7 T and demonstrate the practical applicability of T₂^* and T₂-weighted fMRI acquisitions.

Highlights

• •High resolution regular cortical grids are advantageous for local applications.
• •Equi-volume sampling is slightly advantageous over equidistant sampling in-vivo.
• •Isotropic submillimeter cortical sampling without spatial pooling requires high SNR.
• •9.4 T human T2 and T2* BOLD fMRI are practically feasible and provide high SNR.
• •9.4 T T2*-weighted 0.35 mm iso. res. anatomical images for laminar contrast in vivo.

Introduction

Functional magnetic resonance imaging (fMRI) has rapidly developed since its introduction and has become a key neuroscientific research tool. More and more detailed aspects of the functional architecture of the human brain are unraveled by means of steadily increased data quality (signal-to-noise ratio (SNR), spatial and temporal resolution). As the spatial resolution is pushed from tens of cubic millimeters to fractions of one cubic millimeter (submillimeter), more fine-grained structures such as cortical laminae and columns become accessible in-vivo in humans, closing the gap between invasive (animal) electrophysiological or optical research methods and human fMRI. While high and ultra-high field strengths have played an important role in making higher spatial resolutions accessible, advanced data analysis strategies paralleled this development to exploit the new physical possibilities.

Using exemplary data, we review in this article how ultra-high field fMRI allows investigating the mesoscopic functional organization of the cortex. In this context it is useful to differentiate applications by their purposes leading to different requirements for high-resolution measurements and analysis strategies. A major distinction can be made between applications that allow pooling of measurements along at least one cortical dimension from those applications that consider each voxel’s response individually. Submillimeter resolution fMRI has been, for example, used to separate (coarsely) cortical laminae reporting differential pooled functional (or anatomical) responses along larger stretches of cortex (Koopmans et al., 2010, Koopmans et al., 2011, Ress et al., 2007). Other studies have separated neighboring responses along the cortex while integrating signals across cortical depth, e.g. to map columnar feature representations with non-isotropic voxels (e.g. Cheng et al., 2001; Goodyear and Menon, 2001; Yacoub et al., 2008, 2007). In recent years, studies have started to separate functional responses along both cortical distance as well as cortical depth. These studies are particularly challenging, as they do not allow pooling along cortex either tangentially or radially and thus require high isotropic spatial resolution with sufficient contrast-to-noise ratio (CNR) at each submillimeter voxel. While challenging, this approach is suited for research questions that aim to map feature representations along the cortex while studying the consistency (“columnarity”) of these maps across cortical depth levels simultaneously. In order to properly relate functional responses across depth levels, correspondence between locations in different layers needs to be established. This requires appropriate cortex modeling and analysis strategies that not only separate layers but also link positions across depths. Initial results of this approach include axis-of-motion maps in area hMT or sound frequency maps in area A1 for different cortical layers (De Martino et al., 2015, Zimmermann et al., 2011).

In the sections below, we review the above-mentioned application scenarios and analysis strategies in more detail by the means of recent high resolution 9.4 T data, and illustrate the benefits of ultra-high field imaging (beyond 7 T). Furthermore, we describe in greater detail the regular grid depth sampling approach that allows us to accurately analyze, process, and visualize responses along layers and columns with high spatial resolution in restricted cortical regions. Besides varying cortical thickness and curvature supporting equidistant (i.e. cortical depth preserving) sampling, the analysis method now incorporates an equi-volume (i.e. cortical layer volume preserving, Waehnert et al., 2014) layer separation model.

Standard cortical meshes

A relative strength of fMRI compared to optical imaging methods is the accessibility of areas in cortical sulci. For analysis and visualization purposes, folded, inflated or flattened surface representations of the volumetrically sampled data have proven useful (Dale et al., 1999, Goebel et al., 1998). These triangulated surface representations are typically reconstructed along the white matter/gray matter (or mid-gray matter) boundary of a whole hemisphere from a segmented anatomical whole brain acquisition, e.g. T₁-weighted MPRAGE (Mugler and Brookeman, 1990) or more recent extensions of it (e.g. MP2RAGE, MEMPRAGE, Marques et al., 2010; van der Kouwe et al., 2008). The triangulation approach is the natural choice because it easily allows smoothing (removing its original voxelated appearance) and flattening, allowing visualization of sampled functional data as whole-hemisphere surface maps with minimal distortions. Spatially transforming a reconstructed hemisphere mesh into a sphere facilitates cortex based alignment in multi-subject analysis, usually providing better group-level results than volumetric template-based alignment approaches (e.g. in MNI or Talairach normalization, Frost and Goebel, 2012).

