Skip to content

Smooth Surfaces and the Brainstem Body (audit fix, project 2 of 6)

Date: 2026-10-18 Status: described in discussion; builder check done; awaiting the owner's review Branch: spinal-cord (C2 stays unmerged until every audit item is fixed) Audit: .superpowers/audit-2026-10-13.md items 1, 2, 6 and 9

1. Why

Every shape looks like stacked 1 mm cubes: terraces on the cortex, hippocampus and nuclei (audit item 1). The brainstem has no body: AAL3 has no midbrain, pons or medulla, so the brainstem is a cloud of floating nuclei and the spinal cord hangs below it with a gap (item 2). Sources on different grids also look unevenly smooth (item 6), and 910 zero normals are repaired at each build (item 9).

The goal is surfaces that read as real organ surfaces, and a solid brainstem the nuclei sit inside, continuous with the cord.

2. Smooth surfaces (all builds)

The usual practice in anatomy atlases (owner accepted it as the standard choice):

  • Each structure's binary mask is smoothed before its surface is extracted (a Gaussian of 2 mm, see section 6, on a float copy of the mask, contoured at 0.5), then the surface is lightly evened with the existing Taubin pass. The 1 mm staircase disappears; gyri and sulci stay.
  • Accuracy beats smoothness for small structures: a small nucleus is blurred less, so its enclosed volume stays within 10% of its voxel volume and its centre moves less than 1 mm. Structures that would vanish or split are never blurred into nothing.
  • Every structure is smoothed on its own; neighbours never merge, and touching parts keep their own edges.
  • The same rule applies on every source grid (1 mm, 0.5 mm, 0.1 mm), with the blur expressed in millimetres so smoothness looks the same across sources.
  • Zero normals at pinch points should mostly disappear; the build report keeps counting any left.
  • Level-of-detail settings (lod0_max_faces, lod_fractions, min_faces) are the builder's call if smoothing changes how decimation looks; the face budget of the scene must not grow by more than 10%.

3. Brainstem body (local builds only)

3.1 Source (owner's decision)

  • Iglesias JE et al. (2015), Bayesian segmentation of brainstem structures in MRI, NeuroImage 113: 184-195, doi:10.1016/j.neuroimage.2015.02.065; the MNI ICBM 2009c (symmetric) version from the FreeSurfer wiki page "Subfield atlases in MNI-ICBM 152 space", file Brainstem.zip (sha256 f6ad672136102669955ee803f27c188a5e172a326a61fb9818d12fb4707fe5c2), member BrainstemProbs.MNIsymSpace.nii.gz with BrainstemProbs.MNIsymSpace.names.txt.
  • Grid 53 x 67 x 85 voxels of 1 mm, RAS, 4D probabilities: channel 0 background, 1 medulla (FreeSurfer 175), 2 pons (174), 3 superior cerebellar peduncle (178), 4 midbrain (173).
  • No licence stated: local_only, like the glands, the Brainstem Navigator and PAM50. A permission email to the authors goes into review/emails/ for the owner to send.
  • A voxel belongs to the brainstem when the brainstem as a whole (every non-background channel summed) is at least 0.5; it then goes to the most likely of the three kept parts. This leaves no cracks along the seams and fills the place of the dropped superior cerebellar peduncle channel (the atlas already has that peduncle as a tract from HCP-1065). Parts never overlap.

3.2 Structures

  • Three midline regions, one inside each existing group (my wording, easy to change): Midbrain body (in group 31 Midbrain), Pons body (in 32 Pons), Medulla body (in 34 Medulla). The groups and their nuclei stay as they are; the shareable build leaves the three bodies out and keeps every nucleus.
  • They rank below every existing label in the brain volume, so they never overwrite a nucleus in the slices, and are meshed on their own grid so their 3D shape stays whole.
  • No descriptions yet.

3.3 See-through bodies (owner's decision)

  • The three bodies are see-through by default, like glass, so the nuclei inside stay visible. They use a new structure kind, shell, for context shapes drawn see-through; the bodies are the only shells for now.
  • A click on the brainstem selects the nucleus under the mouse when there is one, and the body only where no nucleus is behind it. The same holds for Find quizzes ("click it in the 3D view").
  • Everything else treats a shell like any structure: the tree, hide, isolate, F to frame, the x-ray outline when it is selected, slices, quizzes and routes.
  • How see-through (opacity) is the builder's call, checked on rendered images: the body's shape must read clearly and the nuclei inside must stay easy to see. Shells must not slow rotation or selection beyond the project 1 budgets.

3.4 Where the bodies meet other shapes

  • The two atlases do not line up exactly: the bodies reach into the cerebellum (about 430 voxels of medulla into lobule IX, 400 of pons into lobules III, IV-V and X, 150 of midbrain into vermis III). The cerebellum wins: the bodies are trimmed against every non-brainstem label, so no two surfaces cross or flicker.
  • The medulla (to z -78) overlaps the top of the PAM50 cord (from about z -67): in a local build with both, the medulla is trimmed where the cord begins, so they meet without a gap or overlap.

4. Out of scope

The bamboo look of the cord segments and the colours (project 3), speed (project 4), slices (project 5), build-check gaps and the booklet (project 6), descriptions of the new parts.

5. Testing

  • Smoothing on a toy volume: a staircase surface becomes smooth (a curvature or normal-angle measure drops), a sphere's enclosed volume stays within 10% of its voxel volume, a small blob keeps volume within 10% and centre within 1 mm, two touching blobs stay two separate shapes.
  • Real build: every existing structure still has a shape; scene faces within 10% of before; zero normals reported; the slice labels unchanged (smoothing only changes surfaces).
  • See-through: shells are drawn see-through; a click over a nucleus inside a body selects the nucleus, a click on the body away from any nucleus selects the body; a Find quiz accepts a 3D click on a brainstem nucleus.
  • Brainstem: a local build adds the three bodies with centres near (0, -23, -13) midbrain, (0, -28, -33) pons and (0, -41, -55) medulla; the normal build leaves them out and keeps every nucleus; no body overwrites a nucleus label in slices; at the junction no medulla vertex lies more than 1 mm inside the cord's top segment.
  • A visual check of rendered close-ups (cortex, deep nuclei, brainstem junction) before handing over.
  • Budgets: first frame under 2 s, own memory under 400 MB, selection budgets from project 1.

6. Calls made by the builder of this document (not the owner)

  • 2 mm of blur (one voxel left the hand-drawn AAL3 ridges on the cortex; checked on renders; owner accepted it, 2026-10-18), less for small structures, volume within 10%, centre (of the enclosed volume) within 1 mm.
  • Names Midbrain body, Pons body, Medulla body.
  • Highest-probability assignment at 0.5 and dropping the peduncle channel.
  • Whole-brainstem assignment (no cracks), trimming against the cerebellum and at the cord.