codepawl/ct-anatomy-glb

CT Anatomy GLB

_{Film: scan s1045 assembling itself, rendered in three.js. Music: "Cephalopod" by Kevin MacLeod (incompetech.com), CC BY 4.0. Sounds: heartbeat by FenrirFangs, bones by spookymodem and Kneeling (Freesound, CC0).}

Open the 3D viewer to orbit any body, pull its systems apart and read each structure's name.

Web-ready 3D anatomy from real CT scans. Each body is one GLB with one named node per structure, grouped by body system, coloured by atlas convention, and compressed with meshopt.

For visualization, education and graphics only. Not for diagnosis or any clinical decision.

Built from the segmentation masks of the TotalSegmentator dataset v2.0.1 by Jakob Wasserthal and colleagues, Department of Research and Analysis, University Hospital Basel, released under CC BY 4.0. Please cite their paper: Wasserthal J. et al., "TotalSegmentator: Robust Segmentation of 104 Anatomic Structures in CT Images", Radiology: Artificial Intelligence, 2023, doi:10.1148/ryai.230024.

Use it in 67js

Every body loads in 67js, our WebGL2 renderer, with its glTF addon. Each structure comes in as its own named part, so you can hide, recolour or animate one organ at a time. The files are meshopt compressed, so pass the three.js meshopt decoder.

npm install 67js@npm:@codepawlhq/67js three

import { Renderer67 } from '67js';
import { loadGltf } from '67js/addons/gltf';
import { MeshoptDecoder } from 'three/addons/libs/meshopt_decoder.module.js';

const url = 'https://huggingface.co/datasets/codepawl/ct-anatomy-glb/resolve/main/bodies/s1045.glb';
const body = await loadGltf(url, { decoders: { meshopt: MeshoptDecoder } });
const renderer = new Renderer67(canvas, body.parts, { ground: false, samples: 4 });
renderer.render({ eye: [0.9, 0.62, 2.3], target: [0, 0.5, 0] });

Filter body.parts by name to show one system, for example body.parts.filter((part) => part.name.startsWith('rib_')). The film above was rendered in three.js, and the files work in any glTF viewer.

For games, WebGPU and engines

Each body also comes in lighter LODs and in uncompressed copies for engines that do not read meshopt. Every file keeps the same 115 named nodes, metres, axes and origin, so swapping a LOD is a plain file swap.

Folder	Use	Compression
bodies/	web, full detail	meshopt
lods/<scan>_lod1, lod2, lod3	web games, LOD swaps	meshopt
engine/<scan>.glb and engine/<scan>_lod1, lod2, lod3	Unreal, Blender, Godot, Unity	none

Scan s1045 at each level:

Level	Triangles	Web size	Engine size
full	361,040	1.80 MB	8.75 MB
lod1	180,602	1.01 MB	3.34 MB
lod2	72,134	0.51 MB	1.39 MB
lod3	21,552	0.27 MB	0.48 MB

Tested: three.js WebGPURenderer drew bodies/, lods/ and engine/ files of scan s1045 on the WebGPU backend with all 115 meshes. Blender 5.2 imported engine/s1045.glb and its LOD3 with all 115 meshes. Unreal Engine 5.8 imported engine/s1045.glb and its LOD2 as 115 static meshes each. Unreal does not read meshopt, so use engine/ there.

import { WebGPURenderer } from 'three/webgpu';
import { GLTFLoader } from 'three/addons/loaders/GLTFLoader.js';
import { MeshoptDecoder } from 'three/addons/libs/meshopt_decoder.module.js';

const loader = new GLTFLoader().setMeshoptDecoder(MeshoptDecoder);
const base = 'https://huggingface.co/datasets/codepawl/ct-anatomy-glb/resolve/main/';
const body = await loader.loadAsync(base + 'lods/s1045_lod2.glb');
scene.add(body.scene);

What is here

• bodies/: 7 bodies, 2,515,782 triangles in all (from 20,238,966 raw marching cubes triangles).
• organs/: 20 single organs or organ groups (aorta_and_veins, brain, digestive_tract, heart, kidneys, liver, lungs, pelvis, ribcage, spine) from scans s0690, s1045, at higher detail.
• lods/ and engine/: lighter and uncompressed variants of every body, listed in lods/variants.jsonl.
• thumbs/: a 512 px WebP per file, with thumbs/metadata.jsonl for the Dataset Viewer above.
• assets/: the banner, preview film, 67js render and gallery shown on this page.
• metadata.jsonl: one row per file with scan id, age and sex from the dataset's metadata, structure names, per-structure volume in millilitres, triangle count, file size and coverage.
• Size of bodies/ and organs/: 20.07 MB.

File	Age	Sex	Coverage	Structures	Triangles	Size
bodies/s1045.glb	62	male	skull to upper thigh	115	361,040	1.80 MB
bodies/s0790.glb	40	female	skull to hip, lower pelvis cut by the scan edge	114	358,714	1.78 MB
bodies/s1159.glb	47	female	skull to hip, lower pelvis cut by the scan edge	114	357,106	1.78 MB
bodies/s1369.glb	73	female	skull to hip, arms raised above the head, lower pelvis cut by the scan edge	113	364,708	1.81 MB
bodies/s0690.glb	94	female	skull to upper thigh	113	356,520	1.78 MB
bodies/s0086.glb	71	male	neck to upper thigh, no head in the scan	109	360,524	1.79 MB
bodies/s0050.glb	31	male	neck to upper abdomen and pelvis, head and arms outside the scan	109	357,170	1.77 MB

How it was made

1. Only the segmentation masks were fetched, with HTTP range requests into the 23.58 GB source zip. No CT image was downloaded, so no skin or face surface can be built from this pack. The dataset has no skin class.
2. Per mask: specks dropped, thin bone closed and its hollow cortex filled, Gaussian blur (sigma 1.0 voxel, 0.62 for thin structures), marching cubes at 1.5 mm voxels, 24 Taubin smoothing steps, quadric decimation with a face budget shared by surface area.
3. GLB written with one node per structure under one node per system, then meshopt compressed with glTF Transform.

Axes: metres, +Y superior, +Z anterior, +X the patient's left (the patient faces +Z). Each body file passed a side check: left kidney, spleen, left lung, left femur and left humerus all sit at higher X than their right-side partner or the liver. Node extras carry label, system and volume_ml.

Limits

• Coverage is what each clinical scan covered. No scan has forearms, hands or legs below the thigh, and some bodies are cut at the pelvis or have the arms raised. Each row's coverage says which.
• Shapes are smoothed segmentations of real patients, not idealised atlas models. Small structures (adrenal glands, thin vessels) are simplified, and segmentation errors in the source carry over.
• Colours are illustrative atlas colours, not measured tissue colour.
• Scans come from clinical routine at University Hospital Basel and were de-identified by the dataset authors.

Licence

CC BY 4.0, same as the source. Credit: TotalSegmentator dataset, Wasserthal et al., University Hospital Basel. Meshes and packaging by CodePawl.