CC-GAN: Coronary Artery 3D Reconstruction From Two X-Ray Views With Mamba-Based Topology-Aware Generator and Dual-Domain Discriminator.

Two non-simultaneously acquired X-ray angiographic views from a single X-ray system provide a practical and low-dose input crucial for 3D coronary artery reconstruction. However, this setting poses significant challenges due to severe geometric ambiguity and motion inconsistency. Existing methods commonly rely on backprojection to lift two views into an initial volumetric estimate; however, the resulting backprojection volume exhibits overlooked challenges, including ambiguous vessel topology, severe noise and artifacts, and frequency-domain characteristics that deviate from true coronary anatomy. To address these challenges, we propose Coronary Completor GAN (CC-GAN), a conditional adversarial framework for reconstructing 3D coronary anatomy from two non-simultaneous X-ray projections. First, we introduce Omni-Orientation Mamba (O2-Mamba) into our designed Mamba-based Topology-Aware (MTA) Generator, enabling linear-complexity global modeling to capture the weak topological cues embedded in the backprojection volume. Second, we introduce the Multi-Scale Adaptive Filtering (MSAF) module into our designed MTA-Generator, an attention-based skip-fusion mechanism that suppresses artifact- and noise-related features. Third, we design a Dual-Domain Discriminator (D2-Discriminator) that jointly analyzes spatial and frequency representations, enhancing the detection of frequency patterns and promoting more realistic reconstructions. Extensive experiments on both synthetic and real clinical datasets demonstrate that CC-GAN consistently outperforms existing state-of-the-art methods. Our results suggest the clinical potential of reconstructing 3D coronary vasculature using only two non-simultaneously acquired projections from a single X-ray system.