A blank voxel is not empty brain

Your cortex is a thin sheet crumpled into your skull. It is where the thinking mostly happens, and on its own terms it is two-dimensional: a place on the cortex is a pair of coordinates across a surface, the way a place on a page is a pair of coordinates across paper. Folding it does not change that. It only hides it.

Modern atlases store it that way. The CIFTI format keeps the cortex as surface vertices and the deeper structures as volume voxels. It follows the real geometry, and it does not waste space sampling the white matter and fluid the sheet is not in.

But most brain software only reads blocks. Volume formats like NIfTI (the standard for SPM, FSL, MRIcroGL and a great deal else) want a 3D grid of cubes. So a lab that wants to use a modern surface atlas inside an older tool has to press the sheet back into a block, and that conversion is lossy in a specific and quietly awful way.

Two things can go wrong, and they pull against each other. Too conservative, and voxels that are genuinely cortex come out unlabeled: blank spaces in the map where the brain obviously is. Too permissive, and labels spill past the cortex into places they have no business being.

A team at South China Normal University built a workflow called Surf2Volume to measure the trade-off rather than eyeball it. They took the Schaefer2018 atlas (100 parcels, 7 networks), which happens to exist in both forms, a surface version and a published volume version. That gave them something rare: a conversion they could score against a reference.

The pipeline routes labels from the surface representation through fsaverage to a surface version of the MNI152 template, then rasterizes them inside a smoothed cortical ribbon mask. One number controls how much of that mask counts as cortex: a threshold on the smoothed mask. Lower thresholds include more voxels; higher thresholds include fewer.

They swept twelve thresholds. At 0.05, the conversion scored a Dice of 0.729 and a Jaccard of 0.580, both slightly better than the best Workbench values (0.709 and 0.556). It left 11.88% of reference voxels unlabeled, against 20.57% for standard Workbench ribbon mapping. Labels outside the reference fell from 11.22% to 9.43%. Among the thresholds tested, 0.05 also produced the lowest average parcel volume distortion.

The good news is real. The other news is that about one in nine reference voxels still comes out blank, and no threshold fixes that, because the problem is not the algorithm. It is the mismatch between a sheet and a grid.

The paper is also careful about something most benchmark papers would rather bury. Workbench at 7 mm had less missing support and higher voxel-wise accuracy than Surf2Volume at 0.05, but substantially more labels outside the reference and greater parcel volume distortion. There is no setting that wins every measure. And the reference volume itself is not biological ground truth. It is another projection of the same surface atlas, done by different hands. They are grading one map against another map, because the actual cortex has no voxels to check against.

What stays with me is what a blank voxel looks like downstream. In the data it is indistinguishable from absence. An unlabeled voxel does not announce that it is a boundary artefact. It sits there being zero, in a file a statistician will read as a place where nothing happened. A shape mismatch can wear the face of a finding.

The paper's own advice is the tell: keep the original surface atlas when fine cortical topography is what you are asking about. The volume version is a convenience, and the convenience costs something. Threshold 0.05 is the best setting for this benchmark, on this atlas, on this grid. It is not a universal constant.

The brain never had voxels. We did.

Sources