Welcome to GOUHFI 2.0, a novel deep learning (DL) toolbox for subcortical segmentation, cortical parcellation and volumetry of brain MR images of any contrast, resolution and field strength.
This README provides a project overview with links to the full Installation and Usage documentation.
Helper commands for neuroscientists/neuroimagers working with label maps are also documented in docs/usage.md.
- 09/05/26: GOUHFI 2.0.1: Docker support added!
- In order to make GOUHFI more accessible to the community, we made a Docker implementation for GOUHFI. That way, GOUHFI can now be run without any Python setup using only a few installation steps. See
docs/installation_docker.mdfor the full guide.
- In order to make GOUHFI more accessible to the community, we made a Docker implementation for GOUHFI. That way, GOUHFI can now be run without any Python setup using only a few installation steps. See
- 13/12/25: GOUHFI 2.0 is finally available!
- In addition of an improved subcortical segmentation tool on clinical cohorts, GOUHFI 2.0 now also offers (a) contrast-, resolution- and field-agnostic cortex parcellation following the Desikan-Killiany-Tourville (DKT) atlas and (b) Total Intracranial Volume (TIV) estimations for normalized volumetry analyses.
- The related arXiv publication for GOUHFI 2.0 is available here.
- 29/09/25: 🎉 GOUHFI's paper has been accepted for publication in the Imaging Neuroscience journal! 🎉 The accepted version is available online here.
GOUHFI is a fully automatic, contrast- and resolution-agnostic, DL-based brain segmentation tool optimized for Ultra-High Field MRI (UHF-MRI), while also demonstrating strong performance at 1.5T/3T compared to other well-established techniques.
Using the domain randomization approach proposed in SynthSeg and a state-of-the-art 3D U-Net with Residual Encoders from the nnUNetv2 framework, GOUHFI is able to handle various contrasts, resolutions and even field strengths without requiring fine-tuning or retraining. Tested on multiple datasets, it showed high accuracy and impressive robustness to noise and inhomogeneities, making it a valuable tool for neuroscientists working at both 3T and UHF-MRI.
For more details on how GOUHFI was developed, please refer to the corresponding paper published in Imaging Neuroscience.
More recently, we released a major update under the name GOUHFI 2.0. This new version retains the original subcortical segmentation approach, based on the same domain randomization principles, but benefits from an improved and expanded training dataset. In addition, GOUHFI 2.0 introduces cortical parcellation and Total Intracranial Volume (TIV) estimation, enabling independent normalized volumetric analyses. Further details about the development of GOUHFI 2.0 are provided in the corresponding ArXiv publication.
Two installation methods are available:
| Method | Best for | Guide |
|---|---|---|
| Docker | Most users who just want to segment their images. Easiest, no Python knowledge required. | docs/installation_docker.md |
| Native (conda + pip) | Developers or users who want to modify the code. | docs/installation_native.md |
Both methods require the user to download the trained model weights from Zenodo or Hugging Face (~7 GB each, 3 models in total).
GOUHFI is operated through command-line tools (run_gouhfi, run_preprocessing, run_renaming, etc.).
The full command reference and detailed examples are in docs/usage.md.
Native installation:
conda activate gouhfi
run_preprocessing -i /path/to/raw_images -o /path/to/preprocessed
run_renaming -i /path/to/preprocessed -o /path/to/renamed
run_gouhfi -i /path/to/renamedDocker (GPU):
docker run \
# ── Docker configuration ──────────────────────────────────────────
--gpus all \ # pass NVIDIA GPU into container
--rm \ # delete container when done
--shm-size=16g \ # shared memory for nnUNet workers
-e HOST_UID=$(id -u) \ # output files will be owned by you
-e HOST_GID=$(id -g) \ # output files will be owned by you
-v /path/to/weights:/opt/gouhfi/trained_model:ro \ # model weights (read-only)
-v antspynet-cache:/opt/keras-cache \ # ANTsPyNet weight cache (persistent)
-v /path/to/input:/input \ # your input NIfTI files
-v /path/to/output:/output \ # where segmentations will be saved
# ── Image and command ─────────────────────────────────────────────
mafortin30/gouhfi:2.0.1 \
run_gouhfi -i /input -o /outputThis project incorporates code from the following projects, used under the Apache License 2.0:
Image preparation/preprocessing:
- ANTsPyNet:
- For brain extraction. Quick and efficient brain extraction tool (
antspynet.brain_extraction) if you need to do this for your images to be segmented. - We provide a script called
brain_extraction_antspynet.pywhere we wrapped an unmodified implementation ofantspynet.brain_extractionto make the repository more self-contained. - If you have an already up and running ANTsPyNet installation, you can use it directly from there. In this repository, the function
run_brain_extractionwill execute this script.
- For brain extraction. Quick and efficient brain extraction tool (
Training:
- nnU-Net v2:
- The nnUNet v2 framework was used for training, inference, post-processing and evaluation of GOUHFI.
- This repository contains the full nnUNetv2 directory (version v2.4.1).
- If you would like to reproduce the full training pipeline as explained in the GOUHFI paper (or retrain a model from scratch), you should be able to do so with GOUHFI's installation alone.
