4D Vessel Reconstruction for Benchtop Thrombectomy Analysis

Ethan Nguyen; Javier Carmona; Arisa Matsuzaki; Naoki Kaneko; Katsushi Arisaka

eess.IVApr 8, 2026

4D Vessel Reconstruction for Benchtop Thrombectomy Analysis

Authors: Ethan Nguyen, Javier Carmona, Arisa Matsuzaki, Naoki Kaneko, Katsushi Arisaka

Organizations: UCLA Health · UCLA Physics and Astronomy · Chan Zuckerberg Biohub Network · Ronald Reagan UCLA Medical Center

Abstract

Introduction: Mechanical thrombectomy can cause vessel deformation and procedure-related injury. Benchtop models are widely used for device testing, but time-resolved, full-field 3D vessel-motion measurements remain limited. Methods: We developed a nine-camera, low-cost multi-view workflow for benchtop thrombectomy in silicone middle cerebral artery phantoms (2160p, 20 fps). Multi-view videos were calibrated, segmented, and reconstructed with 4D Gaussian Splatting. Reconstructed point clouds were converted to fixed-connectivity edge graphs for region-of-interest (ROI) displacement tracking and a relative surface-based stress proxy. Stress-proxy values were derived from edge stretch using a Neo-Hookean mapping and reported as comparative surface metrics. A synthetic Blender pipeline with known deformation provided geometric and temporal validation. Results: In synthetic bulk translation, the stress proxy remained near zero for most edges (median $\approx$ 0 MPa; 90th percentile 0.028 MPa), with sparse outliers. In synthetic pulling (1-5 mm), reconstruction showed close geometric and temporal agreement with ground truth, with symmetric Chamfer distance of 1.714-1.815 mm and precision of 0.964-0.972 at $τ= 1$ mm. In preliminary benchtop comparative trials (one trial per condition), cervical aspiration catheter placement showed higher max-median ROI displacement and stress-proxy values than internal carotid artery terminus placement. Conclusion: The proposed protocol provides standardized, time-resolved surface kinematics and comparative relative displacement and stress proxy measurements for thrombectomy benchtop studies. The framework supports condition-to-condition comparisons and methods validation, while remaining distinct from absolute wall-stress estimation. Implementation code and example data are available at https://ethanuser.github.io/vessel4D

Figures & tables

Figure 1: Overview of the experimental and computational pipeline. Multi-camera chessboard calibration yields camera parameters for nine-view vessel recordings. After SAM2-based segmentation, cropping, and dataset preparation, the deforming vessel is reconstructed with 4D Gaussian Splatting and filtered to obtain a time-varying point cloud. The point cloud is converted into a fixed-connectivity edge graph using clustering, DBSCAN, Delaunay graph pruning, and one-time manual curation. Downstream outputs are region-of-interest displacement magnitude (mm) and edge-based stress-proxy magnitude $|\sigma_{e}|$ (MPa). Square boxes denote inputs/outputs and rounded gray boxes denote processing steps.

Figure 2: Experimental setup and major stages of the reconstruction and analysis pipeline. (A) Nine-camera acquisition rig used for multi-view imaging, shown during chessboard-based calibration; long arrows (shown in blue) indicate the cameras’ viewing directions toward the working volume. (B) Example multi-view raw frames acquired during benchtop thrombectomy imaging under UV illumination, showing the fluorescent-beaded silicone vessel phantom. (C) Corresponding segmented vessel frames after preprocessing with SAM2, used to isolate vessel for downstream reconstruction. (D) Example time-varying 3D point-cloud reconstruction obtained from 4D Gaussian Splatting. (E) Fixed set of tracked surface vertices obtained from point-cloud clustering and used for fixed edge-graph construction and temporal tracking. (F) Edge-based stress-proxy magnitude $|\sigma_{e}|$ (MPa) on the resulting edge graph at the representative timepoint.

Figure 3: Synthetic validation workflow (schematic; no single frame/timepoint metric shown). Ground-truth vessel geometries were deformed in Blender using controlled bulk-translation and localized-pulling conditions, rendered with the same camera geometry as the physical setup, and passed through the same camera geometry and downstream reconstruction pipeline, with segmentation bypassed because the renders use transparent backgrounds. The resulting reconstructed point clouds and fixed edge graphs were compared with ground truth using Chamfer distance, temporal consistency metrics, and deformation/stress-proxy agreement analyses.

Figure 4: Regions of the ICA/MCA surface used for regional aggregation and analysis. R1: M1 mid segment; R2: M1/M2 bifurcation; R3: proximal inferior M2; R4: proximal superior M2; R5: ICA terminus/M1 origin.

Figure 5: Rigid-body translation control (synthetic vessel), evaluated at the final frame of the bulk-translation sequence. (A) Displacement magnitude (mm) for ground truth (GT) and reconstruction (Ours). (B) Edge-based stress-proxy magnitude $|\sigma_{e}|$ (MPa); GT remains near zero under rigid motion, and Ours is near zero for most edges with sparse outliers.

