4D Vessel Reconstruction for Benchtop Thrombectomy Analysis

This paper presents a low-cost, nine-camera multi-view workflow utilizing 4D Gaussian Splatting to reconstruct time-resolved vessel kinematics and derive comparative stress proxies from benchtop thrombectomy experiments, enabling standardized device testing and condition-to-condition analysis.

The Gist

Imagine you are trying to pull a tangled knot out of a very delicate, stretchy garden hose without popping the hose or kinking it. This is essentially what doctors do during a mechanical thrombectomy, a procedure to remove blood clots from the brain. While it saves lives, the tools used can sometimes stretch or damage the blood vessels, leading to complications.

The problem? Until now, scientists testing these tools on fake models (called "phantoms") could only see if the clot came out, but they couldn't really see how the vessel was twisting, bending, or stretching in 3D space during the process. It was like trying to judge how much you're stretching a rubber band by only looking at a single, blurry photo.

This paper introduces a new, low-cost "super-vision" system that lets researchers watch the blood vessel dance in 4D (3D space + time) with incredible detail. Here is how they did it, explained simply:

1. The "Cage of Eyes" (The Setup)

Instead of using one expensive camera, the team built a cage around their fake brain vessel using nine cheap cameras (the kind found in high-end smartphones). They arranged them in a sphere-like shape, like a dodecahedron (a 12-sided die), all pointing at the center.

• The Analogy: Imagine a group of friends standing in a circle around a dancer, all taking video at the same time. By combining their views, you can build a perfect 3D movie of the dancer's every move.

2. The "Digital Dust" (The Reconstruction)

To make the invisible vessel visible, they sprinkled tiny glowing beads on the fake vessel and shined a special UV light on it. The cameras recorded this.

• The Magic: They used a cutting-edge AI technique called 4D Gaussian Splatting. Think of this as turning the video into a cloud of millions of tiny, glowing "digital dust particles." Unlike a rigid 3D model that might crack when it bends, this cloud of dust flows and stretches naturally, just like a real rubber hose. It captures the vessel's shape frame-by-frame.

3. The "Spiderweb" (The Analysis)

Once they had the moving cloud of dust, they needed to measure how much it was stretching. They turned the cloud into a fixed spiderweb (a graph of connected points).

• The Metaphor: Imagine the vessel is a spiderweb. As the doctor pulls the clot, the web stretches. The researchers didn't just measure how far the whole web moved; they measured how much each individual strand of the web stretched.
• The "Stress Proxy": They calculated a "stress score" for every part of the web. This isn't a perfect physics calculation of pressure inside the wall, but it's a relative score. It tells them: "Hey, this part of the vessel is stretching way more than that part."

4. The "Video Game Test" (Validation)

Before trusting the system on real experiments, they tested it in a computer simulation (like a video game). They created a fake vessel in a program called Blender, stretched it by known amounts (e.g., exactly 3 millimeters), and ran it through their system.

• The Result: The system was incredibly accurate. It knew the vessel moved 3mm, and it reported that it moved almost exactly 3mm. It also correctly showed that when the vessel just slid sideways (without stretching), the "stress score" stayed near zero. This proved the system wasn't just making up numbers.

5. The Real-World Discovery

Finally, they tested two different ways doctors place their suction tubes (catheters) to remove clots:

1. Deep placement: Putting the tube right at the split of the artery.
2. Cervical placement: Putting the tube further back in the neck.

The Finding: The "Deep placement" was gentler on the vessel. The "Cervical placement" caused the vessel to stretch and twist much more, resulting in higher "stress scores." This suggests that where you put the tool matters a lot for safety.

Why This Matters

This paper gives interventional neuroradiologists a tool to replay and analyze thrombectomy procedures on a benchtop model, testing different conditions to see which techniques minimize vessel displacements and stresses, helping them determine the safest approaches before ever touching a patient.

• Before: They knew the clot was gone, but they didn't know if they had over-stretched the vessel during the test.
• Now: They can replay the procedure in high definition, see exactly which parts of the vessel were under the most tension, and adjust their technique based on these findings.

It's a low-cost, open-source way to make these life-saving procedures safer by understanding the invisible forces at play. Note that while real-time, live visualization during surgery is a goal for the future, this current system is designed for offline analysis to refine techniques before they are used on patients.

Technical Summary

1. Problem Statement

Mechanical thrombectomy is a primary treatment for acute ischemic stroke, but it carries risks of vessel deformation, perforation, and hemorrhagic complications. While benchtop models are standard for device testing, current measurement techniques are limited:

• Sparse Data: Existing methods often rely on global metrics (e.g., recanalization rates) or sparse landmark tracking, failing to capture full-field, time-resolved 3D vessel surface kinematics.
• Imaging Constraints: Traditional 3D Digital Image Correlation (3D-DIC) requires stable surface textures or speckle patterns and rigorous calibration, which can be difficult to maintain in dynamic, fluid-filled environments.
• Lack of Comparative Stress Metrics: There is a gap in standardized methods to quantify relative vessel surface stress and deformation patterns across different procedural conditions (e.g., catheter placement strategies).

2. Methodology

The authors developed a low-cost, end-to-end workflow combining multi-view optical acquisition, deep learning segmentation, and 4D Gaussian Splatting (4DGS) reconstruction.

A. Experimental Setup

• Hardware: A custom rig comprising nine Arducam IMX586 cameras (2160p resolution, 20 fps) arranged in a dodecahedron-like PVC frame. Total cost < $1,500.
• Phantom: Silicone models of patient-specific Internal Carotid Artery (ICA) and Middle Cerebral Artery (MCA) vessels.
• Surface Preparation: Red fluorescent microspheres (300 µm) applied to the vessel exterior, illuminated by UV light (400 nm) to enhance feature contrast for segmentation.
• Procedure: A linear actuator performed controlled pullbacks (4 mm/s) to simulate thrombectomy retrieval.