For separating cortical layers with high resolution fMRI, the surface approach has been extended by adding multiple cortical meshes at different cortical depths (Fig. 1a). These can be generated by evolving a white matter/gray matter (WM/GM) boundary surface towards the boundary with cerebrospinal fluid (CSF) along vertex normals, in conjunction with a cortical thickness estimate, which requires segmentation of both the WM/GM boundary as well as the GM/CSF boundary (Dale et al., 1999, Fischl and Dale, 2000, Polimeni et al., 2010). This procedure automatically grants correspondence between the same vertices (mesh nodes) at different cortical depths, which would be lost if independent surfaces were created for different cortical depths. Constraints can be applied to enforce smooth curvature (second-order smoothness) and uniform vertex spacing, and to prevent self-intersection (Fischl and Dale, 2000).

Fig. 1.

High-resolution cortical depth analyses using Cartesian grids. (a) The standard mesh approach first reconstructs a triangulated mesh at the voxel boundaries of a WM/GM (or mid-grey matter) boundary of a segmented volume. In order to better fit the underlying smooth cortical sheet, the voxelated mesh is smoothed leading to unequal distances and angles between vertices. In order to keep correspondence across depth, additional meshes at any relative depth level are created following vertex normals. (b) In the Cartesian grid approach, smooth meshes are directly created within grey matter based on a smooth vector field obtained by solving the Laplace equation (Jones et al., 2000). In order to keep correspondence across depth, additional grids at any relative depth level are created by following streamlines. (c) Besides placing grids at equi-distant relative depth levels, depth can be adjusted following the equi-volume model. Using small segments of connected grid points forming depth frustums, grid points can be adjusted in depth in order to keep layer volumes constant in folded cortex; the schematic explanation uses 3 grids and, correspondingly, 2 layers. (d) Results from equidistant and equi-volume sampling using 4 grids (3 layers) in a cut of a three-dimensional section of a segmented brain; left panel shows partial view of created grids while right panel shows filled layers in corresponding slice.

Construction of high resolution cortical depth grids

The surface mesh approach based on triangulation described above is very suitable for whole hemisphere representations because it works robustly in the presence of highly variable curvatures. However, distances and angles between neighboring nodes vary strongly and thereby impede the analysis of (quantitative) spatial relations. This issue results from reconstructing the voxelated boundary of segmented anatomical data followed by mesh smoothing to obtain a model of the cortical sheet (Fig. 1a). A further issue of whole-hemisphere meshes are distortions in local distances and angles introduced when smoothing meshes or, more pronounced, when flattening meshes for visualization.

When only a small region of cortex is of interest, a Cartesian grid approach (introduced in Zimmermann et al. (2011)) promises more practicability because it creates smooth regular cortex grids directly in grey matter at a specified spatial resolution instead of starting with meshes reconstructed at white matter voxel boundaries (Fig. 1b). Such regular 2D grids better reflect local distances and angles when quantitatively analyzing and visualizing sampled (e.g. functional) data since no mesh resampling is necessary. Note that in the standard mesh approach, the number of neighbors of a vertex depends on the originally reconstructed voxel boundaries (e.g. a vertex has less neighbors if located in a straight wall than if located at the fundus of a sulcus). After smoothing the mesh, such differences lead to different angles between edges formed by the triangles, which connect a vertex with its neighbors. These irregularities result in inhomogeneous density sampling of the targeted depth surfaces. Since regular grids are created explicitly following iso-surfaces, each vertex is always surrounded by 4 neighboring vertices at a constant distance forming 90 degree angles between edges (distances stretch or shrink as in the mesh approach when projecting mid-GM surfaces to other depth levels in curved cortex). The explicitly created regular grids simplify accurate post-processing of sampled functional data such as shortest distance, gradient and curvature calculation. Spatial mesh smoothing within layers, for example, can be implemented using standard Gaussian kernels for 2D Cartesian coordinates. While two-dimensional operations such as within-layer (Gaussian) smoothing of sampled functional data is also possible with standard meshes (Ahveninen et al., 2016, Andrade et al., 2001, Jo et al., 2007, Kiebel et al., 2000), they are more straightforward and precise when using regularly designed grids since they do not require resampling. How valuable these expected advantages are in specific cases needs to be shown by direct comparisons of the whole-mesh and cortical grid approach in future work.