Figure 6: Synthetic pulling deformation (1– $5\text{\,}\mathrm{m}\mathrm{m}$ ), evaluated at each condition’s final frame. Rows compare ground truth (GT) and reconstruction (Ours). (A) Displacement magnitude (mm). (B) Edge-based stress-proxy magnitude $|\sigma_{e}|$ (MPa).

Figure 7: GT-versus-reconstruction agreement across synthetic pull magnitudes (condition-level comparison; not a single-frame plot). Pairwise heatmaps report (A) Chamfer distance (CD, mm), (B) temporal Chamfer delta ( $\Delta\mathrm{CD}$ , mm), and (C) F-score; metric definitions follow Sec. 2.2 (Eqs. 12 , 16 , and 19 ). Best agreement occurs along the diagonal (matched pull conditions).

Condition	CD		$\Delta\mathrm{CD}_{\text{rel}}$	Precision	Displacement Error		Stress Bias
	(mm)	$\mathrm{CD}_{\text{norm}}$			(mm)	(%)	(MPa)
Bulk	1.881	3.152	0.085	0.919	0.091	0.64	0.018
1 mm	1.815	3.041	0.139	0.970	0.047	7.17	0.004
2 mm	1.752	2.937	0.100	0.967	0.094	7.82	0.004
3 mm	1.738	2.912	0.114	0.964	0.087	4.98	0.007
4 mm	1.745	2.924	0.100	0.972	0.164	7.38	0.007

Table 1: Synthetic validation metrics by deformation condition. Definitions for CD,

\mathrm{CD}_{\text{norm}}

\Delta\mathrm{CD}_{\text{rel}}

, precision, recall, and F-score follow Sec. 2.2 (Eqs. 12 , 16 , 18 , and 19 ). Displacement error (mm, %) and stress bias (MPa) are reported as Ours

-

GT using the max-median ROI definitions described in Sec. 2.1.11 .

Figure 8: Agreement analyses for max-median ROI metrics between reconstruction and ground truth (condition-level summaries, not single-frame values): correlation and Bland–Altman for displacement magnitude (A–B, mm) and relative surface-based stress-proxy magnitude (C–D, MPa).

Figure 9: Benchtop application demonstration at peak deformation for a single representative trial per AC placement condition (cervical and terminal). (A) Reconstructed displacement magnitude (mm) relative to the initial frame. (B) Corresponding edge-based stress-proxy magnitude $|\sigma_{e}|$ (MPa) at the same peak-deformation timepoint.

Figure 10: Regional max-median displacement and stress-proxy during thrombectomy for a single representative trial per AC placement condition. Bars report the maximum over time of the regional median displacement magnitude (A, mm) and edge-based stress-proxy magnitude (B, MPa) across distal M1 segment, MCA bifurcation (M1/M2), and inferior M2 division.

Condition	CD				$\Delta\mathrm{CD}$		$\tau$ (mm)	P	R	F
Condition	$d_{P\to G}$ (mm)	$d_{G\to P}$ (mm)	Sym (mm)	$\mathrm{CD}_{\text{norm}}$	mm	rel	$\tau$ (mm)	P	R	F
Bulk	0.535	1.345	1.881	3.152	$1.52\times 10^{-2}$	0.085	1	0.919	0.313	0.467
							2	0.983	0.874	0.925
							3	0.998	0.981	0.989
1 mm	0.465	1.349	1.815	3.041	$4.75\times 10^{-4}$	0.139	1	0.970	0.309	0.468
							2	0.998	0.863	0.926

Table S1: Synthetic validation metrics across deformation conditions, including directed and symmetric Chamfer terms, temporal disagreement, and overlap statistics.

Region	Method	Displacement (mm)					Stress (MPa)
Region	Method	1 mm	2 mm	3 mm	4 mm	5 mm	1 mm	2 mm	3 mm	4 mm	5 mm
R1	GT	1.023	2.046	3.068	4.091	5.114	0.005	0.009	0.014	0.018	0.023
R1	Ours	0.999	2.040	3.074	4.116	5.079	0.006	0.010	0.016	0.017	0.022
R2	GT	0.956	1.912	2.868	3.823	4.779	0.007	0.015	0.022	0.029	0.036
R2	Ours	0.922	1.925	2.791	3.731	4.750	0.009	0.015	0.025	0.037	0.034
R3	GT	0.302	0.604	0.906	1.208	1.510	0.010	0.021	0.031	0.042	0.052

Table S2: Regional summary statistics (median) for displacement and stress across sweep settings. Agreement is evaluated primarily via Bland–Altman and correlation in the main manuscript.

Figure S1: Agreement analyses without spatial coherence filtering (SCF) for max-median ROI metrics between reconstruction and ground truth: correlation and Bland–Altman for displacement magnitude (A–B, mm) and relative surface-based stress-proxy magnitude (C–D, MPa).

Figure S2: Qualitative stress-proxy comparison with and without spatial coherence filtering (SCF) at the last frame (maximum deformation) of the 5 mm synthetic pulling condition. GT, reconstruction with SCF, and reconstruction without SCF are shown with a shared color scale. The zoomed view shows more visible stray high-stress points without SCF, whereas SCF suppresses these localized artifacts while preserving the overall stress pattern.

