New AI model eliminates ECG for real-time coronary roadmapping during PCI

A team of researchers from Siemens Healthineers and Friedrich-Alexander-Universität Erlangen-Nürnberg has introduced a unified deep learning model for dynamic coronary roadmapping (DRM) that operates without an electrocardiogram (ECG) signal. During percutaneous coronary intervention (PCI), surgeons rely on repeated contrast injections to visualize arteries, which increases radiation exposure and risks contrast-induced nephropathy—affecting up to 30% of patients with renal impairment. Traditional DRM overlays a precomputed angiographic map onto live fluoroscopy, but requires precise cardiac phase matching and catheter tracking, often dependent on ECG data. The new model eliminates this dependency by learning cardiac motion dynamics directly from X-ray sequences.

The framework uses a large-scale spatio-temporal transformer pretrained on 16 million X-ray frames—the first such application for DRM. It simultaneously performs phase matching and catheter tip tracking through two auxiliary tasks: ECG R-peak detection and tip tracking, which remove the need for costly manual mask annotations. A majority-voting postprocessing strategy aggregates temporal predictions, improving robustness and outputting a confidence score correlated with phase-matching error. Evaluated on clinical datasets, the model achieves state-of-the-art performance with low temporal misalignment, making it suitable for real-time guidance without ECG input. The work is detailed in a preprint on arXiv (2607.09805) and represents a step toward safer, more efficient PCI procedures.