The high resolution cortical grids that the authors have used in recent studies (De Martino et al., 2015, De Martino et al., 2013, Kemper et al., 2015b, Zimmermann et al., 2011) are based on the following procedure: After brain segmentation of WM, GM, and CSF, the mid-GM surface is found via an assessment of cortical thickness using the Laplace method (Jones et al., 2000). The smooth intensity field within GM obtained by the Laplace method allows calculation of a gradient of intensities pointing in the direction of cortical depth at each GM voxel (Fig. 1b). Integrating along these gradient values produces “field lines” or “streamlines” (Jones et al., 2000) establishing one-to-one radial correspondence between the WM/GM and GM/CSF surface for subsequent modelling. A regularly spaced Cartesian grid is created around a seed point in the mid-GM surface traversing in orthogonal direction with respect to the streamlines, i.e. along the tangent plane. Importantly, the mid-GM surface is not directly taken from the middle equipotential isocontour, but generated at middle cortical thickness to yield an equidistant cortical depth grid. Within this surface, regular metric distances can be guaranteed (Fig. 1b, blue grid). Then, grids at any deeper and more superficial relative depths are added by evolving the mid-GM regular grid towards the WM/GM or GM/CSF boundary following streamlines (Fig. 1b, green grid). Because of the unique correspondence between grid points at different cortical depths, this method is particularly suited to study possible columnar organizations of functional responses. To establish correspondence across grid points in a piece of curved cortex, equal sampling distances of sampling points at cortical depths other than the starting one must be, however, violated (indicated by “stretched” green grid in Fig. 1b). Note that the grid resolution oversamples the voxel resolution at least twice in order to ensure sufficiently dense voxel sampling in upper and lower grids despite expansions and shrinkages of distances between corresponding points at different depth levels in folded cortex.

Cortical depth grids provide advantages for visualization of functional depth-sampled data. First, grids can easily be visualized in 2D format without distortions introduced by flattening standard meshes. Second, a stack of depth grids with associated sampled functional data can also be converted in volume space with two dimensions representing the extent of cortex and one dimension representing cortical depth. These grid volumes allow using standard volume tools for analysis and visualization including volume rendering and slicing, which is helpful for inspecting consistency of responses across depth, i.e. columnarity.

An application of this modeling approach is demonstrated in Fig. 2. The results of an ocular dominance experiment at 9.4 T are overlaid on a high resolution cortical depth grid. The direct correspondence of sampling points at different cortical depths allows for inspection and straight-forward quantitative analysis (De Martino et al., 2015) of (dis-)similarity of results at different cortical depths, while the columnar pattern tangentially to the cortex can also be observed.

Fig. 2.

: Results of human 9.4 T visual fMRI experiment targeting ocular dominance. Selectivity index s represents degree of exclusiveness of response to left eye (negative values) or right eye (positive values) in monocular visual stimulation, defined as s = (β_R– β_L)/( β_R+ β_L), where β_R and β_L represent BOLD signal response amplitudes to left and right eye stimulation, respectively. (a) Ocular dominance responses in early visual cortex (V1 and V2 boundary delineated based on pRF mapping experiment) on inflated cortex representation. View from posterior onto the occipital pole. (b) High resolution cortical grid volume representations of regions of interest in V1 (equidistant sampling along cortical depth). The gaps in the rendered representations allow inspection of the consistency of ocular dominance across cortical depth. Results from T₂-weighted BOLD fMRI using 3D-GRASE (0.7 mm isotropic nominal resolution) overlaid on anatomical reference generated from T₁-weighted MPRAGE (see methods section for details).

Equi-volume model

In our previous publications (e.g. De Martino et al., 2015; Zimmermann et al., 2011) grids were created at cortical depth levels following an equidistant approach keeping a specified relative cortical depth level to the cortex boundaries, i.e. the absolute distance at a local region for a specified relative depth level is obtained by multiplying the depth level by the measured local cortical thickness. Here we describe an extension of regular depth grid creation integrating the equi-volume approach (Waehnert et al., 2014), which has also been applied in fMRI (Huber et al., 2015, Kok et al., 2016). The equi-volume model adjusts the thickness of layers in cortical segments to preserve their volume compensating for cortical folding. Using the geometric properties of the grid sampling approach, this principle can be accommodated simply by calculating respective (sub-)volumes in a depth volume frustum (portions of solids that lie between two parallel planes, see Fig. 1c) since varying curvature is explicitly reflected in connections of corresponding points across depth by expanding/shrinking distances to neighboring grid points across layers. In flat cortex, distances to neighboring grid points will be constant and calculated volume fractions correspond to corresponding relative distances, i.e. in flat cortex the results of the equidistant and equi-volume model are identical. In folded cortex, however, the height of layers (i.e. distance of intermediate grids) need to be adjusted in order to keep constant relative volumes which are calculated using the volume equation for frustums. More specifically, a cone frustum is calculated at each grid position by connecting the corresponding points at the WM/CSF and GM/CSF boundary forming a vertical segment vector s and by calculating the area of a disk (indicated in Fig. 1c) at the WM and GM boundary (A_WM, A_GM) oriented orthogonally to the vertical segment vector. A cone frustum (instead of a frustum with a rectangular base) is used, in order to ensure that the frustum volume equation is also applicable in cases with substantially different curvatures in different directions (e.g. saddle points). The volume of the created cone frustum can be calculated based on the height h of the vertical segment and the disk areas at the WM and GM boundary. In order to calculate the sub-volume of any layer, the calculated disk radii at the WM and GM boundaries are interpolated linearly to any intermediate depth level; the interpolated radii are used to calculate adjusted disk areas from which the volume of any sub-frustum (layer) can be obtained (Fig. 1c). A comparison of equidistant and equi-volume grid sampling is shown in Fig. 1d.

