LUMEN_CamSVD_nnUNet

LUMEN CamSVD nnUNet Model

Model Details

Model Description

This repository hosts the nnU-Net model trained for vessel segmentation as part of the LUMEN pipeline. For details, please refer to the LUMEN paper.

• Developed by: Rui Li
• Model type: nnU-Net
• Training dataset: 7T TOF-MRAs from the Cambridge 7T Cerebral Small Vessel Disease (CamSVD) study, consisting of healthy controls, and subjects with lacunar strokes with SVD and non-lacunar strokes without SVD.
• License: CC BY

Model Sources

<!-- Provide the basic links for the model. -->

LUMEN pipeline:

• Repository: https://github.com/RuiLiGitLove/LUMEN
• Paper: Li, Rui, et al. "LUMEN–A Deep Learning Pipeline for Analysis of the 3D Morphology of the Cerebral Lenticulostriate Arteries from Time-of-Flight 7T MRI." NeuroImage (2025): 121377.

nnU-Net:

• Repository: https://github.com/MIC-DKFZ/nnUNet
• Paper: Isensee, Fabian, et al. "nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation." Nature methods 18.2 (2021): 203-211.

How to Use the Model

1. Install nnU-Net: follow instructions at https://github.com/MIC-DKFZ/nnUNet.
2. Configure the nnU-Net environment variables nnUNet_raw, nnUNet_preprocessed, nnUNet_results. Follow instructions at https://github.com/MIC-DKFZ/nnUNet/blob/master/documentation/setting_up_paths.md.
3. Download the pretrained models at LUMEN_CamSVD_nnUNet_ResEncL_3d_fullres_5fold.zip.
4. Activate the nnU-Net environment. Run nnUNetv2_install_pretrained_model_from_zip path/to/zip. This will install the pretrained models, which is an ensemble of 5 models trained from 5-fold cross validation, at the path specified by your nnUNet_results variable.
5. Prepare your dataset in the nnUNet_raw folder following the required format. Here is an example structure.

   nnUNet_raw/
    └── Dataset065_MyDataset
        └── imagesTs
            ├── ID001_tof_0000.nii.gz
            ├── ID002_tof_0000.nii.gz
            └── ID003_tof_0000.nii.gz
   

   The _0000 postfix specifies the input channel, which must be 0000 as the model only expects 1 channel for TOF segmentation. No dataset.json or imageTr is needed, as no training is involved.
6. Run inference with this pretrained ensemble of models with the following command. Remember to replace [INPUT_FOLDER] and [OUTPUT_FOLDER] with your own paths.

   nnUNetv2_predict -d Dataset066_CamSVD \
                   -i [INPUT_FOLDER] \
                   -o [OUTPUT_FOLDER] \
                   -f 0 1 2 3 4 \
                   -tr nnUNetTrainer \
                   -c 3d_fullres \
                   -p nnUNetResEncUNetLPlans
   

   Note that the dataset name Dataset066_CamSVD is used to locate the trained model.

Training Details

Details please refer to Li, Rui, et al. "LUMEN–A Deep Learning Pipeline for Analysis of the 3D Morphology of the Cerebral Lenticulostriate Arteries from Time-of-Flight 7T MRI." NeuroImage (2025): 121377.

Training Data

<!-- This should link to a Dataset Card, perhaps with a short stub of information on what the training data is all about as well as documentation related to data pre-processing or additional filtering. -->

Trained with 9 7T TOF-MRA volumes from lacunar stroke patients from the Cambridge 7T Cerebral Small Vessel Disease (CamSVD) study. For each patient, we defined a cubic region-of-interest (ROI) that covers the lenticulostriate arteries (LSA) region in each hemisphere. Therefore, there were 18 ROIs in total, which were split to perform the standard 5-fold cross validation used in the nnU-Net framework.

Training Procedure

<!-- This relates heavily to the Technical Specifications. Content here should link to that section when it is relevant to the training procedure. -->

Preprocessing [optional]

Image intensity normalisation option (for nnU-Net): rescale_to_0_1

Training Hyperparameters

Changed the following hyperparameters from the nnU-Net source code: In training/nnUNetTrainer/nnUNetTrainer.py line 149-151:

self.num_iterations_per_epoch = 100     # default=250
self.num_val_iterations_per_epoch = 20  # default=50
self.num_epochs = 50                    # default=1000

Citation

<!-- If there is a paper or blog post introducing the model, the APA and Bibtex information for that should go in this section. -->

If you are using this model, please cite:

LUMEN

BibTeX:

@article{li2025lumen, title={LUMEN--A Deep Learning Pipeline for Analysis of the 3D Morphology of the Cerebral Lenticulostriate Arteries from Time-of-Flight 7T MRI}, author={Li, Rui and Chatterjee, Soumick and Jiaerken, Yeerfan and Zhou, Xia and Radhakrishna, Chethan and Benjamin, Philip and Nannoni, Stefania and Tozer, Daniel J and Markus, Hugh S and Rodgers, Christopher T}, journal={NeuroImage}, pages={121377}, year={2025}, publisher={Elsevier} }

APA:

Li, R., Chatterjee, S., Jiaerken, Y., Zhou, X., Radhakrishna, C., Benjamin, P., ... & Rodgers, C. T. (2025). LUMEN–A deep learning pipeline for analysis of the 3D morphology of the cerebral lenticulostriate arteries from time-of-flight 7T MRI. Neuroimage, 121377.

nnU-Net

BibTeX:

@article{isensee2021nnu, title={nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation}, author={Isensee, Fabian and Jaeger, Paul F and Kohl, Simon AA and Petersen, Jens and Maier-Hein, Klaus H}, journal={Nature methods}, volume={18}, number={2}, pages={203--211}, year={2021}, publisher={Nature Publishing Group} }

APA:

Isensee, F., Jaeger, P. F., Kohl, S. A., Petersen, J., & Maier-Hein, K. H. (2021). nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation. Nature methods, 18(2), 203-211.

Model Card Author

Rui Li

Model Card Contact

Rui Li [email protected]