A local, AI-assisted case-analysis workbench for radiology research. It imports CT studies, segments them with TotalSegmentator, lets you correct the results with nnInteractive or direct edits, and keeps an exact history of every change. It sits alongside established DICOM viewers and does not replace them.
Research use only. Not a medical device. 3DScan is not cleared or approved for diagnosis, treatment or any other clinical decision. Its AI models are research tools, and their output can be wrong. It comes with no warranty. See LICENSE.
Everything runs on your own machine. Scans are never uploaded to a model host or any other service.
- Import: browser upload,
scan3d import, an optional Orthanc, or any DICOMweb archive. Static MRI is experimental and opt-in. - Segmentation: selected TotalSegmentator tasks on a GPU or the CPU, plus optional MRSegmentator to compare beside them. With a second GPU, choose which one each AI job runs on.
- Corrections: nnInteractive prompts, direct mask edits, scissors and recipes, with exact undo.
- Library: one library for everyone signed in, with search, rename, Trash with undo, and Delete forever. An administrator adds people with a link.
- Viewer: linked slices and 3D volume rendering.
- Measurements: mask volumes and AI-agreement metrics.
- Exchange: DICOM SEG import and export, native OsiriX ROIs, 3D views as new DICOM series, archiving by DICOMweb, direct DICOM sends, 3D Slicer, PNG and portable offline HTML.
On Ubuntu, with git, Node.js 22.12 or newer, and the build tools listed in
INSTALL.md:
git clone https://github.com/Truth129/3DScan.git
cd 3DScan
scripts/bootstrap.sh --torch cpu
.venv/bin/scan3d serveOpen http://127.0.0.1:9010 and choose Explore synthetic demo, or import a scan.
scripts/bootstrap.sh --help lists the install options. Settings live in a git-ignored
scan3d.env. See configuration.
| Read this | For |
|---|---|
| Install | Linux setup, GPUs, AI models |
| Docker | Running it in a container, on the CPU or a GPU |
| Configuration | Every setting in scan3d.env |
| User guide | Importing, reviewing, correcting and exporting |
| Orthanc | Connecting to an Orthanc archive |
| Archives | DICOMweb archives and direct DICOM sends |
| Architecture | How the system works |
| API | HTTP API and data model |
| Development | Tests, checks and project rules |
3DScan runs on Linux, not macOS or Windows. INSTALL.md lists the tested distributions and GPUs.
- Sign-in follows the address. By default 3DScan listens on
127.0.0.1only, with sign-in off, so anything that can reach the port can read every case. Any other address requires an account, and a sign-in needs HTTPS there. See Sign-in. - Your data directory holds patient images. Protect and back it up like any other clinical data. Never put patient data in a bug report, screenshot or pull request.
- Check that your use of patient data is allowed under your own institution's rules.
3DScan is released under the Apache License 2.0. The AI models are separate downloads with their own terms, which you must accept yourself:
- nnInteractive weights: CC BY-NC-SA 4.0, noncommercial. The engine code is Apache-2.0.
- TotalSegmentator: the code is Apache-2.0, but some subtasks need a separate license and are noncommercial. See the provider's license page.
- MRSegmentator: the code is Apache-2.0. Its repository states no separate license for the weights.
Using these models commercially may require permission from their authors. Third-party
notices are in web/public/licenses/. A viewer build adds
licenses/THIRD-PARTY-NOTICES.txt for the packages it bundles, and portable exports carry theirs.