Skip to main content
[Preprint]. 2023 Dec 29:arXiv:2312.17670v1. [Version 1]

Table 3.

Method summary of all participating teams, ordered alphabatically by team name. For detailed descriptions and relevant references cited, please refer to the Appendix A. Awards lists the number of first (gold), second (silver) and third (bronze) prizes a team has won in the TopCoW challenge. The winning team names are in bold.

Team Track Task Awards Method Highlights
2i_mtl CTA Binary 3D AttentionUNet for segmentation;
3D autoencoder to mitigate false positives

  • - Use provided CoW ROI for training

  • - Image & mask input to autoencoder
agaldran CTA MRA Binary
Multiclass		 3D dynamic UNet

  • - Ensembling

  • - Cross-validation on patch size
EURECOM CTA
MRA		 Binary SynthSeg for Brain mask extraction;
A2V for multi-modal segmentation

  • - Single model for both modalities

  • - 2D axial slice input

  • - Use provided CoW ROI for patch extraction
gbCoW MRA Binary
Multiclass		 3D nnUNet

  • - Only multiclass labels used

  • - Turned off data augmentation
gl CTA
MRA		 Multiclass Atlas registration for custom ROI extraction;
3D MedNexT & UX-Net for binary and subsequent multiclass segmentation

  • - Dataset specific atlas

  • - Binary mask input for multiclass segmentation

  • - Ensembling
IWantToGoToCanada CTA Binary
Multiclass		 3D nnUNet for binary segmentation;
3D Swin-UNETR for subsequent multiclass segmentation

  • - Binary mask input for multiclass segmentation
junqiangchen CTA
MRA		 Binary
Multiclass		 VNet3D for custom ROI extraction;
VNet3D for ROI segmentation

  • - Brain mask extraction

  • - Binary mask for custom ROI
lWM CTA
MRA		 Binary
Multiclass		 3D ResidualUNet

  • - Custom end-to-end UNet
NexToU CTA
MRA		 Binary
Multiclass		 graphic file with name nihpp-2312.17670v1-t0026.jpg 3D nnUNet for low-res binary segmentation;
NexToU architecture for full-res cascading

  • - Centerline boundary Dice (cb-Dice)

  • - Binary topological interaction (BTI) module
NIC-VICOROB-1 CTA
MRA		 Binary
Multiclass		 graphic file with name nihpp-2312.17670v1-t0027.jpg 3D nnUNet

  • - Ensembling

  • - Binary mask input for CT multiclass
NIC-VICOROB-2 CTA
MRA		 Binary
Multiclass		 3D AttentionUNet for binary segmentation;
2D AttentionUNet for subsequent multiclass segmentation

  • - Axial slice input for multiclass segmentation

  • - Binary mask input for multiclass segmentation

  • - Use provided CoW ROI for 3D patch extraction
Organizers CTA
MRA		 Multiclass graphic file with name nihpp-2312.17670v1-t0028.jpg nnDetection for custom ROI extraction;
3D nnUNet for multiclass segmentation

  • - Image registration for data augmentation

  • - ROI object detection

  • - Ensembling
refrain MRA Binary
Multiclass		 graphic file with name nihpp-2312.17670v1-t0029.jpg Atlas registration for custom ROI extraction;
3D nnUNet for segmentation

  • - Data augmentation for rare CoW variants

  • - Segment specific loss weighting
sjtu_eiee_2–426lab CTA Binary
Multiclass		 graphic file with name nihpp-2312.17670v1-t0030.jpg 3D nnUNet for custom ROI extraction;
3D nnUNet for ROI segmentation

  • - Binary mask for custom ROI
UB-VTL CTA
MRA		 Binary Modified 3D Brave-Net

  • - clDice loss for connectedness

  • - Residual connections & PReLu activations

  • - Use provided CoW ROI for patch extraction
UW MRA Binary graphic file with name nihpp-2312.17670v1-t0031.jpg 3D nnUNet

  • - 3-component loss (Dice + CE + TopK)

  • - Ensembling
WilliWillsWissen CTA
MRA		 Binary
Multiclass		 graphic file with name nihpp-2312.17670v1-t0032.jpg 3D nnUNet

  • - clDice/skeleton recall for connectedness

  • - Training on both modalities

  • - Ensembling
ysato MRA Binary Auto vessel thresholding; Region growing

  • - Non-deep learning algorithm

  • - Short inference time (~15s per case)

  • - Little computing power needed (done on CPU)