Fig. 3(a) shows an application of cortical depth sampling using human anatomical data acquired at 9.4 T. High resolution (0.35 mm isotropic) T₂^*-weighted images were acquired and co-registered with a T₁-weighted reference image, which was used for WM/GM/CSF segmentation. Cortical depth profiles were assessed within retinotopically defined primary (V1) and secondary (V2) visual areas (see below) according to the equidistant and equi-volume models. The highly myelinated stria of Gennari, which is known to have short T₂^* and hence low intensity in T₂^*-weighted images (Bridge et al., 2005, Budde et al., 2011, Duyn et al., 2007, Fukunaga et al., 2010, Sánchez‐Panchuelo et al., 2012) is expected to generate a dip in the depth dependent profiles of area V1 (but not V2) at middle cortical depths. Even at this very high resolution the results between the two methods are very similar. Nonetheless, closer inspection unveils that equidistant sampling yields slightly smoother profiles (most noticeable in participant 2, left hemisphere), as a consequence of sampling in convex and concave areas. When sampling concave and convex areas separately (i.e. sulci and gyri, respectively), systematic differences between the two sampling methods emerge (see Fig. 4). In line with the ex-vivo findings of Waehnert et al. (2014), the trough is located at the same middle cortical depth in concave and convex areas of V1 with equi-volume sampling, while it is shifted using equidistant sampling.

Fig. 3.

: (a) Cortical depth profiles of high resolution T₂^*-weighted anatomical data sampled in V1 (green) and V2 (red) according to equi-volume (solid lines) and equidistant (dashed lines) cortical models. Plot represents region averages within one hemisphere, each. Shaded areas represent std. errors. The gray band represents the approximate position of the Stria of Gennari as assessed from Nieuwenhuys (2013). (b) High resolution T₁-weighted reference images (left column, 0.6 mm isotropic interpolated to 0.35 mm isotropic), and T₂^*-weighted images (right column, 0.35 mm isotropic native resolution), superimposed with GM segmentation (middle row) and V1/V2 regions of interest (bottom row). Note the clear appearance of the Stria of Gennari in the T₂^*-weighted image (black arrows) and dark lines just underneath the gray matter in white matter (white arrows). See also supplementary materials S1 and S2 for multi-slice T₂^*-weighted images in two participants.

Fig. 4.

: Cortical depth profiles of high resolution T₂^*-weighted anatomical data sampled in V1 separately for concave areas (light green, sulci), convex areas (dark green, gyri), and combined (medium green), averaged across all hemispheres. Equi-volume (solid lines) and equidistant (dashed lines) sampling are drawn separately. A dotted line at middle cortical depth (50%) is drawn for visual guidance.

Spatially pooled sampling alternatives

When considering the multiple alternatives of depth sampling, the chosen analysis strategy has to serve the specific research question. For example, obtaining regular grids over an entire hemisphere is not possible. In such cases, the triangulation approach seems currently most suitable. For this reason, triangulated surfaces at different cortical depths have been used in anatomical (e.g. De Martino et al., 2014b) and functional (mesoscopic) studies (Goncalves et al., 2015). The latter sampled only middle or deep and middle cortical depths and ignored superficial layers because of the well-known detrimental effect of macrovasculature in these layers (Polimeni et al., 2010, see also in ultra-high field imaging section below).

To study feedforward-feedback relations, the cortical depth profile of a region of interest may be sufficient. This requires an accurate definition of the brain area under investigation and assumes that in a homogeneously tuned cortical area with similar receptive fields (or anatomical properties, De Martino et al., 2014b) in order to average signals (Huber et al., 2015, Kok et al., 2016, Scheeringa et al., 2016) or information (Muckli et al., 2015) across many voxels at the same cortical depth. Evidently, this approach significantly reduces the demand for high SNR because of the averaging across voxels while allowing much finer cortical depth resolution than the voxel size.

