angioinsight/single-vessel-flow

Dataset Card for Single-Vessel Coronary Hemodynamics

4,200 synthetic single-vessel coronary geometries, each paired with a steady-state OpenFOAM simulation.

Dataset Details

Dataset Description

Every case provides a 1D centerline (position, radius, curvature, torsion, arc length), the inlet boundary conditions, and a surface point cloud carrying ground-truth pressure and wall shear stress. The dataset was built to benchmark learned surrogates that predict continuous wall fields from a low-dimensional vessel description, but it is a general paired geometry/field dataset and suits operator learning, point-cloud regression, or reduced-order modelling.

• Curated by: Reza Akbarian Bafghi (University of Colorado, Boulder), Sukirt Thakur (AngioInsight, Inc.), Maziar Raissi (University of California, Riverside)
• License: CC BY 4.0

Dataset Sources

• Repository: https://huggingface.co/datasets/angioinsight/single-vessel-flow
• Paper: https://openreview.net/forum?id=AoJUrVjufP (TMLR 2026)
• Earlier version: https://ml4physicalsciences.github.io/2025/files/NeurIPS_ML4PS_2025_305.pdf (ML4PS @ NeurIPS 2025)

Uses

Direct Use

Training and evaluating surrogate models that map vessel geometry and inlet conditions to wall pressure and WSS; benchmarking neural operators and point-cloud regressors; studying geometry encodings for tubular structures.

Out-of-Scope Use

Not for clinical use. The targets are CFD output, not measurements, and nothing here has been validated against patient data. Any clinical application would require prospective validation, uncertainty quantification, and regulatory review.

Dataset Structure

from datasets import load_dataset
import numpy as np

ds = load_dataset("angioinsight/single-vessel-flow", split="test")  # 200 cases, 117 MB
case = ds[0]

surface  = np.array(case["surface_data"])      # (N, 13)
centers  = np.array(case["sampled_centers"])   # (128, 4) = x, y, z, radius
pressure = surface[:, 0]                       # kinematic, mm^2/s^2
xyz      = surface[:, 8:11]                    # mm

Omit split= to fetch all 4,200 cases (~2.5 GB).

Field	Shape	Description
vessel_id	string	vessel_0001 …, unique within a split
surface_data	(N, 13)	Surface point cloud with fields, see below
sampled_centers	(128, 4)	Centerline x, y, z, radius
s, r, kappa, tau	(128, 1)	Arc length, radius, curvature, torsion
p_inlet	float	Absolute inlet pressure
q	float	Inlet volumetric flow rate

N ranges 6,720–15,744 (mean ~12,300); the centerline is always 128 points.

surface_data columns

Idx	Column	Description
0	p	Kinematic pressure at the wall
1	WSS_magnitude	Wall shear stress magnitude
2–4	Normals:0/1/2	Outward unit surface normal
5–7	wallShearStress:0/1/2	Wall shear stress vector
8–10	Points:0/1/2	Surface point coordinates
11	p_1D	Pressure from a 1D Poiseuille model
12	wss_1D	WSS from the same 1D model

WSS_magnitude is exactly the norm of columns 5–7. Columns 11–12 are an analytical low-fidelity reference, not simulation output: a 1D Poiseuille model evaluated per axial station, so they take only ~200 distinct values across a case's thousands of points. Both papers report them as the "low-fidelity" baseline.

Units. Surface and centerline arrays mix millimetres and metres, so check before computing anything geometric.

Quantity	Unit
Points, sampled_centers (all four columns)	mm
r, s	m
kappa, tau	generator-internal, no fixed physical scale
p, p_1D, WSS_magnitude, wallShearStress, wss_1D	kinematic, mm²/s²
p_inlet	Pa (absolute)
q	mm³/s

Pressure and WSS are kinematic (divided by density). To get Pa, multiply by ρ = 1,060 kg/m³ expressed in these units:

p_pa = surface[:, 0] * 1.06e-3        # gauge pressure, Pa
ffr  = (p_pa + p_inlet) / p_inlet     # fractional flow reserve

The pressure field is gauge, referenced to the inlet: p ≈ 0 at the inlet and decreases downstream. p_inlet sets the absolute level for the FFR conversion only and, as expected for incompressible flow, does not affect the gauge field.

r and sampled_centers[:, 3] both describe radius but are not the same quantity up to a unit conversion: their ratio has a median of 1.19 with a 0.99–1.29 spread. sampled_centers[:, 3] is what the models in both papers consume. Likewise kappa carries no constant conversion to physical curvature, so recompute it geometrically if you need physical units. Surface coordinates are in the original anatomical frame, not centered or normalized.

