A 2D Image Feedback-Based Autonomous Endovascular Navigation Method: A Perception and Control Framework

Sheng Cao; Shuxiang Guo; Yonggan Yan; Jian Guo; Chunying Li; Bin Wang; Mingchao Ding; Yongxin Zhang

cs.ROOct 15, 2026

A 2D Image Feedback-Based Autonomous Endovascular Navigation Method: A Perception and Control Framework

Authors: Sheng Cao, Shuxiang Guo, Yonggan Yan, Jian Guo, Chunying Li, Bin Wang, Mingchao Ding, Yongxin Zhang,

Organizations: Beijing Institute of Technology, Beijing 100081, China · Advanced Institute for Ocean Research, Southern University of Science and Technology, Shenzhen, 518055, China · Shenzhen Institute of Advanced Biomedical Robot Company Limited, Shenzhen 518000, Guangdong Province, China · Department of Electronic and Electrical Engineering, Southern University of Science and Technology, Shenzhen, Guangdong 518055, China · School of Life Science and the Key Laboratory of Convergence Medical Engineering System and Healthcare Technology, Ministry of Industry and Information Technology, Beijing Institute of Technology, Beijing 100081, China · School of Mechanical, Electronic and Control Engineering, Beijing Jiaotong University, Beijing, China · Southern University of Science and Technology, Shenzhen, Guangdong 518055, China · Aerospace Center Hospital and the Key Laboratory of Convergence Medical Engineering System and Healthcare Technology, Beijing Institute of Technology, Beijing 100081, China · Neurovascular Center, Changhai Hospital, Naval Medical University, Shanghai, China

Abstract

In endovascular interventions, physicians rely on planar DSA and X-ray fluoroscopy to interpret vessel anatomy and device motion, yet converting these images into feedback for autonomous navigation remains difficult. A perception and control method is formulated for two-dimensional image feedback. Vessel masks are converted into a labeled centerline graph, and device-tip estimates are projected onto a planned path to obtain progress, deviation, and direction. A rule-based controller uses these quantities for delivery, branch selection, probing, rollback, and recovery. The method was evaluated on a physical dual-device robotic platform in 65 closed-loop runs using simulator-rendered planar images across three paths. All runs reached at least 0.95 maximum observed guidewire progress; two runs (3.1%) met the prespecified failure rule of three or more autonomous stop events. The results establish route-level performance on the tested platform; clinical image sequences and broader anatomy remain to be evaluated.

Figures & tables

Fig. 1: Overall system and human-in-the-loop workflow. Mentice provides the planar image stream, the physical dual-device robot executes autonomous commands, and the physician supervises and can override the process.

Fig. 2: System architecture and data flow. Offline, custom device and vessel SE U-Net models are trained, exported to ONNX, and converted to TensorRT engines. Online, device masks provide frame-wise tip and direction estimates, whereas the averaged vessel output is locked as a labeled centerline graph. Operator-selected start and target points define the topology-constrained route. Operator-marked interaction-risk points form a separate early-trigger channel: proximity invokes priority retreat before the delayed off-path condition, without modifying the graph or replanning the route. This heuristic response is not a formal safety guarantee. Projection of both tips onto the route produces the visual state used by the priority controller. A discrete command drives the robot and changes the physical device configuration. Mentice updates the device–vessel interaction and renders the realized motion in the next simulator-rendered image, which closes the visual feedback loop.

Metric	Device	Vessel
Train/validation/test images	102/22/22	117/15/14
Output channels	2	11
Training crop	$512\times 512$	$512\times 512$
Batch size	1	1
Training epochs	60	60
Mean epoch time (s)	28.6	8.6

TABLE I: Segmentation training summary

Task	Backend	Agreement Dice	Patch (ms)	Full frame (ms)	Size (MB)
Device	PyTorch FP32	reference	30.6	70.0	–
Device	TensorRT FP16	0.9999/0.9997	8.1	32.8	60.1
Device	TensorRT INT8	0.9981/0.9854	4.2	17.1	32.1
Vessel	PyTorch FP32	reference	30.8	72.0	–
Vessel	TensorRT FP16	0.9998	8.3	33.5	59.8
Vessel	TensorRT INT8	0.9812/0.9951	4.2	17.9	32.4

TABLE II: TensorRT benchmark

Fig. 3: Capability boundary and evaluated navigation paths. (a) Manual guidewire-only tests define the conditions that can be completed before autonomous testing; the 30 ∘ and J guidewires do not complete Path 3 under the single-guidewire constraint, whereas the corresponding coordinated guidewire–catheter task remains reachable. (b) Simulator-rendered scenes for Paths 1–3 show the operator-selected start and target points and the planned vessel routes. Paths 1 and 2 include five guidewire conditions, while Path 3 includes only the three conditions that passed the prespecified capability check.

Fig. 4: Closed-loop experimental platform. The platform combines the Mentice VIST G7 simulator, physical dual-device intervention robot, controller interfaces, display terminal, and autonomous-navigation workstation. It links device actuation, planar image generation, visual-state estimation, and autonomous commands in one supervised loop; the present dataset contains simulator-rendered projections rather than clinical DSA acquisitions.

