Open-source toolkits for 3D medical imaging — segmentation, foundation-model pretraining, and the automation layered on top of them.
Everything in this org runs end-to-end — raw data in, trained models and evaluated predictions out — with sensible defaults that work out of the box and a config file for when you need more control. Every tool here runs on a laptop CPU as well as a multi-GPU cluster (NVIDIA or AMD ROCm); nothing requires a specific vendor's hardware to get started.
| Project | What it does |
|---|---|
| MIST | End-to-end 3D medical image segmentation: analyze → preprocess → train → evaluate → predict → postprocess. One config file switches between nnU-Net, MedNeXt, Swin UNETR, and other architectures. |
| MISFIT | Pretrains 3D medical imaging foundation models with masked autoencoders. Point it at unlabeled NIfTI volumes and get back an encoder that transfers to segmentation, classification, and retrieval. |
| skills | Claude Code Agent Skills that give an LLM working knowledge of MIST, MISFIT, and mist-autoresearch, so it can help configure and debug pipelines without re-reading the source every time. |
# Segmentation
pip install mist-medical
# Foundation-model pretraining
pip install misfit-medicalTry MIST first with the end-to-end Colab demo — no install, no GPU required.
Issues and PRs are welcome on any repo. Each project has its own
CONTRIBUTING.md; start there for dev setup and test/lint conventions.
All projects are released under Apache 2.0.