Access the Arterial model weights
These weights are released under CC BY-NC 4.0 for noncommercial research use. They are research software, not an approved medical device.
By requesting access you confirm that you have read and accept the Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) licence, and that you understand these models are provided for research purposes only. They have not been cleared or approved by any regulatory body and must not be used to inform clinical decisions about individual patients.
Log in or Sign Up to review the conditions and access this model content.
Arterial — model weights
Trained model weights for Arterial, an AI framework for automated vascular analysis of CT Angiography (CTA) in the supra-aortic region, developed to support mechanical thrombectomy planning in acute ischaemic stroke.
This repository contains only the weights. The code, installation instructions and documentation live in the GitHub repository.
What is in this repository
| Directory | Task | Architecture |
|---|---|---|
segmentation/ |
Extracranial, intracranial and mandible segmentation from CTA | nnU-Net v2 |
landmark_detection/ |
Anatomical landmark detection at vascular bifurcations | 3D U-Net |
vessel_labelling/ |
Anatomical naming of centerline segments (e.g. LCCA) | Graph neural network |
access_prediction/ |
Catheter accessibility prediction from pathway graphs | ArterialGNet (multi-scale GNN) |
The segmentation weights account for essentially all of the ~1.2 GB; the graph models are a few MB each.
Usage
You must accept the licence above before any download will work. Access is granted automatically the moment you accept — there is no waiting period and no manual approval.
Recommended: the Hugging Face CLI
1. Install the client
pip install huggingface_hub
2. Accept the licence — click Agree and access repository at the top of this page.
3. Log in
Create an access token with the read role at huggingface.co/settings/tokens, then:
hf auth login
Paste the token when prompted. This is a one-off step per machine.
4. Download
If you have the Arterial repository checked out, use the bundled script — it places the weights where Arterial expects them and verifies the result:
bash scripts/download_models.sh
Otherwise, fetch them directly:
hf download FLOWCAT-CV/arterial-models --local-dir arterial-models
or from Python:
from huggingface_hub import snapshot_download
snapshot_download("FLOWCAT-CV/arterial-models", local_dir="arterial-models")
Telling Arterial where the weights are
By default Arterial looks in $arterial_dir/models. To keep the weights anywhere else, set
ARTERIAL_MODELS_DIR, which always takes precedence:
export ARTERIAL_MODELS_DIR="/data/arterial-models"
Offline and air-gapped machines
Clinical environments frequently have no outbound network access. Download on a connected machine,
copy the directory across, and point ARTERIAL_MODELS_DIR at it:
# on a connected machine
hf download FLOWCAT-CV/arterial-models --local-dir arterial-models
tar czf arterial-models.tar.gz arterial-models
# on the target machine
tar xzf arterial-models.tar.gz -C /data
export ARTERIAL_MODELS_DIR=/data/arterial-models
Layout
access_prediction/ dataset.json, fold_{0..4}/model_weights.pth
landmark_detection/ six_landmarks_2ch.pth, six_landmarks_11_7.pth
segmentation/ extracranial_vessels/, intracranial_vessels/,
totalsegmentator_mandible/
vessel_labelling/ extracranial_vessels/
The segmentation/ subdirectories follow the nnU-Net v2 trained-model folder convention
(plans.json, dataset.json and fold_*/checkpoint_final.pth) and are passed directly to
nnUNetPredictor.initialize_from_trained_model_folder.
Full installation instructions, including the rest of the Arterial framework, are in the repository README.
Intended use
These models are intended for research on vascular morphology and endovascular treatment planning, by researchers working with CTA imaging of the supra-aortic region. They are the weights used by the Arterial framework to produce segmentations, labelled centerlines, geometric features and accessibility predictions.
Out of scope
- Clinical decision-making for individual patients. These are research models. They have no regulatory clearance (CE, FDA or otherwise) and must not be used to guide patient care.
- Commercial use of any kind, which the CC BY-NC 4.0 licence prohibits. Contact the authors if you need commercial terms.
- Imaging modalities other than CTA, and anatomy outside the supra-aortic region.
Limitations
- Trained on CTA acquired in the context of acute ischaemic stroke workup. Performance on other populations, indications or acquisition protocols is unknown.
- Sensitive to acquisition parameters, contrast timing and severe artefact. Poor contrast opacification degrades segmentation, and downstream centerline, labelling and accessibility outputs inherit those errors.
- The vessel labelling and access prediction models operate on graphs derived from the segmentation step, so segmentation failures propagate through the pipeline.
- Not prospectively validated as a decision-support tool.
License
Released under CC BY-NC 4.0 — noncommercial use only, with attribution.
Copyright 2022-2026 Stroke Research at Vall d'Hebron Research Institute (VHIR), Barcelona, Spain.
The same terms apply to the Arterial source code. Arterial's dependencies carry their own licences (nnU-Net and MONAI under Apache-2.0, VMTK under BSD, PyTorch Geometric under MIT), which you must comply with independently.
Citation
If you use these models, please cite:
@article{canals2023vascular,
title={A fully automatic method for vascular tortuosity feature extraction in the supra-aortic region: Unraveling possibilities in stroke treatment planning},
author={P. Canals, S. Balocco, O. Díaz, J. Li, A. García-Tornel, A. Tomasello, M. Olivé-Gadea, M. Ribo},
journal={Computerized Medical Imaging and Graphics},
volume={104},
pages={102170},
year={2023},
publisher={Elsevier},
doi={10.1016/j.compmedimag.2022.102170}
}
🔗 https://www.sciencedirect.com/science/article/pii/S0895611122001409
Acknowledgments
Developed by the FlowCAT lab, a division of the Stroke Research group at the Vall d'Hebron Research Institute (VHIR), Barcelona, Spain.
Built on nnU-Net, VMTK, PyTorch Geometric and MONAI. The mandible segmentation model derives from TotalSegmentator.