TopCoW Challenge Data¶

The challenge data cohort is composed of patients admitted to the Stroke Center of the University Hospital Zurich (USZ) in 2018 and 2019. The inclusion criteria for the challenge data are: 1) Both MRA and CTA scans are available for that patient; 2) at least the MRA or CTA allows for an assessment of the CoW anatomy and geometry. The patients of the challenge cohort were recovering from a stroke-related neurological disorder, including ischemic stroke, transient ischemic attack, stroke mimic, retinal infarct or amaurosis fugax, intracerebral hemorrhage, and cerebral sinus vein thrombosis.

Data Acquisition¶

The two imaging modalities used are: Computed tomography Angiography (CTA) and Time of Flight Magnetic Resonance Angiography (TOF-MRA) or MRA. The data was acquired at the University Hospital of Zurich (USZ) during routine examinations following standard procedures for MRA and CTA. MRA scans were typically acquired by SIEMENS Skyra model or Avanto Fit model, with magnetic field strength of 3 Tesla or 1.5 Tesla, and with TOF-3D mode or TOF-3D multi slab mode. CTA is typically acquired by SIEMENS SOMATOM Definition Flash using Dual Energy.

For CTA, the voxel size ranges from around 0.34 to 0.53 mm in the X-Y dimension, and ranges from around 0.62 to 0.75 mm in the Z dimension. For MRA, the voxel size ranges from around 0.29 to 0.35 mm in the X-Y dimension, and ranges from around 0.5 to 0.6 mm in the Z dimension.

Anonymization and Defacing¶

The data is anonymized (removal and anonymization of relevant DICOM patient information). Additional de-facing and cropping procedures are performed to ensure patient privacy in the image data. Specifically, we mask out or shear-cut the facial regions, and then crop the image data to include only the braincase region.

Training and Test Set¶

Training, validation, and test cases all have the MRA and CTA joint-modality pairs, with one scan for each modality. The task is to segment the CoW vessels (Task 1) and anatomical components on the CoW region (Task 2) in either MRA track or CTA track.

• Training dataset: 90 patients (both image and annotation released)
• Validation set: 5 patients (image and roi released to public but without annotations)
• Test set: 35 patients (not released to public)
• In addition fo the above TopCoW challenge data, we also release 20 MRAs from the CROWN challenge for training, and hold 10 cases for internal testing.

All the vessel components of CoW necessary to diagnose the CoW angio-architecture are annotated voxel-wise. Multi-class labels contain a different voxel value for different CoW vessel segments:

• 0: Background, 1: BA, 2: R-PCA, 3: L-PCA, 4: R-ICA, 5: R-MCA, 6: L-ICA, 7: L-MCA, 8: R-Pcom, 9: L-Pcom, 10: Acom,11: R-ACA, 12: L-ACA, 15: 3rd-A2

For binary segmentation task, a separate binary vessel label is provided by combining the multi-class labels:

• 0: Background, 1: CoW Vessel

Access Training Data¶

The training (Tr) data folder has the following sub-folders:

• imagesTr: Angiographic scans in nifti format, 16-bit signed. LPS+ orientation
  • The nifti files are saved with schema: {challenge}_{modality}_{fov}_{pat_id}.nii.gz
    • challenge: topcow or crown
    • modality: mr for MRA, ct for CTA
    • fov (field of view): whole for whole brain-case fov or roi for CoW ROI
    • pat_id: patient ID, 001, 002, ...
• mul_labelsTr: Multi-class segmentation nifti with CoW anatomical labels
• bin_labelsTr: Binary segmentation labels (Background: 0; CoW Vessel: 1)
• roi_size_loc: Size and location info for the CoW region of interest (ROI) i.e. the 3D bounding box:
  • Text file containing the size and location of the 3D bounding box info
    • Size = number of pixels along the x, y, z axis
    • Location = coordinate of the x-min, y-min, z-min (0-indexed)
  • Preview image {challenge}_{modality}_roi_{pat_id}_axial.png showing five sample axial slices with the bounding box drawn in red
• visual_qc: Visual quality control (not for machine learning). For each study, we provide:
  • Two videos of the angiography before and after annotation in MP4 format: {challenge}_{modality}_{pat_id}_{before, after}.mp4
  • Preview png {challenge}_{modality}_{fov}_{pat_id}_preview.png for the ROI and whole nifti

Example tree view of the released data folder:

