This official repository houses baseline methods, training scripts, and pretrained models for the LUMIR challenge at Learn2Reg 2025.
The challenge is dedicated to unsupervised brain MRI image registration and offers a comprehensive dataset of over 5000 preprocessed T1-weighted 3D brain MRI images, available for training, testing, and validation purposes.
Please visit learn2reg.grand-challenge.org for more information.
- Download Training Dataset: Access the training dataset via Google Drive (~51GB).
- Download Validation Dataset: Access the validation dataset via Google Drive (~430MB).
- To support participants, we also provide coarse anatomical label maps (10 classes) for the validation datasets, enabling participants to perform quick self-validation prior to leaderboard submission. However, evaluations on the leaderboard as well as the test phase will be conducted on much finer anatomical structures (over 100 classes). Access the simplified label maps via Google Drive (~14MB).
- Preprocessing: The OpenBHB dataset underwent initial preprocessing by its creators, which included skull stripping and affine registration. For comprehensive details, refer to the OpenBHB GitHub page and their article. Subsequently, we performed N4 bias correction with ITK and intensity normalization using a pre-existing tool.
- Annotation: We conducted segmentation of the anatomical structures using automated software. To enhance the dataset for evaluation purposes, an experienced radiologist and neurologist contributed manual landmark annotations to a subset of the images.
- Image size: The dimensions of each image are
160 x 224 x 192. - Normalization: Intensity values for each image volume have been normalized to fall within the range
[0,255]. - Dataset structure:
LUMIR/imagesTr/------ LUMIRMRI_0000_0000.nii.gz <--- a single brain T1 MR image LUMIRMRI_0001_0000.nii.gz LUMIRMRI_0002_0000.nii.gz ....... LUMIR/imagesVal/------ LUMIRMRI_3454_0000.nii.gz LUMIRMRI_3455_0000.nii.gz - Dataset json file: LUMIR25_dataset.json
Learning-based models: TBA
Learning-based foundation models:
- SynthMorph (Official website | Scripts)
- ConvexAdam (Official website | Scripts)
- SynthSR + VFA (SynthSR official website | VFA official website | Scripts)
Optimization-based methods:
- SyN (ATNs) TBA
- deedsBCV TBA
- TRE (Code)
- Dice (Code)
- HD95 (Code)
- Non-diffeomorphic volumes (NDV) (Code) See this article published in IJCV, and its associated GitHub papge
We expect to provide displacement fields for all registrations in the file naming format should be disp_PatID1_PatID2, where PatID1 and PatID2 represent the subject IDs for the fixed and moving images, respectively. The evaluation process requires the files to be organized in the following structure:
folder.zip
└── folder
├── disp_3455_3454.nii.gz
├── disp_3456_3455.nii.gz
├── disp_3457_3456.nii.gz
├── disp_3458_3457.nii.gz
├── ...
└── ...Submissions must be uploaded as zip file containing displacement fields (displacements only) for all validation pairs for all tasks (even when only participating in a subset of the tasks, in that case submit deformation fields of zeroes for all remaining tasks). You can find the validation pairs for in the LUMIR_dataset.json. The convention used for displacement fields depends on scipy's map_coordinates() function, expecting displacement fields in the format [X, Y, Z,[x, y, z]] or [[x, y, z],X, Y, Z], where X, Y, Z and x, y, z represent voxel displacements and image dimensions, respectively. The evaluation script expects .nii.gz files using full-precision format and having shapes 160x224x196x3. Further information can be found here.
Note for PyTorch users: When using PyTorch as deep learning framework you are most likely to transform your images with the grid_sample() routine. Please be aware that this function uses a different convention than ours, expecting displacement fields in the format [X, Y, Z,[x, y, z]] and normalized coordinates between -1 and 1. Prior to your submission you should therefore convert your displacement fields to match our convention.
@article{chen2025beyond,
title={Beyond the LUMIR challenge: The pathway to foundational registration models},
author={Chen, Junyu and Wei, Shuwen and Honkamaa, Joel and Marttinen, Pekka and Zhang, Hang and Liu, Min and Zhou, Yichao and Tan, Zuopeng and Wang, Zhuoyuan and Wang, Yi and others},
journal={arXiv preprint arXiv:2505.24160},
year={2025}
}
@article{dufumier2022openbhb,
title={Openbhb: a large-scale multi-site brain mri data-set for age prediction and debiasing},
author={Dufumier, Benoit and Grigis, Antoine and Victor, Julie and Ambroise, Corentin and Frouin, Vincent and Duchesnay, Edouard},
journal={NeuroImage},
volume={263},
pages={119637},
year={2022},
publisher={Elsevier}
}
@article{taha2023magnetic,
title={Magnetic resonance imaging datasets with anatomical fiducials for quality control and registration},
author={Taha, Alaa and Gilmore, Greydon and Abbass, Mohamad and Kai, Jason and Kuehn, Tristan and Demarco, John and Gupta, Geetika and Zajner, Chris and Cao, Daniel and Chevalier, Ryan and others},
journal={Scientific Data},
volume={10},
number={1},
pages={449},
year={2023},
publisher={Nature Publishing Group UK London}
}
@article{marcus2007open,
title={Open Access Series of Imaging Studies (OASIS): cross-sectional MRI data in young, middle aged, nondemented, and demented older adults},
author={Marcus, Daniel S and Wang, Tracy H and Parker, Jamie and Csernansky, John G and Morris, John C and Buckner, Randy L},
journal={Journal of cognitive neuroscience},
volume={19},
number={9},
pages={1498--1507},
year={2007},
publisher={MIT Press One Rogers Street, Cambridge, MA 02142-1209, USA journals-info~…}
}
If you have used Non-diffeomorphic volumes in the evaluation of the deformation regularity, please cite the following:
@article{liu2024finite,
title={On finite difference jacobian computation in deformable image registration},
author={Liu, Yihao and Chen, Junyu and Wei, Shuwen and Carass, Aaron and Prince, Jerry},
journal={International Journal of Computer Vision},
pages={1--11},
year={2024},
publisher={Springer}
}
