Title: VDSB-GWSyn: Leveraging Diffusion Schrödinger Bridges for Anatomically Plausible and Controllable Guidewire Generation in Coronary Angiography

Abstract:

As robot-assisted Percutaneous Coronary Intervention (PCI) gains traction to minimize radiation exposure for operators, the ability to precisely localize coronary guidewire endpoints has become a critical component of computer-assisted PCI. However, the field faces significant hurdles, primarily due to the lack of annotated Coronary Angiography (CAG) images containing guidewires and the insufficient adaptability of current synthesis models. To overcome these limitations, we introduce VDSB-GWSyn, a novel framework built on the Diffusion Schrödinger Bridge (DSB) architecture. This system facilitates the generation of high-fidelity, controllable guidewire samples set against complex anatomical backgrounds.

The VDSB-GWSyn pipeline operates in several stages. Initially, a shape prior algorithm is employed to capture the fundamental geometric properties of guidewires. Subsequently, the model generates guidewire masks that adhere to constraints derived from vessel segmentation masks, thereby producing corresponding endpoint coordinates. Finally, utilizing a SPADE-conditioned DSB, the framework synthesizes realistic guidewire overlays onto actual CAG images.

Our experimental results demonstrate that the guidewire samples produced by VDSB-GWSyn yield strong performance metrics, including favorable ROI-FID and ROI-KID scores, alongside high In-Place Revision (IPR) ratings. Furthermore, leveraging this synthetic data for pre-training, followed by fine-tuning on real data, significantly enhances downstream guidewire endpoint localization. This approach reduced the Mean Position Error (MPE) from 16.01 px to 7.71 px and boosted the Percentage of Correct Keypoints (PCK) at a 3 px threshold from 52.63% to 86.27%, thereby supporting more clinically viable deployments of robot-assisted guidewire delivery systems. Beyond this specific application, the core philosophy of VDSB-GWSyn—emphasizing controllable device synthesis while strictly preserving background integrity and ensuring anatomical feasibility—holds promise for broader transfer to other interventional device perception tasks characterized by data scarcity.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC