A 3D thoracic operative atlas linked to a scrollable CT. Built for VATS left upper lobectomy (anterior and posterior approaches); the same data model takes other operations.
- CT and 3D in one frame (RAS mm, carina at the origin). Scroll or click the CT and the 3D follows; click in 3D and the CT jumps there. Labelled structures are outlined on the CT.
- Procedure mode: each step moves the camera to the surgeon's view, highlights the structure being dissected (teal) and what to protect (red), asks the next move before revealing it, and staples at a computed staple line.
- Your CT: drop an uncompressed DICOM series (or .nii / .nii.gz). It is read in the browser and never uploaded. Click the carina once to place it against the 3D reference (translation only).
Needs Python 3.10–3.12, Node 20+, and uv (or plain pip). 16 GB RAM, or --lowmem; an NVIDIA GPU makes step 2 minutes instead of ~45 min.
npm install # app + gltf-transform (mesh compression)
cd pipeline && uv sync --extra segment && cd .. # Python deps incl. TotalSegmentator
uv run --project pipeline python pipeline/segment.py --input /path/to/dicom-folder --device gpu # add --lowmem on 8 GB
uv run --project pipeline python pipeline/build.py # names hilar branches, meshes, writes public/data/
uv run --project pipeline python pipeline/procedures.py # the two approaches, cameras from the computed anatomy
npm run dev # http://localhost:5174
npm run build # static site in dist/, host anywhere
pipeline/hilum.py explains how branches are named (skeleton trees, lobe territory, order along the trunk). Nerves,
the ligamentum arteriosum and node stations are schematic, placed on landmarks, and labelled as such.
Not for clinical use. One reference CT (3D Slicer's CTA-cardio sample, no stated licence: teaching use only). Branch names come from this scan's geometry and need a surgeon's review.