Organizations: Department of Control Science and Engineering, College of Electronics and Information Engineering, and Shanghai Institute of Intelligent Science and Technology, Tongji University, Shanghai 200092, China · Department of Electronic Engineering, Faculty of Engineering, The Chinese University of Hong Kong, Hong Kong SAR 999077, China · Shanghai Operation Robot Co., Ltd., Shanghai 201318, China · Department of Cardiology, Zhongshan Hospital, Fudan University, Shanghai Institute of Cardiovascular Diseases, National Clinical Research Center for Interventional Medicine, Shanghai 200032, China · State Key Laboratory of Cardiovascular Diseases and Medical Innovation Center, Shanghai East Hospital, School of Medicine, Tongji University 200092, Shanghai, China
Robotic-assisted percutaneous coronary intervention (PCI) is constrained by the inherent limitations of 2D Digital Subtraction Angiography (DSA). Unlike physicians, who can directly manipulate guidewires and integrate tactile feedback with their prior anatomical knowledge, teleoperated robotic systems must rely solely on 2D projections. This mode of operation, simultaneously lacking spatial context and tactile sensation, may give rise to projection-induced ambiguities at vascular bifurcations. To address this challenge, we propose a two-stage framework (SCAR-UNet-GAT) for real-time robotic path planning. In the first stage, SCAR-UNet, a spatial-coordinate-attention-regularized U-Net, is employed for accurate coronary vessel segmentation. The integration of multi-level attention mechanisms enhances the delineation of thin, tortuous vessels and improves robustness against imaging noise. From the resulting binary masks, vessel centerlines and bifurcation points are extracted, and geometric descriptors (e.g., branch diameter, intersection angles) are fused with local DSA patches to construct node features. In the second stage, a Graph Attention Network (GAT) reasons over the vessel graph to identify anatomically consistent and clinically feasible trajectories, effectively distinguishing true bifurcations from projection-induced false crossings. On a clinical DSA dataset, SCAR-UNet achieved a Dice coefficient of 93.1%. For path disambiguation, the proposed GAT-based method attained a success rate of 95.0% and a target-arrival success rate of 90.0%, substantially outperforming conventional shortest-path planning (60.0% and 55.0%) and heuristic-based planning (75.0% and 70.0%). Validation on a robotic platform further confirmed the practical feasibility and robustness of the proposed framework.
Figures & tables
Fig. 1: The robotic intervention procedure relies on real-time 2D DSA images acquired by a C-arm X-ray system for path planning, providing precise navigation guidance for the robotic operation. (a) The intraoperative setup shows the endovascular robot operating under DSA guidance. (b) An apparent crossing point appears in the 2D projection when vessels at different depths visually overlap, which may be misinterpreted as a true bifurcation. (c) Correct path planning avoids such projection-induced crossings, while (d) invalid planning erroneously treats them as real intersections, potentially leading to navigation errors. The schematic was created using BioRender ( https://biorender.com ).
Fig. 2: Overview of the proposed SCAR-UNet-GAT framework for robotic vascular path planning. (a) The pipeline combines SCAR-UNet-based segmentation of DSA image sequences with topology-aware GAT-based path planning, and is deployed on a robotic interventional system to provide navigation guidance for intervention procedures. (b) The SCAR-UNet segmentation module uses an encoder-decoder architecture with spatial-coordinate attention, integrated with multiple attention mechanisms and ASPP, enabling accurate extraction of vascular masks from DSA images. (c) The topology-aware path planning module extracts vessel centerlines and crossing features from segmentation results, fuses image patch features at intersections, constructs a vessel graph, predicts path probabilities using GAT reasoning, and then undergoes post-processing to determine the planned route.
Fig. 3: Validation of Robotic Vascular Path Planning on Clinical Data and Robotic Systems. (a) Results on clinical DSA images, illustrating the details of apparent crossing points and the visual comparison among three methods: shortest-path, heuristic-based, and the proposed approach. (b) The robotic intervention platform, consisting of a robotic arm, gripper, monocular camera, pulse pump, and vascular phantom. The monocular camera provides DSA-like vascular images as inputs for the path planning system. (c) Results on vascular phantom images, further demonstrating the disambiguation of apparent crossing points and the comparative visualization of the three methods. (d) Validation on robotic systems, where YOLOv5 was employed for real-time guidewire tip detection and PID-based control was used to regulate advancement and rotation of the gripper, enabling the guidewire to follow the planned trajectory and successfully reach the target point.
Method
Dice (%)
Precision (%)
Recall (%)
FPS
U-Net
89.4
87.7
90.9
68
Attention U-Net
90.5
89.0
92.0
61
DeepLabv3+
90.8
89.1
92.0
62
TransUNet
91.6
90.8
92.5
49
Swin-UNet
92.0
91.2
93.1
45
SCAR-UNet
93.1
92.0
94.5
58
TABLE I: Comparative Analysis of Vessel Segmentation Performance
Model Variant
Dice
Precision
Recall
FPS
w/o CoordAttention
91.7
90.4
92.8
60
w/o SimAM
92.0
90.7
93.2
60
w/o SE
92.2
91.0
93.3
59
w/o ASPP
92.5
91.4
93.6
59
w/o Hybrid Loss (Dice only)
92.4
91.2
93.4
58
Full SCAR-UNet
93.1
92.0
94.5
58
TABLE II: Ablation study of SCAR-UNet on the coronary DSA dataset. Dice, Precision, and Recall are reported in %; FPS measures inference speed.
Method
Disambiguation Success Rate (Successful Cases / Total Cases)
Target Arrival Rate (Successful Cases / Total Cases)
Shortest Path (Dijkstra on vessel skeleton)
12 / 20 (60.0%)
11 / 20 (55.0%)
Heuristic-based (Topology features + heuristic rules)
15 / 20 (75.0%)
14 / 20 (70.0%)
Proposed Method (Deep learning + GAT reasoning)
19 / 20 (95.0%)
18 / 20 (90.0%)
TABLE III: Comparison of Apparent Crossing Disambiguation and Target Arrival on DSA Images
School of Mathematical Sciences, Ocean University of China, Qingdao, 266100, Shandong, China. · Department of Cardiology, Qilu Hospital (Qingdao), Cheeloo College of Medicine, Shandong University, No. 758 Hefei Road, Qingdao, 266000, Shandong, China.