Beyond Segmentation:
Enriching ImageCAS with Coronary Ostia Landmarks, Rooted Centerline Graphs, and Open-Weight Localization Models

Matteo Leccardi1 ORCID Google Scholar, Andrea Moglia1 ORCID Google Scholar, Marco Morandini1, Michele Brembilla1,
Alessandro Nicosia1, Andrea Penserini1, Christian Brandini1,
Ender Konukoglu2 ORCID, Marco Marcon1 ORCID Google Scholar,
Luca Mainardi1 ORCID Google Scholar, Pietro Cerveri1,3 ORCID Google Scholar

1: Politecnico di Milano, 2: ETH Zurich, 3: Università di Pavia
For correspondence: matteo.leccardi@polimi.it

Paper arXiv GitHub Dataset (Zenodo) Model weights (Zenodo)
B3Lab CVL - ETHZ
Ostia mini-challenge
Overview of the work: datasets, annotation protocol, benchmark and public release

Abstract

Coronary ostia and centerlines are key reference points for automated analysis of coronary computed tomography angiography (CCTA), from surgical planning to intraoperative vessel tracking. However, the only public ostium annotations cover 77 non-gated aortic scans, not CCTA, while CAT08 and ASOCA (32 and 40 CCTA scans) mark the ostia only implicitly, as centerline start points. This work enriches ImageCAS, the largest public dataset for coronary artery segmentation, with ostium annotations for its 1000 patients and centerline graphs rooted at these new ostia, making it, to the authors' knowledge, the largest public dataset for ostium localization and the only one with centerlines following the official ImageCAS masks and starting at annotated ostia. Three annotators marked each ostium in the first 500 volumes; in the next 500, the candidates of a model trained on the first batch were corrected (human-in-the-loop).

Six 3D localization models, trained on ImageCAS, CAT08 and ASOCA, were benchmarked on 58 held-out volumes (50 ImageCAS, 4 CAT08, 4 ASOCA). SwinUNETR-V2 and nnResUNet localized all ostia without failure (>5 mm error), with median errors of 0.70–0.87 mm. Trained on CAT08 and ASOCA alone, SwinUNETR-V2 fails on 34% of right and 53% of left ostia, whereas adding 250 ImageCAS volumes removes all failures.

The centerlines were extracted by a deterministic, quality-controlled minimal-path procedure; on ASOCA, they cover a median of 99.5% of the reference length, with a median per-patient mean distance of 0.22 mm between matched reference points and centerlines. Annotations, centerlines, code, and weights are available at https://hal9000-lab.github.io/.

At a glance

1000ImageCAS patients annotated
6annotators, rotating groups of three
~5 hinstead of ~144 h for the semi-automatic half
999centerline graphs out of 1000 patients
4open-weight benchmark models

Ostia annotations

The coronary ostia are the points where the left and right coronary arteries originate from the aorta. ImageCAS provides 1000 CCTA volumes with coronary segmentation masks but no ostium annotation; we add one left and one right ostium landmark to every volume.

Left and right ostium landmarks on axial and sagittal CCTA slices
1 of 100
Left (navy) and right (firebrick) ostium landmarks on an axial and a sagittal slice through each ostium, for 100 volumes of ImageCAS, ASOCA and CAT08. The examples change every few seconds; use the arrows or the slider to browse them.

From annotation to benchmark

How the ostia dataset was assembled and how the benchmark was trained and evaluated, step by step: the two annotation stages, the assistant model, the test set, the benchmark models and the ablation studies. Use the chapter buttons to jump to a step.

Want to try it yourself? Place the ostia on real CT scans and compare your accuracy with our models: Ostia mini-challenge

0:00
Schematic illustration: every square stands for one CCTA volume, and the order of the annotator groups is illustrative.

Centerline graphs

Centerline graphs rooted at the annotated ostia are extracted from the ImageCAS segmentation masks with a deterministic minimal-path tree-growing procedure. They are validated against the expert centerlines of ASOCA.

Centerline extraction pipeline
Centerline extraction pipeline: from the segmentation mask and the annotated ostia to one rooted tree per side, with per-node position, radius, topological type and side.

Centerline tracking, in brief

A simplified 2D illustration of how a centerline tree is traced: from the ostia placed by the model, through the distance to the vessel wall, to the rooted graph.

0:00
Illustration on a synthetic vessel tree. The actual procedure works in 3D on the ImageCAS masks and includes further refinement steps that are left out here.

Centerline tracking, in detail

The same procedure with every step of the pipeline: mask repair, ostium and gap bridging, distance and cost maps, the two minimal-cost fronts, the two growing passes, tip trimming, smoothing and the final typed graph. Use the chapter buttons to jump to a step.

0:00
Illustration on a synthetic vessel tree with a 0.5 mm voxel, using the parameters of the procedure (cost exponent, 5 mm minimum branch length, 2 mm trimming, 2 mm smoothing sigma). The real procedure works in 3D on the ImageCAS masks.