Limitations

A common limitation of the various cortical depth sampling approaches described above is that they rely on the segmentation of the cortex derived from an anatomical reference image, most commonly a T₁-weighted image at submillimeter resolution. Two error sources are therefore common to various cortical sampling algorithms: (1) Segmentation errors of the WM/GM boundary and the GM/CSF boundary lead to an erroneous definition of the cortical depth. Additional anatomical contrasts (e.g. additional T₂^*-weighted acquisition) may help to reduce this ambiguity (De Martino et al., 2014a). (2) Motion and distortion artefacts cause misregistration to the anatomical reference across extended areas. Advanced coregistration techniques such as boundary-based registration may improve the results (Greve and Fischl, 2009). Independently of the cortical sampling technique, these two aspects will always limit the correctness of assigned cortical depth and surface position. To avoid some of the issues present when aligning to a standard anatomical data set, a reference image with the same distortions as the functional data can be used (Renvall et al., 2016). This can be achieved either by acquiring additional images with strong anatomical contrast (e.g. T₁-weighed EPI references), or by segmenting the functional data directly (e.g. in combination with complex phase information, Fracasso et al., 2016).

Another limitation common to all sampling methods is the intrinsically limited effective spatial resolution of the functional data, which is below the nominal resolution. Not only do interpolation and misregistration impede the effective spatial resolution, but so do image blurring and functional (BOLD) signal point-spread, which will be discussed in the following section.

Our data (see Fig. 3) support the previously expressed notion (Fracasso et al., 2016, Guidi et al., 2016, Waehnert et al., 2014) that although equi-volume depth sampling according to the Bok model (Bok, 1929, Waehnert et al., 2014) yields a small advantage, the differences with equidistant sampling are limited in the context of currently feasible in-vivo functional experiments. This is also true for the ocular dominance data (Fig. 2): If resampled according to the equi-volume model, only 0.2% of the labeled grid points switch ocular dominance from left to right or vice versa, and 90.7 % of the labels (right/left/undefined) in the sampled grid stay identical. These results suggest that previous fMRI studies using equidistant sampling would yield only slightly different results when using equi-volume sampling.

Ultra-high field for high SNR at high isotropic resolution

As discussed above, high SNR is a prerequisite for approaches that consider each voxel’s response individually, and ultra-high field scanners can provide this. As in structural imaging, the baseline signal has better SNR because of the improved physical spin statistics and more favorable interaction between the RF field and the biological tissue (Ugurbil et al., 2003). In addition, an approximately linear increase in dynamic blood oxygenation level dependent effect (BOLD effect) size is theoretically expected (Uludag et al., 2009) and experimentally observed with stronger magnetic fields (Yacoub et al., 2003, Yacoub et al., 2001). Both effects, better image SNR and larger BOLD effect size, independently improve CNR and counterbalance the reduction of SNR due to higher spatial resolution.

The aforementioned improvements in CNR are expected to continue also beyond the field strength of 7 T, which is currently the highest field strength accessible for most human fMRI. Improvements in SNR have also been demonstrated experimentally (Budde et al., 2014, Pohmann et al., 2015). Fig. 5 showcases the CNR benefit in results of a routine human 9.4 T experiment. We acquired population receptive field (pRF, Dumoulin and Wandell, 2008) data to map the retinotopic organization in human visual cortex (Goebel, 2015). The figure shows example pRF mapping results from 9.4 T and histograms of the voxelwise correlation values, R, in comparison to data from a previous 7 T experiment with similar imaging properties (see methods section for details). R values close to 1 reflect good agreement between voxel’s time courses and the modelled pRF responses taking into account its horizontal and vertical position and (Gaussian-shaped) receptive field in the visual field. Median R in the 9.4 T participants is significantly higher than at 7 T in V1 and V2 (0.68 ± 0.07 vs 0.52 ± 0.10 at p<0.02 (two-sided Wilcoxon rank-sum test considering all 7 T hemispheres and median results for each 9.4 T hemisphere)). This demonstrates the feasibility of routine experiments employing relatively high spatial resolution (1.05 mm isotropic) using GE-EPI at 9.4 T.

Fig. 5.

(a–d) Population receptive field mapping results from 9.4 T human fMRI experiments displayed on inflated cortical mesh representations. View from posterior onto the occipital pole. Eccentricity (a), visual angle (b), pRF sizes (c), and resulting delineation of V1 and V2 (d) are shown for 4 hemispheres of two participants. e) Comparison of voxelwise correlation values between 9.4 T and similarly acquired 7 T data (red dashed line and blue solid line, respectively). Shaded areas indicate standard deviation across hemispheres (12 measurements at 9.4 T, 22 hemispheres at 7 T, see Methods section for details).

