Spatial image preprocessing (distortion correction and image alignment) was carried out using the HCP’s spatial minimal preprocessing pipelines5 (
link). This included steps to maximize alignment across image modalities, to minimize distortions relative to the subject’s anatomical space, and to minimize spatial smoothing (blurring) of the data. The data were projected into the 2 mm standard CIFTI grayordinates space, which includes cortical grey matter surface vertices and subcortical grey matter voxels5 (
link). This offers substantial improvements in spatial localization over traditional volume-based analyses, enabling more accurate cross-subject and cross-study registrations and avoiding smoothing that mixes signals across differing tissue types or between nearby cortical folds. Additionally, we did minimal smoothing within the CIFTI grayordinates space to avoid mixing across areal borders prior to parcellation.
For cross-subject registration of the cerebral cortex, we used a two-stage process based on the multimodal surface matching (MSM) algorithm14 (
link) (see
Supplementary Methods 2.1–2.5). An initial ‘gentle’ stage, constrained only by cortical folding patterns (FreeSurfer’s ‘sulc’ measure), was used to obtain approximate geographic alignment without overfitting the registration to folding patterns, which are not strongly correlated with cortical areas in many regions. Previously, we found that more aggressive folding-based registration (either MSM-based or FreeSurfer-based) slightly decreased cross-subject task-fMRI statistics, suggesting that aligning cortical folds too tightly actually reduces alignment of cortical areas14 (
link). A second, more aggressive stage used cortical areal features to bring areas into better alignment across subjects while avoiding neurobiologically implausible distortions or overfitting to noise in the data. The areal features used were myelin maps, resting state network maps computed with weighed regression (an improvement over dual regression34 (
link) described in the
Supplementary Methods 2.3) and resting state visuotopic maps (see
Supplementary Methods 4.4). Areal distortion was measured by taking the log base-2 of the ratio of the registered spherical surface tile areas to the original spherical surface tile areas. The mean (across space) of the absolute value of the areal distortion averaged across subjects from both registration stages was 30% less than the standard FreeSurfer folding-based registration and the maximum (across space) of this measure was 54% less. Despite less overall distortion, the areal-feature-based registration delivers substantially more accurate registration of cortical areas than does FreeSurfer folding-based registration as judged by cross-subject task fMRI statistics, an areal feature that was not used to drive the registration14 (
link). Because MSM registration preserves topology and is relatively gentle (it does not tear or distort the cortical surface in neurobiologically implausible ways), it is unable to align some cortical areas in some subjects where the areal arrangement differs from the group average (see
Supplementary Results and Discussion 1.3–1.4 for more details on atypical areas). Group average registration drift away from the gentle folding-based geographic alignment was removed from the surface registration35 (
link) (see
Supplementary Methods 2.5) to enable comparisons of this dataset with datasets registered using different areal features (for example, post-mortem cytoarchitecture). Group average registration drift is any consistent effect of the registration during template generation on the mean size, shape, or position of areas on the sphere (as opposed to the desired reductions in cross-subject variation). An obvious example is the 37% increase in average brain volume produced by registration to MNI space4 (
link). Uncorrected drifts during surface template generation can cause apparent changes in cortical areal size, shape, and position when comparing across studies.
Resting state fMRI data were denoised for spatially specific temporal artefacts (for example, subject movement, cardiac pulsation, and scanner artefacts) using the ICA+FIX approach, which includes detrending the data and aggressively regressing out 24 movement parameters36 (
link),37 (
link). We avoided regressing out the ‘global signal’ (mean grey-matter time course) from our data because preliminary analyses showed that this step shifted putative connectivity-based areal boundaries so that they lined up less well with other modalities, likely because of the strong areal specificity of the residual global signal after ICA+FIX clean up. Task fMRI data were temporally filtered using a high pass filter. More details on resting state and task fMRI temporal preprocessing are described in the
Supplementary Methods 1.6–1.8. Substantial spatial smoothing was avoided for both datasets, and all images were intensity normalized to account for the receive coil sensitivity field. Artefact maps of large vein effects, fMRI gradient echo signal loss, and surface curvature were computed as described in
Supplementary Methods 1.9.
Glasser M.F., Coalson T.S., Robinson E.C., Hacker C.D., Harwell J., Yacoub E., Ugurbil K., Andersson J., Beckmann C.F., Jenkinson M., Smith S.M, & Van Essen D.C. (2016). A multi-modal parcellation of human cerebral cortex. Nature, 536(7615), 171-178.