A 3D promptable segmentation model adapted from SAM for volumetric medical images. Its original workflow uses iterative foreground/background point prompts and returns a volumetric mask.
At a glance
| Field | Details |
|---|---|
| Model type | 3D point-prompt segmentation model |
| Research team / organization | Shanghai Jiao Tong University; Shanghai Artificial Intelligence Laboratory; Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences. |
| Input / modality | 3D medical volumes; the project benchmarks CT, MRI and other volumetric modalities. |
| Prompt / interaction | Interactive 3D point prompts; iterative corrections. |
| Code license | Code: Apache-2.0. Check the checkpoint page for the weight-specific terms. |
📚 Paper and publication
SAM-Med3D: Towards General-purpose Segmentation Models for Volumetric Medical Images — arXiv:2310.15161 (2023); the project README reports acceptance as an ECCV BIC 2024 Oral.
📦 Checkpoints and access
The upstream project currently highlights the newer SAM-Med3D-turbo checkpoint. Its README links Hugging Face, Google Drive and HF-Mirror. The Hugging Face model blueyo0/SAM-Med3D exposes sam_med3d_turbo.pth and is referenced by the upstream MedIM example.
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 | git clone https://github.com/uni-medical/SAM-Med3D.git |
The demo is an evaluation example and expects a paired ground-truth mask. For model construction and checkpoint selection, see the official MedIM example.
⚠️ Version and practical notes
The project currently highlights the newer SAM-Med3D-turbo checkpoint. Keep it distinct from the original paper checkpoint, and record the exact weight and prompt protocol when reproducing published results.
📜 License
Code: Apache-2.0. Check the checkpoint page for the weight-specific terms.
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.