TopCoW Training Data and External Testsets

Training Data and External Testsets from the TopCoW Challenge

This Zenodo upload accompanies the TopCoW challenge summary paper (see Citation). It contains the training data released by the TopCoW challenge and four external testsets used in the paper's analysis and benchmark. We create this Zenodo record as a permanent release of the TopCoW challenge data. The external testsets, which came from public sources and have all been annotated in the same fashion as the TopCoW training data, serve as continued benchmark for anyone interested in reproducing the evaluations in our paper. The combined data also serve as a jumping-off point to generalize the TopCoW labels to other datasets.

For more details and analysis on the data in this upload, please refer to our challenge summary paper. If you use the data from this Zenodo upload, please cite the TopCoW challenge paper (see Citaion).

We also have a Zenodo "Software" upload with the best performing dockers from the TopCoW challenge. Check it out:

Best Performing TopCoW Segmentation Dockers

Contents of Data

Each zip contains the MRA and/or CTA images (in LPS+ orientation), and the labels: CoW segmentation masks, CoW region of interest (ROI), and CoW graphs. The data ZIPs typically have the following sub-folders:

• All the images are stored in the sub-folder `imagesTr`. The images are angiographic scans in nifti format,  LPS+ orientation.
• The CoW segmentation masks are in the sub-folder `cow_seg_labelsTr`.
• Size and location for the CoW ROI in text file format are in `roi_loc_labelsTr`. The x,y,z of the ROI correspond to LPS+ orientation.
• Yml files indicating the presence of edges (0: absent, 1: present) of CoW graphs are stored in sub-folder `antpos_edges_labelsTr`.

Additionally, for each external testset, we also provide a `raw/` sub-folder with raw images and their braincase cropping and CoW labels:

• `raw/raw_imgs`: raw images from the public external datasets
• `raw/cow_seg_labels_for_raw_imgs`: CoW masks that have reversed the crop-orient steps to map back to the raw images (they are otherwise identical to the masks in `cow_seg_labelsTr` except the coordinate and orientation)
• `raw/braincase_ROI_after_LPS_reorient`: braincase ROI to crop the LPS-reoriented raw images

Note: The contents under `raw/` in external testsets are for transparency purpose and help demonstrate the steps to convert external dataset to be compatible with TopCoW training data. Conversely, the `raw/` data also serve as plug-and-play TopCoW annotations for the respective external datasets. But for TopCoW context, you do not need to use `raw/`.

(Optional: Raw to TopCoW Compatible)

The external testsets in this release were pre-processed in the same manner as the TopCoW training data. These steps include LPS+ reorientation, cropping to the braincase region, and defacing, as described in our paper. The original raw images from the external testsets are stored in `raw/`. You can use similar steps to convert any new dataset to be compatible with TopCoW training data. Here are some tips:

As highlighted in our Best Performing TopCoW Segmentation Dockers zenodo upload, the only pre-processing needed for TopCoW-compatibility for any new dataset is the LPS+ orientation. We have already uploaded the `reorient_nii.py` in the best-dockers zenodo link. You can use it to re-orient new images.

The external testsets in this release have been cropped to the braincase region using the braincase ROI text files in the sub-folder `raw/braincase_ROI_after_LPS_reorient`. Braincase cropping is optional for docker inference, but it will make the data more similar to the TopCoW training data. You can use our example code snippet here to crop the braincase region (after you have reoriented the raw image to LPS+):

https://github.com/CoWBenchmark/TopCoW_Eval_Metrics/blob/master/topcow24_eval/utils/crop_sitk.py 

Info on Each ZIP

TopCoW2024_Data_Release.zip

• Training data release zip used in the TopCoW challenge
• License: "Open use. Must provide the source. Use for commercial purposes requires permission of the data owner." as defined by the OpenData Swiss
• Refer to the README and License in the zip for details and citation

CTA_ISLES2024_TUM.zip

• 26 CTA cases from the training data of the ISLES 2024 challenge. The cases were from TUM center.
• Original License: Creative Commons license CC BY-NC (Attribution-NonCommercial)
• On top of the TopCoW citation, please also cite the ISLES24 works (data descritptor and benchmarking preprints) when using our derived and annotated version

CTA_LargeIA.zip

• 20 CTA cases from the "LargeIA" dataset on Zenodo and paper. Our 20 selected cases do NOT have aneurysms in the CoW ROI.
• As requested by the LargeIA authors, please acquire the original images from their Zenodo link above.
• We provide the CoW masks corresponding to the raw images in the sub-folder `raw/cow_seg_labels_for_raw_imgs`.
• You can also use the `reorient_nii.py` and `crop_sitk.py` (or similar) to reorient the raw image to LPS and crop to the braincase region (the braincase ROI is provided in the `braincase_ROI_after_LPS_reorient` sub-folder under `raw/`). After LPS-reorientation and braincase cropping with the provided braincase ROI, the resulting cropped image should match the CoW masks provided in sub-folder `cow_seg_labelsTr`.
• On top of the TopCoW citation, please also cite their Patterns 2021 paper when using our derived and annotated version

MRA_Lausanne.zip

• 20 MRA cases from the OpenNeuro Lausanne TOF-MRA Aneurysm Cohort dataset. Our 20 selected cases do NOT have aneurysms as they are from the control group.
• Original License: CC0 “No Rights Reserved”
• On top of the TopCoW citation, please also cite their Neuroinformatics 2022 paper when using our derived and annotated version

MRA_IXI_HH.zip

• 20 MRA cases from the IXI dataset. Our 20 selected cases were from the HH hospital.
• Original License: Creative Commons CC BY-SA 3.0 license
• On top of the TopCoW citation, please also cite the IXI dataset website when using our derived and annotated version

Citation

The dataset in this Zenodo upload were released by the TopCoW challenge and used for benchmarking analysis. If you use the TopCoW released data, or our derived and annotated versions of the external testsets, please cite our TopCoW challenge summary paper:

Yang, K., Musio, F., Ma, Y., Juchler, N., Paetzold, J. C., Al-Maskari, R., ... & Menze, B. (2024). Benchmarking the cow with the topcow challenge: Topology-aware anatomical segmentation of the circle of willis for cta and mra. ArXiv, arXiv-2312.

On top of the TopCoW paper citation, please also cite the relevant external dataset publications from the above bullet-points in "Info on Each ZIP".