Bending-Aware Vision Co-Pilot for Intelligent Robotic Assistance in Endovascular Intervention

Visual feedback obtained from digital subtraction angiography (DSA) imaging is indispensable for intraoperative decision-making during endovascular intervention, because real-time DSA provides critical spatial information about the position and deformation of interventional devices within the vasculature. Excessive bending of the guidewire or microwire tip may increase procedural risk and potentially injure the vessel wall. To address this issue, we propose a vision-based collaborative assistance mechanism (Co-Pilot) for a robotic endovascular intervention system. The proposed framework integrates DSA-based device segmentation, centerline reconstruction, curvature-derived bending-risk estimation, and nonlinear assistive regulation into a closed-loop perception-to-control pipeline. We further implement the method on a modular distributed robotic platform that supports low-latency data exchange and teleoperation in a clinical angiography suite. The system was evaluated in six vascular phantom paths and nine in vivo vessel-navigation trials. Across both settings, enabling the proposed vision co-pilot consistently reduced bending-risk-related metrics, with the average maximum risk value decreasing from 113.0 to 49.0 in phantom paths (56.6%) and from 239.8 to 80.5 in animal trials (66.4%).