A medical adaptation of SAM 3 with separate 2D benchmark tooling and a newer 3D v2 training/evaluation pipeline. The current 3D path runs text-prompted inference over full volumes in independent four-slice clips.
At a glance
| Field | Details |
|---|---|
| Model type | Medical SAM3 2D/3D adaptation |
| Research team / organization | The paper lists affiliations at University of Central Florida, University College London, University of Illinois Urbana-Champaign, University of Trento, The Hong Kong Polytechnic University and University of Pennsylvania. |
| Input / modality | 2D medical benchmarks and 3D NPZ volumes; check which code path/version a result uses. |
| Prompt / interaction | 2D box/text prompts; current 3D v2 inference is text-prompted and full-depth. |
| Code license | The repository includes Meta’s SAM License for SAM materials. Check the current Hugging Face model card for the fine-tuned checkpoint’s terms. |
📚 Paper and publication
Medical SAM3: A Foundation Model for Universal Prompt-Driven Medical Image Segmentation — arXiv:2601.10880 (2026 preprint).
📦 Checkpoints and access
The project links model checkpoints on Hugging Face, including the June 2026 3D v2 weights. Select the 2D or 3D asset explicitly; the 3D CLI expects a local medical_sam3_3d.pt checkpoint.
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/AIM-Research-Lab/Medical-SAM3.git |
The 2D evaluation path is separate: cd inference && python run_medsam3_evaluation.py --checkpoint ... --model-name medsam3. Outputs from the 3D path are NPZ files containing segs with original (D,H,W) geometry.
⚠️ Version and practical notes
The repository has separate 2D and 3D workflows. For 3D v2, the upstream README says inference covers the full depth without ground-truth prompts or ground-truth-selected slices.
📜 License
The repository includes Meta’s SAM License for SAM materials. Check the current Hugging Face model card for the fine-tuned checkpoint’s 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.