ImageCAS-X: A dataset and benchmark for validating coronary vessel segmentation and centerline extraction in computed tomography angiography

Kit M. Bransby1, Esther Øksnebjerg1, Kristoffer Kjær1, Jacob Kirkeby1, Yasmin El Youssef2, Aïda Jiménez1, Philip R. Pedersson2, Martina C. de Knegt2, Klaus F. Kofoed2, Rasmus R. Paulsen1

1 DTU Compute, Technical University of Denmark, Kongens Lyngby, Denmark
2 Cardiovascular Research Unit, Copenhagen University Hospital – Rigshospitalet, Copenhagen, Denmark

Paper Code Download Dataset & Weights
Graphical abstract: the ImageCAS-X annotation types and benchmark.

Highlights

  • Voxel-wise annotations of the vessel lumen and coronary segments, alongside centerlines and mesh surfaces, for 800 scans of the publicly available ImageCAS cohort.
  • Benchmarked 5 coronary-specific segmentation methods and 3 general-purpose segmentation methods, measured against the agreement between expert analysts.
  • Performance stratified by disease, image quality, coronary dominance, coronary segment, vessel diameter, and lumen attenuation.
  • The labels support development and validation of methods for lumen segmentation, plaque and perivascular quantification, and haemodynamic modelling.

Benchmark

Table 1. Quantitative comparison of vessel segmentation algorithms (top) and expert analysts (bottom) on the held-out test set.

Method Lumen Segmentation Lumen Centerline
DSC ↑ HD95 ↓ βerr ↓ clDice ↑ ASSD ↓ HD95 ↓
TotalSegmentatorWasserthal et al., Radiology: AI, 2023 70.5 ± 6.2 19.59 ± 6.45 4.6 ± 2.9 76.0 ± 5.3 2.73 ± 0.90 23.26 ± 7.17
3D-FFR-UNetSong et al., IEEE J-BHI, 2022 84.9 ± 5.5 15.93 ± 20.42 4.8 ± 3.5 89.8 ± 5.0 1.71 ± 1.56 17.00 ± 18.23
ADE-HTLZhang et al., IEEE TMI, 2023 87.7 ± 2.8 2.97 ± 3.74 1.5 ± 1.5 93.2 ± 3.2 0.74 ± 0.39 5.61 ± 4.85
Swin-UNETRHatamizadeh et al., MICCAI BrainLes, 2021 87.9 ± 2.7 3.18 ± 3.68 3.8 ± 2.3 92.5 ± 3.0 0.78 ± 0.36 6.16 ± 5.11
ImageCASZeng et al., CMIG, 2023 87.9 ± 2.9 4.45 ± 4.90 4.7 ± 3.1 91.7 ± 3.5 0.90± 0.46 7.81 ± 6.03
nnU-NetIsensee et al., Nature Methods, 2021 89.8 ± 3.2 7.08 ± 12.65 5.6 ± 3.5 92.3 ± 3.6 1.02 ± 0.75 10.41 ± 13.41
nnU-Net + clDiceShit et al., CVPR, 2021 90.0 ± 3.5 9.70 ± 15.36 8.0 ± 4.4 91.7 ± 3.9 1.20 ± 0.99 12.95 ± 14.20
CAS-NetDong et al., Medical Image Analysis, 2023 91.2 ± 2.8 2.99 ± 3.47 1.9 ± 1.5 93.3 ± 3.2 0.73 ± 0.36 5.75 ± 4.98
Inter-observer 92.8 ± 3.1 2.46 ± 3.62 0.4 ± 0.4 95.4 ± 3.6 0.53 ± 0.33 4.58 ± 5.75
ImageCAS (labels)Zeng et al., CMIG, 2023 41.8 ± 6.7 16.15 ± 8.25 7.0 ± 6.7 78.2 ± 6.9 2.23 ± 0.94 18.89 ± 8.91

Values are mean ± standard deviation. Bold indicates the best performing algorithm. DSC Dice similarity coefficient (%). HD95 95th percentile Hausdorff distance (mm). βerr Betti number error. clDice centerline Dice (%). ASSD average symmetric surface distance (mm).

Citation

@article{bransby2026imagecasx,
  title   = {ImageCAS-X: a dataset and benchmark for coronary artery segmentation and centerline extraction in coronary CT angiography},
  author  = {Bransby, Kit M. and {\O}ksnebjerg, Esther and Kj{\ae}r, Kristoffer and
             Kirkeby, Jacob and El Youssef, Yasmin and Jim{\'e}nez, A{\"i}da and
             Pedersson, Philip R. and de Knegt, Martina C. and Kofoed, Klaus F. and
             Paulsen, Rasmus R.},
  journal = {arXiv preprint},
  year    = {2026}
}

Contact