VoxTell v1.1

English简体中文

A free-text promptable 3D vision-language segmentation model that maps natural-language descriptions, from short labels to clinical phrases, to volumetric masks.

Official code Paper Weights

← Model atlas

At a glance

Field Details
Model type Free-text 3D segmentation model
Research team / organization German Cancer Research Center (DKFZ) / Helmholtz Imaging
Input / modality 3D CT, MRI and PET volumes; current v1.1 weights are distributed by the authors.
Prompt / interaction Free-text prompt only; no point/box interaction is required by the standard predictor.
Code license Code: Apache-2.0. Check the Hugging Face repository for model-weight terms.

📚 Paper and publication

VoxTell: Free-Text Promptable Universal 3D Medical Image Segmentation — IEEE/CVF CVPR 2026, pp. 37538–37557; arXiv:2511.11450.

📦 Checkpoints and access

Official Hugging Face repository: mrokuss/VoxTell; v1.1 is downloaded automatically on first use and cached. The model includes the frozen Qwen3-Embedding-4B text encoder.

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
2
3
4
5
conda create -n voxtell python=3.12
conda activate voxtell
# Install PyTorch matching the available CUDA driver, then:
pip install voxtell
voxtell-predict -i case.nii.gz -o output -p "liver" "spleen"

Use NIfTI input and preserve orientation metadata. VoxTell expects RAS orientation, does not automatically resample unusual voxel spacing and recommends checking orientation/spacing when results look anatomically misplaced.

⚠️ Version and practical notes

This is a text-only volumetric model. Preserve the exact English prompt, image orientation, voxel spacing and model tag; prompt wording and preprocessing can affect predictions.

📜 License

Code: Apache-2.0. Check the Hugging Face repository for model-weight 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.