A cortical surface template for human neuroscience

Neuroimaging data analysis relies on normalization to standard anatomical templates to resolve macroanatomical differences across brains. Existing human cortical surface templates sample locations unevenly because of distortions introduced by inflation of the folded cortex into a standard shape. Here we present the onavg template, which affords uniform sampling of the cortex. We created the onavg template based on openly available high-quality structural scans of 1,031 brains—25 times more than existing cortical templates. We optimized the vertex locations based on cortical anatomy, achieving an even distribution. We observed consistently higher multivariate pattern classification accuracies and representational geometry inter-participant correlations based on onavg than on other templates, and onavg only needs three-quarters as much data to achieve the same performance compared with other templates. The optimized sampling also reduces CPU time across algorithms by 1.3–22.4% due to less variation in the number of vertices in each searchlight.

participants).One participant (221319) was excluded from the 7 T data analysis due to data quality (including the participant does not change any conclusions).

Demographics of validation data
The validation data are also datasets that have been previously collected.
The mean age ± standard deviation was 27.29 ± 2.35 yo.
The validation data used in some Extended Data Figures were from the Human Connectome Project (HCP) dataset.The 888 participants of the 3 T data were 469 females and 419 males.The age range was 22-37, and the mean age ± 2/13 standard deviation was 28.65 ± 3.72 yo.The 176 participants of the 7 T data were 106 females and 70 males.The age range was 22-36, and the mean age ± standard deviation was 29.36 ± 3.30 yo.

Resting-state functional connectivity
For the resting-state functional connectivity analysis, we computed the connectivity between each vertex and its neighbors, and averaged it across neighbors, scans, and participants.We refer to this average value as the local connectivity of the vertex.Local connectivity is known to be biased by the geometric shape of the cortex 63 .This is because vertices nearby are largely resampled from the same group of voxels, and their time series are linear combinations of the same set of basis.This issue is more severe when the vertices are closer, as in this case their weights of the linear combination are more similar to each other.
We performed the analysis using both synthesized noise and real fMRI data.
The real fMRI data are the 4 resting-state fMRI scans of the 3 T HCP dataset, and the synthesized data are volumes of similar shape and resolution, except that the numbers were randomly drawn from a standard Gaussian distribution.This allows us to separate the effects of cortical resampling from the effects of MRI's point spread function and the "true" local connectivity.We found that the onavg template reduced the bias of local functional connectivity for both synthesized noise and real fMRI data (Extended Data Fig. 7).

Functional contrast maps
The 3 T task fMRI data of the HCP dataset comprises 7 different tasks, from which 47 unique contrasts can be computed 64 .Some contrasts, such as "FACES-SHAPES" and "SHAPES-FACES," are actually the same contrast with flipped signs, 3/13 and in such cases we only used one of them.We computed each contrast in each participant's native space, and resampled the contrast to different template spaces.
The transform used for resampling was the sum of two barycentric transforms, that is, the transform from the native space to the template space, and the transpose of the transform from the template space to the native space.This ensures that every vertex in the native space is used in the resampling process.After deriving the transform, we normalized it so that the sum of each column equals one.This normalization ensures that the resampled data (template space) has the same scale as the original data (native space).
We computed the ISC of contrast maps in a similar manner as the RSA-ISC analysis.That is, for each participant, we computed the correlation between the participant's own map and the average of others' maps.The only difference was that this analysis was based on the contrast map of the entire cortex, whereas the RSA-ISC analysis was based on the vectorized RDM of each searchlight.Similar to the RSA-ISC analysis, we also estimated how this correlation would change with different amounts of data.When we compared the ISC across template spaces, we averaged the ISC across the 47 contrasts, and tested whether the average ISC differed between onavg and other template spaces (n = 888).For each comparison (onavg vs. fsavg, onavg vs. fslr), we also performed a permutation test with 100,000 repetitions, which allowed us to compare the actual differences with the corresponding null distributions (Extended Data Fig. 8).

Individual differences in functional architecture
We used neural responses to the movie from 178 participants of the HCP 7 T dataset for this analysis.We used 89 participants' data (training data) to build a functional template and performed the analysis using the remaining 88 participants 4/13 (test data; excluding 1 participant for data quality).Specifically, we modeled the neural responses to the movie of different participants as the same functional template transformed with individualized transforms.After applying the individualized transform to the template, we converted the transformed functional template into a tuning matrix, which depicts the neural tuning of the specific participant 36 .The tuning matrix was designed so that (a) vertices (or individuals) that are more different in response time series are also more different in their tuning profiles (i.e., columns of the tuning matrix), (b) the same tuning matrix can be estimated from responses to different stimuli.This allowed us to compute two estimates of the tuning matrix for each participant by dividing the participant's movie data into two halves and estimating the tuning matrix from each half.
For each pair of participants, we computed the similarities between their tuning matrix estimates.The similarity was always based on tuning matrices estimated from different halves of the movie, that is, the first half of the movie for one participant, and the second half of the movie for the other.Therefore, for each pair of participants, we had two tuning matrix similarities, and we used the average of them as the between-participant similarity for this pair of participants.Similarly, for each participant, we computed a within-participant similarity based on the two tuning matrix estimates of the participant.
After looping through all participants and participant pairs, for each participant, we had 1 within-participant similarity and 87 between-participant similarities.We computed the distinctiveness of the participant as the difference between the withinparticipant similarity and average between-participant similarity, divided by the standard deviation of between-participant similarity 36 .In other words, the distinctiveness measures how far away the within-participant similarity was from the 5/13 distribution of between-participant similarities, and higher distinctiveness means the differences between the participant and others are more prominent.
We repeated the analysis with different surface templates, and thus each participant had three distinctiveness values, one for each template.We compared the distinctiveness between onavg and the other two templates.We also repeated the analysis with different amounts of data, that is, 7 minutes, 14 minutes, …, up to the entire movie data (60.91 minutes).The onavg template consistently demonstrated higher distinctiveness than other templates across different amounts of data (Extended Data Fig. 9).