Notice: Our latest contributions on the topic are available through our new repository at GitHub.
 |
 |
 |
 |
| TOF-MRA | PDw | T1w | T2w |
This repository contains data and models to reproduce the results of the 2023 publication "Segmentation and Anatomical Annotation of Cerebral Arteries in Non-Angiographic MRI" (full Reference). Our trained models can be used with the nnU-Net Framework to predict anatomical label for new data.
The models have been trained to predict arterial vessels around the Circle-of-Willis (CoW) for angiographic and non-angiographic MRI. The dataset contained TOF-MRA, T1w, T2w and PDw scans, but the models generated reasonable predictions for different modalities as well (e.g., T1w MPRAGE MRA). The models have been trained on all four modalities in a 5-fold cross validation, which is why no specific model needs to be selected for a given modality. During inference, you may also choose to predict using a single model from one fold only. Our recommendation (and the default configuration during inference) is to use all folds in an ensemble.
Please refer to our publication if you are interested in the details, e.g., how the ground-truth annotations have been generated.
 |
 |
|
(a) Anatomical outline of the Circle-of-Willis (CoW): Arterial vessels and their flow directions are indicated through red lines, and abbreviations of the respective vessel segments are shown (e.g., ICA, MCA, PCA, ...). Segments which may not be present in every patient are shown as dashed lines. Marker points used for the Ground-Truth generation algorithm are shown in blue. The resulting four circulatory regions (Left, Right, Anterior, Posterior) are labeled and their borders indicated in the background colors. |
(b) Exemplar ground-truth annotation for IXI002: circulatory regions and marker points match the anatomical outline described in (a). |
 |
 |
 |
 |
| TOF-MRA | PDw | T1w | T2w |
| Predictions for IXI002 of the cross-validation split (top) and IXI347 of the test-split (bottom). | |||
Our work fully relies on the nnU-Net framework for training and inference. You can
checkout their Github repository or install the package from PyPI:
pip install nnunetv2. However, be careful to set up Pytorch correctly for your system before installing nnU-Net.
For more information, check the training.ipynb and the nnU-Net Instructions.
We provide our nnU-Net compatible dataset, named Dataset600_IXI, used for training our models in the subfolder data.
To reduce the size of the repository, all data derived from the IXI dataset has been omitted and needs to be downloaded separately.
The IXI dataset is publicly available at brain-development.org.
Follow the instructions at prepare.ipynb to easily reconstruct the full dataset.
With the reconstructed dataset you can subsequently train models using the nnU-Net Framework as described in training.ipynb.
We provide pretrained models for our dataset at Google Drive. Follow the instructions at inference.ipynb to use our models with the nnU-Net framework and generate predictions for your data.
The output labels predicted by our models are as follows:
|
|
Our paper will be published at the proceedings of the DMIP conference at ACM. We also share the accepted version of our paper.
If you use our results in your research, we would appreciate you citing the following conference paper:
- B. Sabrowsky-Hirsch, P. Moser, S. Thumfart and J. Scharinger. Segmentation and Anatomical Annotation of Cerebral Arteries in Non-Angiographic MRI. In: Proceedings of the 2023 6th International Conference on Digital Medicine and Image Processing. 2023.
@inproceedings{SabrowskyCerebral2023,
author = {Bertram Sabrowsky-Hirsch, Philipp Moser, Stefan Thumfart and Josef Scharinger},
title = { Segmentation and Anatomical Annotation of Cerebral Arteries in Non-Angiographic MRI},
booktitle = {Proceedings of the 2023 6th International Conference on Digital Medicine and Image Processing},
month = nov,
year = 2023,
publisher = {Association for Computing Machinery},
isbn = {979-8-4007-0942-5}
}
If you have any inquiries, please open a GitHub issue.
This project is financed by research subsidies granted by the government of Upper Austria. RISC Software GmbH is Member of UAR (Upper Austrian Research) Innovation Network.