AI-Driven Robotic PCI: A Perception-to-Action Framework for Coronary Guidewire Control

Robotic percutaneous coronary intervention (PCI) remains predominantly teleoperated, while the most demanding part of the procedure, guidewire navigation through a moving coronary tree under fluoroscopy, still depends on the continuous human interpretation of vessel anatomy, cardiac phase, guidewire position, and target location. We present a preclinical perception-to-action framework for AI-driven robotic PCI that integrates fluoroscopic perception, dynamic coronary vessel memory, vessel-coordinate state estimation, and robot-executable guidewire command generation on a robotic PCI platform. During contrast angiography, phase-indexed vessel-catheter templates and a dynamic vessel-coordinate coronary map are generated. During guidewire manipulation without contrast injection, live fluoroscopy is segmented into catheter and guidewire structures, cardiac phase is estimated by overlap between the live catheter-guidewire skeleton and stored vessel-catheter templates, the guidewire is assigned to the most likely vessel branch, and the distal tip is projected to a vessel-coordinate target representation for robotic action generation. The segmentation dataset contained 6269/1567/957 vessel images, 3857/964/537 guidewire images, and 4796/1199/667 catheter images for training/validation/test splits, respectively. Test-set Dice scores were 90.7% for vessels, 92.6% for catheters, and 89.0% for guidewires. In 623 real-time fluoroscopy frames from physician-supervised animal experiments, phase selection accuracy was 589/623 (94.6%; 95% CI, 92.5-96.1%) and vessel assignment accuracy was 575/623 (92.3%; 95% CI, 89.9-94.1%). Across 34 scenario-level episodes and their associated command-level decisions, correct-command rates ranged from 83.6% to 95.6% across target navigation and safety scenarios. These results provide preclinical evidence that live fluoroscopic perception can be converted into robot-executable coronary guidewire actions within an integrated AI-robotic PCI workflow. The study was designed to establish early system feasibility rather than to prove clinical superiority, large-scale generalization, or comparative advantages over manual or teleoperated robotic PCI.