While high isotropic resolution sampling of fMRI data yields additional advantages over anisotropic sampling (e.g. pooling across depth or pooling across a cortical area), the underlying fMRI data has to support the spatial specificity to make this evident. To this end, gradient-echo (T₂^*-weighted) BOLD fMRI data as typically acquired, comes short compared to more specific acquisition techniques such as T₂-weighted BOLD imaging, cerebral blood volume (CBV) weighted or cerebral blood flow (CBF) based methods (Duong et al., 2001, Duong et al., 2003, Duong et al., 2002, Goense et al., 2016, Goense and Logothetis, 2006, Yacoub et al., 2007). A widespread use of these latter methods is hindered by the significantly smaller sensitivity compared to T₂^*-weighted BOLD imaging. Nevertheless, our early 9.4 T results (Fig. 2) show that T₂-weighted BOLD fMRI can be used to unveil fine-grained patches of differential ocular dominance. Differently from earlier studies targeting ocular dominance columns, which used anisotropic resolution with thick slices spanning the entire cortical thickness (Cheng et al., 2001, Goodyear and Menon, 2001, Yacoub et al., 2007), the consistency across cortical depth can be probed directly due to the high isotropic spatial resolution (0.7 mm isotropic). Moreover, recent modeling results suggest that in T₂-weighted BOLD fMRI, the laminar location of measured signal is linked more closely to the actual lamina, whereas T₂^*-weighted BOLD signals drift towards superficial laminae as they go downstream in the draining veins (Markuerkiaga et al., 2016). This insight extends the previously demonstrated, well established finding that T₂^*-weighted BOLD signals increase towards the surface (e.g. De Martino et al., 2013; Goense and Logothetis, 2006; Kok et al., 2016; Polimeni et al., 2010; Zhao et al., 2004), in that it identifies the increased superficial T₂^*-weighted signal as the additive signal stemming from deep and superficial layers. Nonetheless, (purely) T₂-weighted signal contains also unspecific macrovascular contributions at ultra-high field (Uludag et al., 2009), and results with regard to the laminar origin of signals have to be interpreted carefully.

Besides relatively low sensitivity, T₂-weighted imaging has other limitations: (1) Pure T₂-weighting without T₂^* contamination is difficult to achieve in echo-train imaging pulse sequences such as EPI (whereas non-EPI approaches are limited by their temporal inefficiency) and (2) the RF energy deposition requires compromises between spatial resolution, imaging volume, SNR, and temporal resolution. We address these challenges at 9.4 T with the application of the 3D-GRASE imaging pulse sequence and the use of a local, RF transmit efficient surface array in combination with a high SNR receive array. 3D-GRASE utilizes a slab selective excitation pulse and orthogonal refocusing pulses (Feinberg et al., 2008, Feinberg and Oshio, 1991). This creates an inner volume selection, reducing the necessary phase-encoding steps, while also obviating the need for multiple excitation pulses as in 2D multi-slice imaging. The local RF transmit coil reduces RF power deposition in unrelated portions of the head (see supplementary Figure S3). Parallel transmit techniques promise to further increase homogeneity in the imaging FoV (De Martino et al., 2012, Poser et al., 2014 and references therein; Tse et al., 2016; Wu et al., 2013).

Summary

The relatively new field of high field high resolution fMRI allows investigation of cortical columns and cortical laminar structures in humans. Considerations with regard to analysis strategies for high resolution functional data continue to play an important role in the literature. In this article, we propose that, depending on the neuroscientific application, high resolution data may be pooled across a cortical patch to study cortical depth dependent effects (i.e. laminar profiles). If the intention is to map the tangential layout of a cortical region without investigation of depth-dependent (i.e. columnar) information, data may be pooled across cortical depth (while potentially ignoring signals in superficial laminae). If individual voxel’s responses are of interest and may not be pooled across any dimension, a set of cortical depth meshes may be a solution for sampling information tangentially while keeping correspondence across depth. For quantitative analysis of the spatial organization in fine-grained structures, a cortical grid approach is advantageous with respect to calculations and visualizations. Combined with a cortical layer volume-preserving (equi-volume) model, the cortical grid approach can also be applied to research questions that allow for spatial smoothing within or across layers. We demonstrated and discussed that the difference between equidistant and equi-volume sampling is minor given the current limitations of participant motion, coregistration, segmentation, and fMRI voxel size, however, equi-volume sampling yields an advantage when investigating subtle differences in cortical depth profiles. Our 9.4 Tesla human anatomical and functional data indicate the advantage over lower fields including 7 T and demonstrate the practical applicability of T₂^* and T₂-weighted fMRI acquisitions.

