๐งฉ SegVol
A universal promptable volumetric segmentation model trained primarily on CT data, combining a 3D image encoder with text and spatial prompts.
At a glanceโ
| Field | Details |
|---|---|
| Model type | Text and spatial prompt 3D segmentation |
| Research team / organization | Beijing Academy of Artificial Intelligence (BAAI), DCAI |
| Input / modality | 3D CT; the upstream paper/release emphasizes more than 200 anatomical categories. |
| Prompt / interaction | Text, point and box prompts; optional zoom-in/zoom-out refinement. |
| Code license | Code: MIT. Verify the checkpoint-specific terms on the model-hosting page. |
๐ Paper and publicationโ
SegVol: Universal and Interactive Volumetric Medical Image Segmentation โ NeurIPS 2024 Spotlight; arXiv:2311.13385.
๐ฆ Checkpoints and accessโ
Official checkpoints and demo data are linked from the BAAI Hugging Face repository and Google Drive; the demo script also expects a local CLIP text-encoder folder.
The code, model checkpoint and any required foundation-model checkpoint can have different terms. Review each linked source before use or redistribution.
โ๏ธ Official setup and inferenceโ
# Follow the repository's legacy-compatible environment (PyTorch >=1.11,
# MONAI 0.9.0, transformers 4.18.0, einops 0.6.1).
pip install 'monai[all]==0.9.0' einops==0.6.1 transformers==4.18.0
# Edit config/config_demo.json and script/inference_demo.sh:
# set CT, ground truth, SegVol checkpoint, CLIP checkpoint and work directory.
bash script/inference_demo.sh
Important: the upstream demo/evaluation script reads ground-truth masks to synthesize point/box prompts. Those are GT-derived/oracle prompts; they are not a human-input inference recipe. See the official inference guide.
โ ๏ธ Version and practical notesโ
The official evaluation demo synthesizes point/box prompts from ground-truth masks. Such runs are GT-derived/oracle; for deployment or human-in-the-loop evaluation, provide prompts independently of the ground truth.
๐ Licenseโ
Code: MIT. Verify the checkpoint-specific terms on the model-hosting page.
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.