SDF-CAR: 3D Coronary Artery Reconstruction from Two Views with a Hybrid SDF-Occupancy Implicit Representation

Ahmed Reda, Mohamed Ashraf, Mohamed G. Abdelkader, Yasser Abuouf, Muhamed Albadawi

Keywords: 3D Reconstruction, Coronary Artery, Neural Implicit Representations, Signed Distance Fields, X-ray Angiography, Self-Supervised Learning, Sparse-View Reconstruction

Abstract: Three-dimensional (3D) reconstruction of coronary arteries is crucial for accurate diagnosis and treatment planning of cardiovascular diseases. Although Coronary Computed Tomography Angiography (CCTA) can generate 3D models, it involves high radiation doses, is costly, and cannot be used during real-time interventions. In contrast, Invasive Coronary Angiography (ICA), the standard imaging procedure during interventions, typically provides only a few 2D projections, making 3D geometry difficult to interpret. To bridge this gap, we propose a novel self-supervised framework that leverages a Signed Distance Field (SDF)-based neural implicit representation for reconstructing high-quality 3D coronary artery geometry from only two conventional 2D ICA projections. The proposed framework eliminates the need for 3D ground truth or large training datasets. Combining SDF-based geometric priors with an occupancy-based rendering loss, we achieve more stable optimization and higher-fidelity reconstructions than purely occupancy-based methods. Extensive experiments show that the proposed method outperforms the state-of-the-art baseline. It improves topological accuracy (cIDice) by over 16% and significantly reduces surface error. Source code is available at https://github.com/reda1609/SDF-CAR.

Primary Subject Area: Image Acquisition and Reconstruction

Secondary Subject Area: Application: Cardiology

Registration Requirement: Yes

Reproducibility: https://github.com/reda1609/SDF-CAR

Visa & Travel: Yes

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 167