Fig. 5: Representative autonomous navigation sequences across the three paths. Each row shows one recorded physical-robot run and follows the same four stages: initial state, autonomous advance, branch probing or recovery, and catheter-follow action. The values below each frame are taken from the synchronized action and state records.

Path	$n$	Stops/run	Failed runs	Max. wire progress	Cath.-follow trigger	Full duration (s)	Actions	Forward fraction (%)
Path 1	25	$0.00\pm 0.00$	0/25	$1.000\pm 0.000$	25/25 (100%)	$111.7\pm 37.7$	$79.9\pm 33.4$	$76.0\pm 11.6$
Path 2	25	$0.56\pm 0.87$	2/25	$1.000\pm 0.000$	25/25 (100%)	$162.6\pm 75.4$	$112.1\pm 64.3$	$71.2\pm 10.6$
Path 3	15	$0.07\pm 0.26$	0/15	$1.000\pm 0.000$	15/15 (100%)	$148.2\pm 34.3$	$118.8\pm 28.3$	$66.2\pm 5.0$

TABLE III: Run-level outcomes.

Path	Wire IoU	Catheter IoU	Vision time (ms)
Path 1	$0.890\pm 0.013$	$0.949\pm 0.006$	$70.5\pm 0.9$
Path 2	$0.882\pm 0.016$	$0.943\pm 0.012$	$70.2\pm 0.8$
Path 3	$0.897\pm 0.013$	$0.947\pm 0.010$	$71.1\pm 0.7$

TABLE IV: Imaging and timing.

Fig. 6: Autonomous stop counts for individual runs. The dashed line marks the prespecified three-stop failure threshold; crosses identify the two failed runs.

Fig. 7: Mean per-run composition of guidewire advance, catheter advance, backward movement, and turning or rotation across the three paths.

Fig. 8: Complete trial matrices. Panels A–C show full run duration and panels D–F show action count; each cell is one logged repeat. Asterisks mark the two stop-based failures and N/A denotes the excluded Path 3 conditions.

Path	Guidewire	$n$	Duration (s)	Actions	Forward (%)	Total stops; failed runs
1	30 ∘	5	$86.1\pm 16.0$	$56.8\pm 8.3$	$87.3\pm 3.2$	0; 0
1	45 ∘	5	$105.3\pm 33.7$	$76.6\pm 30.5$	$75.6\pm 8.6$	0; 0
1	60 ∘	5	$134.0\pm 44.3$	$101.6\pm 41.2$	$66.7\pm 11.4$	0; 0
1	90 ∘	5	$98.5\pm 31.0$	$70.4\pm 23.0$	$77.5\pm 9.4$	0; 0
1	J	5	$134.6\pm 43.0$	$94.2\pm 42.7$	$73.1\pm 14.9$	0; 0
2	30 ∘	5	$217.5\pm 125.6$	$160.8\pm 108.0$	$66.1\pm 8.3$	3; 0

TABLE V: Condition-level summary.

Robotic Endovascular Intervention

Robotic endovascular intervention uses robotic systems to navigate vascular instruments and perform procedures with greater precision, control, or autonomy.

See all 24 papers

Vision-Based Reasoning with Topology-Encoded Graphs for Anatomical Path Disambiguation in Robot-Assisted Endovascular Navigation
Feb 23, 2026Jiyuan Zhao, Zhengyu Shi, Wentong Tian +5Robotic Endovascular Intervention Coronary Angiography
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 +2
Contrast-Free Autonomous Navigation of Untethered Endovascular Microrobots Using Single-Plane Fluoroscopy
Aug 31, 2026Husnu Halid Alabay, Tuan-Anh Le, Ping Wang +1Robotic Endovascular Intervention Digital Twins
Department of Physiology and Biomedical Engineering, Mayo Clinic, Scottsdale, Arizona, USA · School of Biological and Health Systems Engineering, Arizona State University, Tempe, AZ, USA
Toward Safe Autonomous Robotic Endovascular Interventions using World Models
Apr 22, 2026Harry Robertshaw, Nikola Fischer, Han-Ru Wu +6World Models for Robotics Robotic Endovascular Intervention
Surgical & Interventional Engineering, School of Biomedical Engineering & Imaging Sciences, King’s College London, UK · Department of Radiology, National Taiwan University Hospital, Taiwan · Department of Neuroradiology, King’s College Hospital, UK

Feb 23, 2026J/K move · Enter open · S save

Vision-Based Reasoning with Topology-Encoded Graphs for Anatomical Path Disambiguation in Robot-Assisted Endovascular Navigation

Jiyuan Zhao, Zhengyu Shi, Wentong Tian, Tianliang Yao+4

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+2

Robotic Endovascular Intervention Coronary Angiography Coronary Artery Segmentation

Open paper AskPDF arXiv

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.

Abstract

Robotic Endovascular Intervention

Explore similar work

Vision-Based Reasoning with Topology-Encoded Graphs for Anatomical Path Disambiguation in Robot-Assisted Endovascular Navigation

Contrast-Free Autonomous Navigation of Untethered Endovascular Microrobots Using Single-Plane Fluoroscopy

Toward Safe Autonomous Robotic Endovascular Interventions using World Models