Materials and methods

Experimental setup and participants

All experiments were performed at the research facilities of Maastricht University and were in accordance with the local ethical board. After giving informed consent, two healthy female participants (mean age 25 ± 1) were scanned in two sessions on a 9.4 T MRI scanner (Siemens, Erlangen, Germany) equipped with a head-only gradient coil (max. 80 mT/m at 400 mT/m/s slew rate) and an 8-channel transmit 24-channel receive surface coil (Life Services, Minneapolis, MN, USA) covering the posterior part of the participants head. A static, non-subject-specific B₁⁺ phase shim was used throughout the measurements, after it had been validated that this approach generated sufficiently efficient and homogeneous RF fields in early visual areas in various participants (Kemper et al., 2015a) using DREAM B₁ mapping (Nehrke and Börnert, 2012, Tse et al., 2014). Only the global scaling factor (reference voltage) was adjusted to the individual participants after acquiring a pre-saturation approach based flip angle map. SAR supervision was maintained by imposing the maximum local SAR level observed in simulations as global RF power deposition limits.

Second order shimming was applied to homogenize the B₀ magnetic field in a manually drawn region of interest in early visual areas using a custom-made dual-echo field mapping GRE sequence and MATLAB (The MATHWORKS Inc., Natick, MA, USA) routines (Tse et al., 2016).

Visual stimulation

Visual stimulation was provided using a projector and a frosted screen near the participant’s head, seen via a tilted mirror. PRF mapping was performed using randomly moving bars (Senden et al., 2014) in the freely available stimulation software BrainStim (https://github.com/svengijsen/BrainStim). Horizontal, vertical, and oblique full contrast flickering checkerboard bars (flickering rate 7.5 Hz) were presented at 12 different positions spanning 0.63° visual angle each with maximum eccentricity of 7.5° (square-shaped). During anatomical acquisitions, participants watched animated cartoon movies.

For binocular eye stimulation using PsychoPy (version 1.82, Peirce, 2007), participants wore anaglyph glasses (Lee filters, Andover, Hampshire, UK; red #026, cyan #116) throughout the entire second experimental session. The stimulus colors were adjusted such that the eyes could be stimulated individually using the pass-band optical wavelength of the filter plates. Participants were instructed to fixate a central fixation dot, while full field concentric flickering checkerboard patterns were presented. The fixation dot was moved 2.5° upwards, such that more of the lower visual field was stimulated. A block design switching between right and left eye every 48 s was chosen (Yacoub et al., 2007). In the beginning and end 24 seconds of gray screen were presented at the same luminance as the non-stimulated eyes perceived. Seven runs of 7:12 min duration each were performed.

Imaging pulse sequences

Anatomical reference images were acquired with a T₁-weighted MPRAGE sequence (TE/TI/TR = 2.5/1200/3600 ms; nominal flip angle = 4°; matrix size = 384×384×256; GRAPPA acceleration factor 3, partial Fourier factor 6/8 (partition direction); total duration 8:54 min.). In addition, a proton density weighted MPRAGE without the inversion module (identical imaging parameters except TR = 1620 ms and total duration = 4:01 min.) was acquired to correct for transmission and receive coil biases (Van de Moortele et al., 2009). TI and flip angle were adapted from 7 T protocols based on literature values for T₁ and proton density in WM and GM at 7 and 9.4 T and the equations in Deichmann et al. (2000) and Marques et al. (2010).

PRF mapping was performed using a gradient-echo EPI sequence (1.05 mm isotropic nominal resolution; TE/TR = 14.6/2000 ms; in-plane FoV 105×105 mm; 54 coronal slices; nominal flip angle 70°; echo spacing 0.64 ms; GRAPPA 2, partial Fourier 7/8; 13:20 min. duration, 3 repetitions).

Anatomical high resolution T₂^*-weighted images were acquired using a 2D GRE sequence with limited FoV (0.35 mm isotropic nominal resolution (0.043 mm³); TE/TR = 20/1500 ms; in-plane FoV 100×100 mm; phase-encoding direction left-right; 48 slices; nominal flip angle; 70°; bandwidth 80 Hz/Pixel; no parallel imaging, no partial Fourier; 7:09 min. duration, 6 averages). The narrow FoV in phase-encoding direction causes aliasing of anterior lateral portions of the images, which are not of interested and discarded in the analysis (see below).

Ocular dominance imaging was performed using a high-resolution inner-volume 3D-GRASE sequence (TE/TR =30.85/2000 ms; in-plane FoV 22.4×105 mm; 12 slices; nominal flip angles (90-107-60-56-53-54-54-58-70-107)°; echo spacing 0.75 ms; partial Fourier 6/8 (partition direction); 8:00 min. duration, 7 repetitions). Images were acquired at a nominal isotropic resolution of 0.7 mm, and were then interpolated to 0.35 mm isotropic resolution using the complex, uncombined coil data. This was done in order to minimize resolution losses due to coregistration or motion-correction. Refocusing flip angles were adjusted based on extended phase graph theory simulations to yield reduced across-slice blurring of estimated 2.3 voxels (Kemper et al., 2016).