Benchmark

We release open-weight models for automatic ostia localization: SwinUNETRv2, ResNet50, ResNet101 and nnResUNet (MedNeXt is part of the benchmark, but its weights are not released), trained with a differentiable coordinate-regression loss and evaluated on a held-out test set from ImageCAS, CAT08 and ASOCA. Errors are graded against thresholds derived from the ostial anatomy and from the variability between annotators.

The best models, SwinUNETRv2 and nnResUNet, localized all the ostia without any failure, with median errors below one millimetre.

Why it matters

An ostium and a rooted centerline tree with a lumen radius at every node are the starting point of patient-specific models of the coronary circulation. The ostia mark where blood enters each tree and where an inflow condition is applied; the centerline graphs give the topology and the calibre of the vessels that a flow model needs. Such models underlie computational fluid dynamics of coronary blood flow, reduced-order (one-dimensional) hemodynamic models, and the estimation of the functional significance of a stenosis, such as fractional flow reserve (FFR) computed from CT angiography. Reliable, openly available ostia and centerlines make these pipelines easier to build, compare and reproduce.

Illustration: a real-time WebGL simulation of pulsatile flow in an unfolded left (blue) and right (red) coronary tree, with dye released at each ostium. The heartbeat is 60 per minute and the right tree pulses 0.25 s after the left one. The flow is the incompressible, inviscid solution in the vessel geometry scaled by the pulse, with a parabolic profile across each vessel; it is illustrative, not patient-specific and not a clinical model.

Full solver (experimental)

The same trees with a real-time two-dimensional Navier-Stokes solver instead of a precomputed flow. The walls are rigid, the heartbeat drives a periodic pressure at each ostium, the endpoints are resistive outlets, and you can drag over the vessels to place 2x2 obstacles in the flow (Shift-drag erases them).

lowhigh
Experimental, in real time (WebGL2): a viscous, incompressible solver on the unfolded left (blue) and right (red) trees, with rigid no-slip walls. The blood pressure (120/80 mmHg by default) is applied at the ostium (green disk) with the heartbeat, and each endpoint (red dot) is an outlet with a resistance, so the flow divides between the branches according to their geometry and a higher pressure gives a higher flow. The right tree lags the left one by a delay that depends on the heart rate. The pressure view colours the range of pressure present in the trees at each moment (low to high, in mmHg). It is a two-dimensional illustration: not patient-specific and not a clinical model.

Resources

Citation

Paper Preprint

@misc{leccardi2026segmentationenrichingimagecascoronary,
  title         = {Beyond Segmentation: Enriching ImageCAS with Coronary Ostia Landmarks, Rooted Centerline Graphs, and Open-Weight Localization Models},
  author        = {Matteo Leccardi and Andrea Moglia and Marco Morandini and Michele Brembilla and Alessandro Nicosia and
                   Andrea Penserini and Christian Brandini and Ender Konukoglu and Marco Marcon and
                   Luca Mainardi and Pietro Cerveri},
  year          = {2026},
  eprint        = {2611.09714},
  archivePrefix = {arXiv},
  primaryClass  = {eess.IV},
  url           = {https://arxiv.org/abs/2611.09714}
}

Dataset (Zenodo)

@dataset{leccardi_awakening_imagecas_dataset,
  title     = {Awakening ImageCAS: coronary ostia and centerline annotations for ImageCAS},
  author    = {Leccardi, Matteo and Moglia, Andrea and Morandini, Marco and Brembilla, Michele and Nicosia, Alessandro and
               Penserini, Andrea and Brandini, Christian and Marcon, Marco and
               Mainardi, Luca and Cerveri, Pietro},
  publisher = {Zenodo},
  year      = {2026},
  doi       = {10.5281/zenodo.23243093},
  url       = {https://doi.org/10.5281/zenodo.23243093}
}

Original ImageCAS dataset

Our annotations extend ImageCAS: please also cite the original dataset paper. The original images and segmentation masks are available on Kaggle.

@article{zeng2023imagecas,
  title   = {ImageCAS: A large-scale dataset and benchmark for coronary artery segmentation based on computed tomography angiography images},
  author  = {Zeng, An and Wu, Chunbiao and Lin, Guisen and Xie, Wen and Hong, Jin and Huang, Meiping and
             Zhuang, Jian and Bi, Shanshan and Pan, Dan and Ullah, Najeeb and Khan, Kaleem Nawaz and
             Wang, Tianchen and Shi, Yiyu and Li, Xiaoming and Xu, Xiaowei},
  journal = {Computerized Medical Imaging and Graphics},
  volume  = {109},
  pages   = {102287},
  year    = {2023},
  doi     = {10.1016/j.compmedimag.2023.102287}
}