B. Computational Pipeline

1. Acquisition & Calibration: Multi-view videos were captured and calibrated using OpenCV with a chessboard pattern (mean reprojection error ~0.134 pixels).
2. Segmentation: SAM2 (Segment Anything Model 2) was used to segment the vessel from the background in each view, creating masked video datasets.
3. 4D Reconstruction: The masked multi-view data was processed using 4D Gaussian Splatting (4DGS) to generate a time-varying 3D point cloud representing the deforming vessel surface.
4. Graph Construction:
   • Clustering: Points were clustered using K-Means (color-based) followed by DBSCAN (spatial) to create a fixed set of vertices.
   • Topology: A static edge graph was generated via 3D Delaunay triangulation and edge-length pruning. This "fixed-connectivity" graph ensured consistent tracking across time.
   • Curation: Manual curation removed topological artifacts (e.g., extraneous clusters) on the initial frame, which were then locked for all subsequent frames.
5. Metric Extraction:
   • Displacement: Calculated as the centroid movement of clusters relative to the baseline ($t=0$).
   • Stress Proxy: Edge stretch ($\lambda$) was calculated and mapped to a Neo-Hookean uniaxial stress model ($\sigma = \mu(\lambda^2 - \lambda^{-1})$) to generate a relative, surface-based stress metric. This is a comparative proxy, not an absolute wall-stress calculation.
   • Spatial Regularization: A robust Laplacian smoothing step was applied to suppress local tearing artifacts while preserving large-scale motion.

C. Validation Strategy

• Synthetic Ground Truth: A Blender pipeline generated synthetic vessel meshes with known deformations (bulk translation and localized pulling of 1–5 mm). These were rendered with the same camera geometry and processed through the pipeline (bypassing segmentation) to validate geometric and temporal accuracy.
• Benchtop Application: Comparative trials were conducted on silicone phantoms with two aspiration catheter (AC) placement conditions: Cervical ICA vs. ICA Terminus.

3. Key Contributions

• Standardized 4D Workflow: A novel, accessible protocol integrating low-cost multi-view imaging with 4D Gaussian Splatting for dynamic vessel reconstruction.
• Fixed-Connectivity Graph: A method to convert time-varying point clouds into a static edge graph, enabling consistent region-of-interest (ROI) tracking and stress calculation without topological drift.
• Relative Stress Proxy: A biomechanical metric derived from surface geometry that allows for comparative analysis of vessel loading between different procedural strategies.
• Synthetic Validation Framework: A rigorous benchmark using synthetic ground truth to quantify geometric error (Chamfer distance) and temporal consistency before applying the method to physical experiments.

4. Results

A. Synthetic Validation

• Rigid Motion: Under bulk translation, the stress proxy remained near zero (median $\approx$ 0 MPa), confirming the system does not generate spurious stress signals during rigid movement.
• Deformation Fidelity:
  • Geometric Accuracy: Symmetric Chamfer Distance (CD) ranged from 1.714 to 1.815 mm for 1–5 mm pulls.
  • Temporal Consistency: High precision (0.964–0.972 at $\tau=1$ mm) and strong correlation ($R^2 = 0.992$) between reconstructed and ground-truth displacement.
  • Stress Accuracy: Strong agreement in stress proxy values ($R^2 = 0.969$) with minimal bias (0.004 MPa).

B. Benchtop Comparative Application

• Cervical vs. Terminal Placement: The system detected distinct differences between placing the aspiration catheter at the cervical ICA versus the ICA terminus.
  • Displacement: Cervical placement resulted in significantly higher max-median displacement in the distal M1 segment (3.776 mm vs. 1.151 mm) and MCA bifurcation (3.144 mm vs. 0.846 mm).
  • Stress Proxy: Similarly, cervical placement induced higher relative stress (Distal M1: 0.109 MPa vs. 0.058 MPa).
• Interpretation: These results suggest that proximal catheter placement (cervical) induces greater vessel deformation and loading compared to distal placement, a finding consistent with clinical concerns about vascular injury.

5. Significance and Limitations

• Significance:
  • Provides spatially resolved, time-resolved kinematic data for thrombectomy research, moving beyond sparse point measurements.
  • Enables comparative hypothesis testing for device design and procedural strategies (e.g., access routes) without requiring expensive clinical trials or invasive sensors.
  • Offers a low-cost, reproducible framework (under $1,500 hardware) that can be adopted by other labs.
• Limitations:
  • Relative Metric: The stress proxy is a comparative surface metric, not an absolute wall-stress estimate (does not account for wall thickness, lumen pressure, or 3D continuum mechanics).
  • Single-Trial Benchtop: The benchtop results are based on one trial per condition, serving as a demonstration rather than a statistical inference.
  • Synthetic vs. Real: Synthetic validation bypassed the segmentation step; real-world performance depends on the robustness of SAM2 against complex backgrounds and occlusions.
  • Phantom Material: Silicone models do not fully replicate the viscoelastic properties of living human tissue.

Conclusion

This paper presents a robust, standardized framework for 4D vessel reconstruction in thrombectomy studies. By combining 4D Gaussian Splatting with fixed-connectivity graph analysis, the authors successfully quantified vessel deformation and relative stress. The method effectively distinguished between different catheter placement strategies, demonstrating its potential as a vital tool for optimizing device safety and procedural techniques in stroke intervention.