Metric	Condition	Slope	Intercept	$R^{2}$	BA Bias	Lower LoA	Upper LoA
Displacement (mm)	w/ SCF	0.987	-0.112	0.992	-0.142	-0.385	0.100
Displacement (mm)	w/o SCF	0.986	-0.108	0.991	-0.140	-0.388	0.107
Stress proxy (MPa)	w/ SCF	1.068	0.002	0.969	0.004	-0.002	0.010
Stress proxy (MPa)	w/o SCF	1.083	0.008	0.964	0.010	0.003	0.016

Table S3: Agreement summary with and without spatial coherence filtering (SCF) for max-median ROI synthetic validation metrics.

Figure S3: Nine-view chessboard calibration comparison between physical and synthetic setups. (A) Real-camera views from the nine-camera rig using the $6\times 8$ chessboard used to estimate camera extrinsic parameters. (B) Corresponding views from the nine synthetic Blender cameras using the same pattern.

4D Reconstruction

4D reconstruction recovers changing 3D scenes, shapes, and motion over time from images, videos, or other sparse measurements.

See all 139 papers

A Position Statement on Endovascular Models and Effectiveness Metrics for Mechanical Thrombectomy Navigation, on behalf of the Stakeholder Taskforce for AI-assisted Robotic Thrombectomy (START)
Mar 30, 2026Harry Robertshaw, Anna Barnes, Phil Blakelock +20Robotic Endovascular Intervention Surgical Robotics
School of Biomedical Engineering & Imaging Sciences, Kings College London, London, UK · King’s Technology Evaluation Centre (KiTEC), Kings College London, London, UK · Patient & Public Involvement, Kings College London, London, UK +18
An Exploratory Study into using Machine-Learning for Fast Step-by-step Emulation of Numerical Mechanical Thrombectomy Simulations for Ischemic Stroke
May 30, 2026Thijs StessenNeural Surrogate Modeling PDE Surrogate Modeling
University of Amsterdam · MSc Artificial Intelligence
Vesselpose: Vessel Graph Reconstruction from Learned Voxel-wise Direction Vectors in 3D Vascular Images
May 1, 2026Rajalakshmi Palaniappan, Christoph Karg, Nemesio Navarro-Arambula +4Image Segmentation Medical Image Analysis
Max-Delbrueck-Center for Molecular Medicine in the Helmholtz Association (MDC) · Humboldt University of Berlin · University of Potsdam +4

Mar 30, 2026J/K move · Enter open · S save

A Position Statement on Endovascular Models and Effectiveness Metrics for Mechanical Thrombectomy Navigation, on behalf of the Stakeholder Taskforce for AI-assisted Robotic Thrombectomy (START)

Harry Robertshaw, Anna Barnes, Phil Blakelock, Raphael Blanc+19

School of Biomedical Engineering & Imaging Sciences, Kings College London, London, UK · King’s Technology Evaluation Centre (KiTEC), Kings College London, London, UK · Patient & Public Involvement, Kings College London, London, UK+18

Robotic Endovascular Intervention Surgical Robotics Robotics

Open paper AskPDF arXiv

While we are making progress in overcoming infectious diseases and cancer; one of the major medical challenges of the mid-21st century will be the rising prevalence of stroke. Large vessels occlusions are especially debilitating, yet effective treatment (needed within hours to achieve best outcomes) remains limited due to geography. One solution for improving timely access to mechanical thrombectomy in geographically diverse populations is the deployment of robotic surgical systems. Artificial intelligence (AI) assistance may enable the upskilling of operators in this emerging therapeutic delivery approach. Our aim was to establish consensus frameworks for developing and validating AI-assisted robots for thrombectomy. Objectives included standardizing effectiveness metrics and defining reference testbeds across in silico, in vitro, ex vivo, and in vivo environments. To achieve this, we convened experts in neurointervention, robotics, data science, health economics, policy, statistics, and patient advocacy. Consensus was built through an incubator day, a Delphi process, and a final Position Statement. We identified that the four essential testbed environments each had distinct validation roles. Realism requirements vary: simpler testbeds should include realistic vessel anatomy compatible with guidewire and catheter use, while standard testbeds should incorporate deformable vessels. More advanced testbeds should include blood flow, pulsatility, and disease features. There are two macro-classes of effectiveness metrics: one for in silico, in vitro, and ex vivo stages focusing on technical navigation, and another for in vivo stages, focused on clinical outcomes. Patient safety is central to this technology's development. One requisite patient safety task needed now is to correlate in vitro measurements to in vivo complications.

Abstract

4D Reconstruction

Explore similar work

A Position Statement on Endovascular Models and Effectiveness Metrics for Mechanical Thrombectomy Navigation, on behalf of the Stakeholder Taskforce for AI-assisted Robotic Thrombectomy (START)

An Exploratory Study into using Machine-Learning for Fast Step-by-step Emulation of Numerical Mechanical Thrombectomy Simulations for Ischemic Stroke

Vesselpose: Vessel Graph Reconstruction from Learned Voxel-wise Direction Vectors in 3D Vascular Images