# Note that TopCoW data has two modalities <mr> and <ct>
topcow_batch-1_40pairMRCT_30062023/
├── bin_labelsTr
│   ├── topcow_ct_roi_024.nii.gz
│   ├── topcow_ct_whole_024.nii.gz
│   ├── topcow_mr_roi_024.nii.gz
│   └── topcow_mr_whole_024.nii.gz
├── imagesTr
│   ├── topcow_ct_roi_024_0000.nii.gz
│   ├── topcow_ct_whole_024_0000.nii.gz
│   ├── topcow_mr_roi_024_0000.nii.gz
│   └── topcow_mr_whole_024_0000.nii.gz
├── mul_labelsTr
│   ├── topcow_ct_roi_024.nii.gz
│   ├── topcow_ct_whole_024.nii.gz
│   ├── topcow_mr_roi_024.nii.gz
│   └── topcow_mr_whole_024.nii.gz
├── roi_size_loc
│   ├── topcow_ct_roi_024_axial.png
│   ├── topcow_ct_roi_024.txt
│   ├── topcow_mr_roi_024_axial.png
│   └── topcow_mr_roi_024.txt
└── visual_qc
    ├── topcow_ct_024_after.mp4
    ├── topcow_ct_024_before.mp4
    ├── topcow_ct_roi_024_preview.png
    ├── topcow_ct_whole_024_preview.png
    ├── topcow_mr_024_after.mp4
    ├── topcow_mr_024_before.mp4
    ├── topcow_mr_roi_024_preview.png
    └── topcow_mr_whole_024_preview.png

# CROWN subset
crown_20_mr_01062023/
├── bin_labelsTr
│   ├── crown_mr_roi_001.nii.gz
│   └── crown_mr_whole_001.nii.gz
├── imagesTr
│   ├── crown_mr_roi_001_0000.nii.gz
│   └── crown_mr_whole_001_0000.nii.gz
├── mul_labelsTr
│   ├── crown_mr_roi_001.nii.gz
│   └── crown_mr_whole_001.nii.gz
├── roi_size_loc
│   ├── crown_mr_roi_001_axial.png
│   └── crown_mr_roi_001.txt
└── visual_qc
    ├── crown_mr_001_after.mp4
    ├── crown_mr_001_before.mp4
    ├── crown_mr_roi_001_preview.png
    └── crown_mr_whole_001_preview.png

Data usage License¶

Our training and validation data are released under the CC BY-NC (Attribution-NonCommercial) license.

By downloading the data you agree with the license terms.

Subset from CROWN Challenge¶

For the subset shared by the CROWN Challenge , please refer to their data usage agreement for more details.

If you use TopCoW or CROWN challenge data in your work, please cite our challenge pre-prints:

  @misc{topcowchallenge,
        title={Benchmarking the CoW with the TopCoW Challenge: Topology-Aware Anatomical Segmentation of the Circle of Willis for CTA and MRA}, 
        author={Kaiyuan Yang and Fabio Musio and Yihui Ma and Norman Juchler and Johannes C. Paetzold and Rami Al-Maskari and Luciano Höher and Hongwei Bran Li and Ibrahim Ethem Hamamci and Anjany Sekuboyina and Suprosanna Shit and Houjing Huang and Chinmay Prabhakar and Ezequiel de la Rosa and Diana Waldmannstetter and Florian Kofler and Fernando Navarro and Martin Menten and Ivan Ezhov and Daniel Rueckert and Iris Vos and Ynte Ruigrok and Birgitta Velthuis and Hugo Kuijf and Julien Hämmerli and Catherine Wurster and Philippe Bijlenga and Laura Westphal and Jeroen Bisschop and Elisa Colombo and Hakim Baazaoui and Andrew Makmur and James Hallinan and Bene Wiestler and Jan S. Kirschke and Roland Wiest and Emmanuel Montagnon and Laurent Letourneau-Guillon and Adrian Galdran and Francesco Galati and Daniele Falcetta and Maria A. Zuluaga and Chaolong Lin and Haoran Zhao and Zehan Zhang and Sinyoung Ra and Jongyun Hwang and Hyunjin Park and Junqiang Chen and Marek Wodzinski and Henning Müller and Pengcheng Shi and Wei Liu and Ting Ma and Cansu Yalçin and Rachika E. Hamadache and Joaquim Salvi and Xavier Llado and Uma Maria Lal-Trehan Estrada and Valeriia Abramova and Luca Giancardo and Arnau Oliver and Jialu Liu and Haibin Huang and Yue Cui and Zehang Lin and Yusheng Liu and Shunzhi Zhu and Tatsat R. Patel and Vincent M. Tutino and Maysam Orouskhani and Huayu Wang and Mahmud Mossa-Basha and Chengcheng Zhu and Maximilian R. Rokuss and Yannick Kirchhoff and Nico Disch and Julius Holzschuh and Fabian Isensee and Klaus Maier-Hein and Yuki Sato and Sven Hirsch and Susanne Wegener and Bjoern Menze},
        year={2024},
        eprint={2312.17670},
        archivePrefix={arXiv},
        primaryClass={cs.CV}
  }