Vision-Based Reasoning with Topology-Encoded Graphs for Anatomical Path Disambiguation in Robot-Assisted Endovascular Navigation
New AI framework uses Graph Attention Networks to solve critical 2D-to-3D ambiguity in robotic heart surgery.
A research team from multiple institutions has published a breakthrough paper, "Vision-Based Reasoning with Topology-Encoded Graphs for Anatomical Path Disambiguation in Robot-Assisted Endovascular Navigation," accepted by IEEE ICRA 2026. The work addresses a fundamental limitation in robotic-assisted percutaneous coronary intervention (PCI), where surgeons teleoperating robots must navigate complex 3D vascular networks using only 2D Digital Subtraction Angiography (DSA) images. This lack of spatial context and tactile feedback creates dangerous ambiguities at vessel bifurcations, where a robot might mistake overlapping 2D projections for a single path. The team's proposed SCAR-UNet-GAT framework directly tackles this by providing real-time, AI-powered 3D reasoning to guide robotic catheters.
The technical solution is a two-stage pipeline. First, a Spatial-Coordinate-Attention-Regularized U-Net (SCAR-UNet) segments coronary vessels from noisy DSA images with a 93.1% Dice coefficient, excelling at thin vessel delineation. Second, and most critically, a Graph Attention Network (GAT) reasons over a graph constructed from the segmented vessels. This graph encodes topology and geometric features like branch diameters and intersection angles. The GAT disambiguates true 3D paths from false 2D projections, achieving a 95.0% path disambiguation success rate and a 90.0% target-arrival success rate in validation—dramatically outperforming shortest-path (60%) and heuristic-based (75%) planners. This validated framework represents a significant step toward fully autonomous, reliable robotic surgery by closing the perception-reasoning gap caused by 2D imaging.
- SCAR-UNet achieves 93.1% Dice coefficient for segmenting thin coronary vessels from 2D DSA images.
- The Graph Attention Network (GAT) stage boosts path disambiguation success to 95%, a 35-point jump over conventional shortest-path planning.
- The full system validated on a robotic platform enables safer navigation by resolving critical 2D-to-3D projection ambiguities at vascular branches.
Why It Matters
This AI-driven 3D reasoning could make complex robotic heart surgeries significantly safer and more accessible by reducing surgeon cognitive load and error risk.