Data analysis

Functional data analysis was performed using BrainVoyager QX 2.8.4 (Brain Innovation, Maastricht, the Netherlands) and MATLAB routines. Preprocessing of the pRF data included slice-time correction, 3D-motion correction, and high-pass filtering. Additionally, data was distortion-corrected using acquisitions with reversed phase encoding direction and the COPE plug-in (v0.5) in BrainVoyager (Andersson et al., 2003, Fritz et al., 2014). The three runs were averaged and analyzed to generate pRF estimates for each voxel. Model time courses for pRFs of receptive field positions between ± 7.5° eccentricity in horizontal and vertical position and size (Gaussian full width at half maximum) between 0.2° and 10° were calculated, to which the experimental data was compared to find the best match (highest correlation). Functional data were co-registered to the intra-session reference anatomical data using positional and edge information. Anatomical alignment across sessions was used to align all (functional and anatomical) data to the anatomical reference image of the second (ocular dominance) session. The resulting maps for voxel-wise retinotopic eccentricity, polar angle, and pRF size were projected onto the surface representations of the hemispheres (see below). Visual areas V1 and V2 were delineated by the reversal of polar angle at the vertical meridian. Delineation was possible down to an eccentricity of approximately 2°. Below this limit, the polar angle did not yield sufficient information.

After delineation of visual areas V1 and V2, the voxelwise correlation values R were analyzed and compared to data acquired similarly in 11 participants of a previous study at 7 T (Emmerling et al., 2016). The spatial resolution of the 7 T datasets was 1.1 mm isotropic. The same number of volumes (first 300 out of the 400 available per run for the 9.4 T data) was used in all pRF mapping runs, and preprocessing and analysis were matched closely (see Emmerling et al. (2016) for further details).

Preprocessing of the ocular dominance functional data only included 3D-motion correction and linear trend removal. Individual runs were aligned with one another and co-registered with the T₁-weighted anatomical acquisition using positional information. Fine-tuning of the alignment was performed manually using image edge information. A general linear model was calculated using the right and left eye condition predictors after convolution with a standard 2-gamma-function hemodynamic response function. Additionally, motion parameters were used as confound predictors (three translational and three rotational per run). As a measure for eye preference, we calculated selectivity, s, as the normalized relative contribution between the BOLD responses to the right and left eye (s = (β_R−β_L)/( β_R + β_L)). The following three criteria were used to identify voxels with reasonable characteristics to analyze eye preference on: (1) Voxels had to have a small eccentricity in the pRF measurement below 5° visual angle. This limit was chosen because participants reported ocular rivalry “creeping in” from peripheral visual field areas, such that, at times, participants would perceive the color-shaded gray background in the non-stimulated eye more pronounced than the checkerboards in the stimulated eye. Such effects are more likely to occur at higher eccentricity (Blake et al., 1992). (2) BOLD responses to right or left eye individually had to be significant (p < 0.001) and both positive to rule out areas of negative BOLD responses. (3) BOLD signal changes above a threshold of 7 % were discarded as likely macrovascular contributions. The group of voxels that fulfilled the criteria 1–3 (small eccentricity, small BOLD % signal change, significant activation for right or left eye individually) was considered to have sufficient signal quality for ocular dominance to be observed if present.

Permuting the spatial distribution of the selectivity index s allowed us to assess the null distribution of the median absolute s (median(|s|)) in V1 and V2 and thus the significance of the difference between V1 and V2. Spatial permutations were created by swapping at random N voxels from V1 and V2 (where N was determined as 70% of the smaller number of voxels fulfilling criteria 1–3 in V1 and V2). This procedure was repeated 1000 times to obtain the empirical null distribution of selectivity values against which the selectivity of V1 and V2 were compared (see Supplementary Fig. S4 for results).

The anatomical high-resolution T₂^*-weighted images were co-registered to one another and then averaged. Prior to that, the images were cropped such that no aliasing in phase-encoding direction was present. The averaged dataset was divided by a strongly spatially smoothed version of it (Garcia, 2010). This step retained local anatomical contrast but eliminated low-spatial-frequency image inhomogeneity stemming from the RF coil profiles. The brain was masked for this step based on intensity so that the local average was not reduced at the edges. The resulting image was co-registered with the MPRAGE anatomical reference.

The T₁-weighted anatomical reference images were divided by the proton density weighted images. Further, low-spatial frequency inhomogeneity correction was applied. Automatic image segmentation of WM and GM was performed and refined manually. Finally, the hemispheres were separated and individual surface representations were created from the mid-GM sheet.

Appendix A. Supplementary material

Supplementary figure S1 (online only): Animated visualization of T₂^*-weighted multi-slice data of participant 1. .

.

Supplementary figure S2 (online only): Animated visualization of T₂^*-weighted multi-slice data of participant 2.

.

Supplementary material

.