๐งฌ BiomedParse v2
The v2 release extends BiomedParse to end-to-end 3D segmentation and adds built-in object-existence detection. It is distinct from the original 2D BiomedParse v1.
At a glanceโ
| Field | Details |
|---|---|
| Model type | Text-guided biomedical parsing model |
| Research team / organization | Microsoft Research |
| Input / modality | 3D CT, MRI, ultrasound, PET and 3D microscopy according to the v2 README; dedicated pretrained v2 weights are for 3D. |
| Prompt / interaction | Text-concept prompts; slice-wise model inference with neighboring volume context encoded into RGB input. |
| Code license | Code: Apache-2.0. The Hugging Face model card lists CC BY-NC-SA 4.0 for model artifacts and requires acceptance of the gated-model conditions. Do not conflate v1 and v2 model terms. |
๐ Paper and publicationโ
A Foundation Model for Joint Segmentation, Detection, and Recognition of Biomedical Objects Across Nine Modalities โ Nature Methods 22, 166โ176 (2025), DOI: 10.1038/s41592-024-02499-w. The v2 3D checkpoint also builds on the CVPR 2025 Text-guided 3D Segmentation Challenge release.
Version pin: this card follows the upstream v2 branch.
๐ฆ Checkpoints and accessโ
biomedparse_v2.ckpt (~4.46 GB) is hosted in the gated Hugging Face repository microsoft/BiomedParse; access may require accepting model conditions and authentication. The upstream v2 README also provides a challenge-trained Docker image.
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
git clone --branch v2 https://github.com/microsoft/BiomedParse.git
cd BiomedParse
conda create -n biomedparse_v2 python=3.10.14
conda activate biomedparse_v2
python -m pip install -r assets/requirements/inference.txt
# Download biomedparse_v2.ckpt after completing HF access requirements:
python inference.py --input_dir examples/imgs --output_dir outputs/3d \
--checkpoint /path/to/biomedparse_v2.ckpt --slice_batch_size 4
The script expects challenge-format preprocessed NPZ volumes. Raw DICOM/NIfTI still needs modality-appropriate intensity preprocessing; reduce --slice_batch_size if GPU memory is limited.
โ ๏ธ Version and practical notesโ
Use the v2 branch and biomedparse_v2.ckpt for the 3D workflow. The v1 branch and weights are a separate 2D workflow; raw scanner images require modality-appropriate preprocessing.
๐ Licenseโ
Code: Apache-2.0. The Hugging Face model card lists CC BY-NC-SA 4.0 for model artifacts and requires acceptance of the gated-model conditions. Do not conflate v1 and v2 model 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.