- However, we recommend the users to refer to the nnUNet documentation for more information on how to proceed since the documentation is not included in this repository.
Generating synthetic images for training:
- SynthSeg:
- The synthetic images used to train GOUHFI were generated from the generative model proposed in SynthSeg.
- Since SynthSeg is a complex beast on its own, we have decided to not include it in this repository.
- However, the approach used to create the synthetic training data for GOUHFI is very similar to the one shown in the 2-generation_explained.py tutorial script available in the SynthSeg GitHub repository. Thus, we recommend to people interested in reproducing the full pipeline with the synthetic image generation process to install SynthSeg on its own and follow their well designed tutorials.
- Basically, by
- Swapping the original labels_classes_priors files from SynthSeg in the 2-generation_explained.py file by the ones shared in the /misc/ subdirectory here in this repository (the four
.npyfiles), - setting the variable
n_neutral_labelsto 6 andrandomise_resto False, and - using the model parameters described in the appendices of the paper (under submission) for the generative model,
- You can create synthetic images for label maps containing the 'Extra-Cerebral' label (see run_add_label for how to perform this).
- Swapping the original labels_classes_priors files from SynthSeg in the 2-generation_explained.py file by the ones shared in the /misc/ subdirectory here in this repository (the four
- Basically, by
- More details about the generative model can be found in the brain_generator.py script and questions about the generative model should be addressed to the SynthSeg team.
If you use GOUHFI in your research, please cite the following:
- For the original GOUHFI paper published in Imaging Neuroscience:
Marc-Antoine Fortin, Anne Louise Kristoffersen, Michael Staff Larsen, Laurent Lamalle, Rüdiger Stirnberg, Pål Erik Goa; GOUHFI: A novel contrast- and resolution-agnostic segmentation tool for ultra-high-field MRI. Imaging Neuroscience 2025; 3 IMAG.a.960. doi: https://doi.org/10.1162/IMAG.a.960
- Bibtex entry:
@article{fortin2025gouhfi,
author = {Fortin, Marc-Antoine and Kristoffersen, Anne Louise and Larsen, Michael Staff and Lamalle, Laurent and Stirnberg, Rüdiger and Goa, Pål Erik},
title = {GOUHFI: A novel contrast- and resolution-agnostic segmentation tool for ultra-high-field MRI},
journal = {Imaging Neuroscience},
volume = {3},
pages = {IMAG.a.960},
year = {2025},
month = {10},
issn = {2837-6056},
doi = {10.1162/IMAG.a.960},
url = {https://doi.org/10.1162/IMAG.a.960},
eprint = {https://direct.mit.edu/imag/article-pdf/doi/10.1162/IMAG.a.960/2556634/imag.a.960.pdf},
}
- For GOUHFI 2.0, currently on ArXiV:
Marc-Antoine Fortin, Anne Louise Kristoffersen, Pål Erik Goa; GOUHFI 2.0: A Next-Generation Toolbox for Brain Segmentation and Cortex Parcellation at Ultra-High Field MRI. ArXiv. doi: https://doi.org/10.48550/arXiv.2601.09006
- Bibtex entry:
@misc{fortin2026gouhfi,
title = {GOUHFI 2.0: A Next-Generation Toolbox for Brain Segmentation and Cortex Parcellation at Ultra-High Field MRI},
author = {Fortin, Marc-Antoine and Kristoffersen, Anne Louise and Goa, Paal Erik},
year = {2026},
eprint = {2601.09006},
archivePrefix = {arXiv},
primaryClass = {eess.IV},
doi = {10.48550/arXiv.2601.09006},
url = {https://arxiv.org/abs/2601.09006}
}
- For the original GOUHFI subcortical segmentation model:
@misc{fortin2025gouhfi,
author = {Fortin, M.-A. and Larsen, M. and Kristoffersen, A. L. and Goa, P. E.},
title = {GOUHFI: Generalized and Optimized segmentation tool for Ultra-High Field Images},
year = {2025},
publisher = {Zenodo},
version = {1.0.0},
doi = {10.5281/zenodo.15255556},
url = {https://doi.org/10.5281/zenodo.15255556}
}
- For GOUHFI 2.0 subcortical segmentation and/or cortical parcellation models:
@misc{fortin2025gouhfi_2p0,
author = {Fortin, M.-A. and Larsen, M. and Kristoffersen, A. L. and Goa, P. E.},
title = {GOUHFI 2.0: Generalized and Optimized segmentation tool for Ultra-High Field Images},
year = {2025},
publisher = {Zenodo},
version = {2.0.0},
number = {3},
doi = {10.5281/zenodo.17920473},
url = {https://doi.org/10.5281/zenodo.17920473}
}
We welcome contributions. If you find bugs, have suggestions, or would like to extend the tool, feel free to open an issue or submit a pull request.
This project is licensed under the Apache 2.0 License. See the LICENSE file for details.
Marc-Antoine Fortin
Norwegian University of Science and Technology (NTNU)
Contact: marc.a.fortin@ntnu.no