3Dentes

Interactive 3D atlas of human oral anatomy, in the browser — segmented from a real cone-beam CT, and it takes its own radiograph. One codebase runs on Linux, macOS, Windows and iPad — install it to the home screen and it works offline.

Live: https://natesaindon.github.io/3Dentes/ · What changed: CHANGELOG.md (currently v0.9.2)

The anatomy is segmented from one cone-beam CT at 0.16 mm isotropic — hard tissue, pulp, periodontal ligament, gingiva and the nerve supply, all of one person. Fading the teeth back is what the layer opacity is for.

What it does

• 28 individually selectable permanent teeth, plus the mandible, the maxillary alveolar process and palate, and upper and lower gingiva.
• Internal anatomy: the pulp chamber and canal system of every tooth, the periodontal ligament space, and the nerve supply — the inferior alveolar nerve in its measured canal, a branch to all 14 lower apices, the mental and incisive terminal branches, and the superior dental plexus with PSA, MSA, ASA and the infraorbital nerve above.
• Clinical notation — every tooth labelled in Universal, FDI and Palmer, all derived from arch/side/position rather than hand-entered.
• Odontogram in the standard clinical layout (upper arch on top, patient's right on the viewer's left), synced both ways with the 3D view. Selecting a tooth on the chart flies the camera to it, approaching from outside the arch so the tooth is actually exposed — near-frontal for an incisor, swinging buccal for a molar. This is how you reach the posterior teeth, since from a frontal view a molar really is hidden behind the premolars in front of it.
• Radiograph mode — a simulated periapical of the selected tooth, computed by ray-integrating the atlas's own meshes against published tissue densities. Not a picture of the scan: the receptor and the beam are placed in the mouth in the 3D view, the sensor is a real size-2 sensor (36 × 25.6 mm active in a 43 × 31.2 × 7.5 mm housing) that cannot pass through bone or teeth, and paralleling versus bisecting, foreshortening, elongation and cone cut all fall out of the geometry rather than being drawn on. The film reports what the tooth measures on it against its true length.
• The bone on that film carries this patient's own trabecular pattern, measured from the scan rather than drawn on — the mandible is no longer a flat slab. That it is trabecular structure and not scanner noise is established against soft-tissue and air controls, and its amplitude is fitted against a real periapical of the same three teeth. The pattern is measured; the mapping from an uncalibrated grey value to attenuation is derived, and the atlas says so.
• Every tooth opens at a placement chosen for it — the sensor clear of bone, the whole tooth on the active area, and the least angulation that achieves both. The mouth opens only as far as that exposure needs. Where no placement works, it says so instead of pretending.
• Layer visibility and opacity, so you can fade the gingiva and alveolar bone back to expose the roots in situ, cut the teeth to 15% to see the pulp inside them, or isolate a single structure.

Radiograph mode

The sensor and the beam are placed in the mouth in the 3D view, and the film beside them is computed from the same geometry. Selecting any part of a tooth — including its enamel or its crown — aims the tube at it.

Nothing about the following is drawn on. The projection is built from a source point and the receptor rectangle, so the errors of the technique are consequences of where the tube and the film were put — the film even reports what the tooth measures on it against its true length.

Where the geometry comes from, and how far to trust it

Different parts of this model stand on very different evidence, and the atlas is built to keep them apart rather than render them alike:

Measured	Teeth, mandible, maxilla, the mandibular canal, and the pulp — the pulp traced by hand on the slices, because no threshold separates it from dentin at this resolution.
Derived	Gingiva, lofted from the measured cementoenamel junction and refitted to the published cervical-line curvature. Enamel, a depth-limited shell whose extent is measured and whose thickness is literature. The periodontal ligament and the lamina dura, both measured in position but drawn thicker than the truth — the ligament is barely one voxel, and the socket wall is drawn continuous rather than only where cortex was resolved.
Schematic	The two zirconia crowns: their outline is measured, but this scan saturates and cannot measure their density, which is the material's from the literature. The nerve inside the mandibular canal — CBCT resolves the canal, not its contents. The mental and incisive branches. Every maxillary nerve: the superior alveolar canals are thin, often dehiscent, and not reliably visible at 0.16 mm.

The app says this per structure, not just here. Select anything and the detail panel states its tier — measured, derived or schematic — the method that produced it, and a source where the shape came from the literature rather than the scan. That is a build-enforced field on every structure rather than a note in a README, because the question "how do you know?" gets asked about the thing you just clicked, not about the project.

