Spatio-temporal carotid artery segmentation for 4D Flow MRI.

Use the package manager pip to install the packages in requirements.txt.

pip install -r requirements.txt

The following models are available in models.py:

• 2D U-Net - unet2d
• 3D U-Net - unet3d
• UNETR unetr
• Spatio-Temporal Transformer spatio_temporal_transformer

The following temporal dimension processing strategies (transforms) are available:

• Interpolation - interpolate
• Padding with cropping - pad-crop
• Padding with averaging - pad-average

Image sizes, keys for transforms and training parameters can be changed in constants.py. Images can either have 1 (magnitude) or 3 channels (magnitude + velocities).

To train individual folds for cross-validation

python train.py -dp <DATASET FOLDER> -o <OUTPUT FOLDER> -m <MODEL> -t <TRANSFORM> -c <CHANNELS> -d <DEPTH> -f <FOLD> 

A splits_final.json file specifying splits in the nnU-Net convention [1] is required for cross-validation, and should be placed in the dataset folder.

To evalate trained models on all folds of the validation data:

python validation.py -m <MODEL FOLDER> -n <MODEL> -d <DATASET FOLDER> -o <OUTPUT FOLDER> -c <CHANNELS> -i <DEPTH> -t <TRANSFORM> -p

A splits_final.json file specifying splits in the nnU-Net convention [1] is required, and should be placed in the dataset folder.

To perform ensemble inference with majority voting:

python ensemble.py -m <MODEL FOLDER> -o <OUTPUT FOLDER> -n <MODEL> -i <IMAGE FOLDER> -c <CHANNELS> -d <DEPTH> -t <TRANSFORM> -p

A splits_final.json file specifying splits in the nnU-Net convention [1] is required, and should be placed in the dataset folder.

To perform inference with a single model:

python inference.py -i <IMAGE FOLDER> -o <OUTPUT FOLDER> -m <MODEL CHECKPOINT PATH> -c <CHANNELS> -d <DEPTH> -t <TRANSFORM> -p

To save predictions without temporal post-processing discard the -p.

To evaluate predictions:

python evaluation.py -i <IMAGE FOLDER> -gt<GROUNT TRUTH LABELS> -s <SAVE FOLDER> -p <PREDICTED LABELS> -v <VELOCITY IMAGE FOLDER> -p

If evaluating predictions that have not been post-processed, omit -p. Velocity images are required for velocity based metrics.

[1] Isensee F. nnU-Net dataset format https://github.com/MIC-DKFZ/nnUNet/blob/master/documentation/dataset_format.md (Accessed 19 June 2025)

The code in this repository is licensed under the GNU-GPL license.