From a few segmented projections to a branch-structured coronary artery — with smooth centrelines and detailed radius profiles.
Two projections of the same artery
3D reconstruction
CT-derived simulated projections and their predicted 3D anatomy. Colours in the reconstruction indicate local vessel radius.
Sparse-view 3D coronary reconstruction commonly relies on cross-view correspondence and triangulation, which are vulnerable to vessel overlap and foreshortening, or on volumetric prediction followed by vascular-graph extraction, which does not directly provide centrelines and radii. We introduce PPCAR-Net, a projection-refined parametric coronary artery reconstruction network that directly predicts a branch-structured centreline-and-radius representation, termed vessel code, without explicit point matching, triangulation, or an intermediate volume.
From a variable number of segmented views, a coarse predictor combines frozen VGGT features with learned branch queries to estimate active branches, their B-spline centreline trajectories, and approximate dense radius profiles. Projection-guided geometry and radius refiners sample input view evidence and predict residual corrections learned with 3D supervision. We evaluate representation fidelity and sparse-view 3D reconstruction quantitatively and qualitatively. On simulated angiographic masks generated from CT-derived coronary anatomy, PPCAR-Net produces better connected artery reconstructions and achieves strong centreline accuracy, particularly for RCA, while maintaining competitive volumetric overlap. Coarse-to-fine inference takes 121 ms, enabling real-time reconstruction.
Sparse angiographic views can hide depth, overlap branches and shorten their apparent length. PPCAR-Net combines anatomical patterns learned across training cases with evidence from the input projections to recover a plausible 3D artery.
PPCAR-Net directly predicts vessel code, our explicit artery representation, which comprises branch presence, B-spline control points and dense radii. B-spline control points define each branch’s centreline, which we sample densely, with a radius associated with each sampled point. Learned refinement then adjusts the geometry and radius profiles using local projection evidence.
PPCAR-Net uses a coarse-to-fine feed-forward pipeline to reconstruct an explicit 3D artery from calibrated coronary masks.
A frozen VGGT backbone extracts features from the input views. Learned branch queries combine these features with coronary anatomical patterns learned during training to predict a coarse vessel code: branch existence, the overall shape of each branch through B-spline centreline curves, and dense radius profiles.
We project the coarse prediction back into the input views and sample the corresponding local 2D evidence around the projected points. The refiners use this evidence to apply learned residual corrections to both the centreline geometry and the radii.
Choose an artery and a case. Compare the CT-derived 3D label with the PPCAR-Net prediction from two segmented views.
Radius colours use the scale shown for each case. Rotations are pre-rendered visualisations; this page does not run model inference.
Browse the complete comparison gallery, including 1-, 2- and 4-view results ↗The RCA and LCA examples use coronary CT angiography annotations from ImageCAS (Zeng et al.). The input masks are simulated projections of the annotated 3D arteries.
Recovering narrowing from 2D evidence
Two simulated projections of a diseased coronary artery, its 3D ground truth and our reconstruction.
The ground-truth surface is shown in red; the prediction is coloured by local radius. The fixed views below highlight the local narrowing.
Fixed views with the selected region circled and enlarged below.
A qualitative example of recovering visible radius reduction, rather than an evaluation of clinical stenosis detection.
This diseased coronary CT anatomy is from the ASOCA dataset (Gharleghi et al.), which provides CT coronary angiography images and artery annotations.
Exploratory reconstruction from real angiographic images, using segmented artery masks as input.
Acquired angiography and selected input frames are shown with their artery masks above; prediction overlays and the 3D reconstruction are shown below.
Radius-coloured prediction from the two views. Paired 3D ground truth is unavailable.
These real X-ray examples lack paired 3D ground truth for accuracy assessment and do not constitute systematic clinical validation.
The real X-ray angiography example is from the data released in the AutoCAR paper repository: Zhu et al., Sparse and transferable three-dimensional dynamic vascular reconstruction for instantaneous diagnosis (2025). Reconstructions shown here are produced by PPCAR-Net.
If you use this work, please cite:
@misc{ren2026ppcarnet,
title={{PPCAR-Net}: Projection-Refined Parametric 3D Coronary Artery Reconstruction from Sparse X-ray Angiographic Views},
author={Ren, Yu and Lee, Hwee Kuan and Cham, Tat-Jen and Yap, Jonathan and Yeo, Khung Keong},
year={2026},
eprint={2610.09383},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2610.09383}
}