What it is not

• No third molars — 28 teeth, not 32.
• One person's anatomy. Root count, canal configuration and curvature are this individual's, not a textbook composite.
• Resolution limits are real. Canals under about 0.5 mm and apical deltas fall below the voxel size and are not drawn. Apical deltas occur in roughly 10% of teeth; drawing one at this scale would render it several times its true calibre.
• Two teeth carry metal restorations whose beam-hardening haloes threshold exactly like a pulp lumen. Their pulp is partly inference and is flagged.
• The simulated radiograph is computed, not photographed. Bone carries no trabecular pattern, because the model has no trabecular geometry, and the crowns clip to flat white, because the beam is modelled at a single effective energy. Both are stated on screen.
• No palatal or sublingual mucosa, so sensor placement is tested against a small allowance standing in for the tissue a real film rests on.
• No tongue, TMJ disc, or salivary glands.

This is a study and visualisation aid, not a complete clinical reference. The app states this on screen; please keep it that way in any fork.

Still on the list — including a caries slider, an age-calcification slider, and arteries and veins — is in docs/wishlist.md. The pathology feature has its own tracker in docs/pathology-wishlist.md: none of it is built, and everything in it would be authored disease on a healthy dentition, labelled as such.

Running it

npm install
npm run build:assets   # -> public/dentition.glb + teeth.json
npm run dev            # http://localhost:5173/3Dentes/

The STLs are vendored in assets/cbct/stl/, so the build works offline. Earlier versions had a second, BodyParts3D-derived asset tree selected by a TOOTH_SOURCE environment variable; that tree has been removed and there is one source now.

The CBCT pipeline itself is not part of npm — it is the Python in tools/cbct/, and it needs the imaging data, which is not in this repo.

To reach the dev server from an iPad on the same network, npm run dev -- --host.

How it fits together

tools/manifest.mjs      which structures to use, their layer, anatomical side,
                        and the notation derivation (single source of truth)
tools/build-assets.mjs  STL -> welded, smooth-shaded, y-up, centred .glb
tools/cbct/             the Python that produced assets/cbct/ from the scan:
                        segmentation, arch split, pulp tracing, gingiva, nerves
src/scene.js            renderer, lighting, camera, camera flights
src/picking.js          raycast selection, drag-vs-click, see-through handling
src/odontogram.js       the tooth chart
src/ui.js               layer, notation and detail panels
src/main.js             wiring and selection state

CLAUDE.md carries the long-form record of this pipeline — every rule that cost something to learn, and the things already ruled out. Read it before changing anything under tools/cbct/.

FMA identifiers are the join key throughout: a structure is FMA55697 everywhere — the source filename, the glTF node name, and the key in teeth.json.

The build refuses to produce a mislabelled model. Two assertions run on every build. The first is laterality: anatomical right is negative x, and a structure labelled left that sits on the right fails the build. The second checks the tooth numbering against the tooth shapes — Universal, FDI and Palmer are all derived from one arch/side/position triple, so they always agree with each other, and agreeing with each other is not the same as being right. So the teeth are ordered around each arch geometrically and that order must match the numbering, and molars and canines are identified from volume, root count and length and must land on the positions claiming them. node tools/tooth-morphology.test.mjs corrupts the mapping in six plausible ways and confirms each is caught.

Notes from building it

A few things that were not obvious and are easy to regress:

• Binary STL is a triangle soup with no shared vertices and no vertex normals. Converting it naively gives ~1M vertices and hard faceted shading. Welding bitwise-identical vertices recovers the original topology exactly (1,043,478 → 173,956 vertices, 83% fewer) because these STLs came from an OBJ whose shared vertices have identical float bits. Welding with a tolerance would round off cusp tips and occlusal fissures, so it is deliberately exact.
• @gltf-transform's normals() generates flat normals, not smooth ones, hence the hand-rolled area-weighted computeSmoothNormals.
• Indices dominate the file. At 32 bits they were larger than positions and normals combined; every structure welds to well under 65,536 vertices, so 16-bit indices cut the build from 8.4MB to 6.3MB for free.
• The build fails if the model is mirrored. Anatomical right is negative x in this dataset; checkLaterality asserts every structure labelled left or right sits on the expected side. An atlas that confidently mislabels a side is worse than no atlas.
• See-through structures don't absorb clicks. With gingiva at its default 45%, the teeth behind it are plainly visible, so a click there should select the tooth. Above 60% opacity a structure reads as solid and does take the click.

Licensing

Two things with two different standings:

• Code (src/, tools/) — MIT. See LICENSE.
• The anatomy (assets/cbct/, and the dentition.glb built from it) — Nate Saindon's own scan, published with his explicit consent, and licensed CC BY-NC 4.0. Use, adapt and redistribute it non-commercially, including in teaching, with attribution: "Anatomy from 3Dentes by Nate Saindon, CC BY-NC 4.0." Commercial use is not granted here — ask.

The atlas previously vendored BodyParts3D meshes under CC BY-SA 2.1 Japan. All of them have been replaced by measured CBCT anatomy and removed from the repository, so no ShareAlike obligation reaches any part of this project. Full detail, including what that inheritance used to cover, is in ATTRIBUTION.md.

This repo contains one individual's medical imaging derivatives, with consent. It contains no DICOM, no raw volumes, and no third-party patient data — and it never should.