Splits. Fixed, and the same splits used for every result in both papers.

Split	Cases	Download
train	3,600	2.1 GB
validation	400	235 MB
test	200	117 MB

Dataset Creation

Curation Rationale

Paired geometry/hemodynamics data at this scale is not publicly available for coronary arteries. Simulating each case is expensive, so a fixed released benchmark makes surrogate models comparable across papers.

Source Data

Data Collection and Processing

Vessels are grown from atlas seed centerlines by perturbing curvature and torsion and integrating the Frenet–Serret equations, with Gaussian stenoses at random axial positions and the radius profile lofted into a tubular surface. Each geometry is simulated in OpenFOAM 11 with the steady-state incompressible simpleFoam solver. Blood is Newtonian, ρ = 1,060 kg/m³, μ = 4.0 mPa·s; walls are rigid and no-slip; flow is laminar (Re ~ 10²). Pressure and WSS are computed in kinematic form, WSS magnitude via the wallShearStress function object.

Measured across all 4,200 released cases: p_inlet 10,002–15,998 Pa (≈75–120 mmHg), q 2,000–3,000 mm³/s, length 29.6–76.5 mm (median 64.7), stenosis severity 32–74% by area reduction (median 55%), outlet-to-inlet radius ratio 0.78–1.04 (median 0.90).

Who are the source data producers?

The authors. All geometries are synthetically generated and all fields are simulated; no third-party data is redistributed.

Personal and Sensitive Information

None. No patient data, imaging, or identifiers are included. The geometries are synthetic and do not correspond to any individual.

Bias, Risks, and Limitations

None of these are real patient anatomies.

Recommendations

Treat this as a benchmark for simulation surrogacy, not as evidence of clinical accuracy. Models validated only here should not be described as validated for hemodynamic assessment in patients.

Citation

BibTeX:

@article{akbarianbafghi2026centerlines,
  title   = {From Centerlines to Hemodynamics: Anisotropic {RBF} Decoders for Coronary Arteries},
  author  = {Akbarian Bafghi, Reza and Thakur, Sukirt and Raissi, Maziar},
  journal = {Transactions on Machine Learning Research},
  issn    = {2835-8856},
  year    = {2026},
  url     = {https://openreview.net/forum?id=AoJUrVjufP}
}

@inproceedings{akbarianbafghi2025geometry,
  title     = {Geometry-Aware Hemodynamics via a Transformer Encoder and Anisotropic {RBF} Decoder},
  author    = {Akbarian Bafghi, Reza and Thakur, Sukirt and Raissi, Maziar},
  booktitle = {Machine Learning and the Physical Sciences Workshop, NeurIPS 2025},
  year      = {2025},
  url       = {https://ml4physicalsciences.github.io/2025/files/NeurIPS_ML4PS_2025_305.pdf}
}

APA:

Akbarian Bafghi, R., Thakur, S., & Raissi, M. (2026). From centerlines to hemodynamics: Anisotropic RBF decoders for coronary arteries. Transactions on Machine Learning Research. https://openreview.net/forum?id=AoJUrVjufP

Akbarian Bafghi, R., Thakur, S., & Raissi, M. (2025). Geometry-aware hemodynamics via a transformer encoder and anisotropic RBF decoder. Machine Learning and the Physical Sciences Workshop, NeurIPS 2025. https://ml4physicalsciences.github.io/2025/files/NeurIPS_ML4PS_2025_305.pdf

Dataset Card Contact

sukirt@angioinsight.com