A promptable foundation model adapted from SAM 2 for 3D medical images and medical videos, using memory-based propagation across slices/frames.
At a glance
| Field | Details |
|---|---|
| Model type | 3D and video promptable segmentation |
| Research team / organization | University Health Network, University of Toronto and collaborating institutions |
| Input / modality | 3D CT/MRI/PET volumes and medical videos such as ultrasound or endoscopy. |
| Prompt / interaction | Spatial prompts such as boxes/masks on a slice or frame, followed by propagation and interactive correction. |
| Code license | Code repository: Apache-2.0. Check each Hugging Face checkpoint page and the original SAM 2 terms before redistribution or deployment. |
📚 Paper and publication
MedSAM2: Segment Anything in 3D Medical Images and Videos — arXiv:2504.03600 (2025 preprint).
📦 Checkpoints and access
The official Hugging Face collection includes MedSAM2_latest.pt, MedSAM2_2411.pt and task-specific CT lesion, MRI liver lesion and ultrasound heart checkpoints. bash download.sh downloads all listed weights.
The code, model checkpoint and any required foundation-model checkpoint can have different terms. Review each linked source before use or redistribution.
⚙️ Official setup and inference
1 | conda create -n medsam2 python=3.12 -y |
That CLI example is CT-specific; follow the official MRI/video notebooks or scripts for those modalities.
⚠️ Version and practical notes
The upstream repository provides general and task-specific checkpoints. Record the exact checkpoint, training data, modality and prompt configuration; the latest release is not automatically identical to earlier published weights.
📜 License
Code repository: Apache-2.0. Check each Hugging Face checkpoint page and the original SAM 2 terms before redistribution or deployment.
Citation: Cite the paper above and the upstream project as directed by its authors. This repository links to the original paper; it does not redistribute the code or weights.
Source pages checked: 27 October 2026. Upstream environments, weights and access policies may change; use the linked project as the source of truth.