Synthetic Angiography AI beats data scarcity for 3D coronary reconstruction

Accurate correspondence matching across multiple angiographic views is the foundation of 3D coronary reconstruction and interventional guidance, but developing robust deep learning models has been limited by a fundamental data bottleneck. Obtaining ground truth for matching in angiography pairs is prohibitively expensive and difficult to scale. To overcome this barrier, researchers In Kyu Lee, Sumin Seo, and Jaesik Min from South Korea introduce a physically-grounded data generation framework that synthesizes high-fidelity Digital Reconstructed Radiographs (DRRs) from 3D Coronary CT Angiography (CCTA) volumes. Their framework generates dense, highly accurate 3D-to-2D projection labels by simulating realistic C-arm acquisition geometry on patient anatomy, producing paired images with ground truth correspondence at zero human cost.

Leveraging this dense supervision, the team proposes a Geometry-Informed Matching Module (GIMM) that integrates global features and anatomical structure into correspondence learning. Unlike real angiography, where evaluation relies on subjective human annotation, their dataset enables human-free evaluation. The approach, accepted at MICCAI 2026, demonstrates improvements over other matching baselines on the CT-derived DRR dataset. This work has significant implications for reducing the annotation bottleneck in medical imaging AI, enabling faster development of 3D reconstruction tools for coronary artery disease diagnosis and treatment planning. The code and dataset are publicly available.