Comparative Study of Ultrasound Shape Completion and CBCT-Based AR Workflows for Spinal Needle Interventions

This study compares AR-guided spinal needle intervention workflows using ultrasound shape completion versus CBCT, finding that while both are viable, the CBCT-based approach offers superior efficiency, precision, and user trust, whereas the ultrasound method provides a radiation-free alternative with limitations in deep anatomical reconstruction, suggesting a hybrid workflow as an optimal solution.

The Gist

Imagine you are a surgeon trying to perform a delicate procedure on a patient's spine, like inserting a needle to relieve pain or collect fluid. The spine is buried deep inside the body, hidden by layers of muscle and skin. To do this safely, you need a "GPS" to see exactly where the bones are and guide your needle without hitting anything important.

This paper compares two different types of "GPS" systems for this job, both using Augmented Reality (AR)—think of it as wearing smart glasses that project a 3D map directly onto your view of the patient.

The researchers wanted to know: Which GPS is better?

1. The "X-Ray" GPS (CBCT): Uses a special 3D scanner that takes a snapshot of the spine before you start. It's like taking a high-resolution photo of the terrain before a hike.
2. The "Sonar" GPS (Ultrasound): Uses sound waves to "see" the spine in real-time, but since sound waves bounce off bones weirdly, the image is often incomplete. To fix this, the computer uses AI to "guess" and fill in the missing parts of the bone, kind of like an artist finishing a sketch based on a few rough lines.

The Experiment: A Training Simulation

The researchers didn't use real patients. Instead, they built a realistic "fake spine" (a plastic model soaked in gelatin to feel like real tissue) and put it in a water tank. They recruited 20 people (mostly engineers and researchers, not yet expert surgeons) to act as the doctors.

Each person had to use one of the two GPS systems to:

1. Plan the route: Decide where to scan and where to aim.
2. Drive the needle: Actually insert a needle into the fake spine to hit two specific targets: a "facet joint" (a small hinge in the back) and a "lumbar puncture" (a deeper spot in the middle).

What They Found: The Race Results

1. The Planning Phase: The "X-Ray" GPS was faster.

• CBCT (X-Ray): Participants planned their route much faster. It was like looking at a complete, detailed map of a city. They knew exactly where the streets were.
• Ultrasound (Sonar): Planning took longer. It was like trying to navigate a city using a map that only shows half the streets, forcing the driver to guess where the missing roads might be.
• The Trade-off: The Ultrasound system is radiation-free (safe for the patient), while the X-Ray system uses a tiny bit of radiation (though still very low and safe).

2. The Needle Insertion Phase: The "X-Ray" GPS was more precise.

• Accuracy: When it came to hitting the target, the X-Ray group was much more accurate, especially for the deep, tricky target (lumbar puncture). Their errors were small.
• Confidence: The X-Ray users felt more confident and trusted their glasses more. They felt like the map was "real."
• Ultrasound Struggles: The Ultrasound group did okay for the shallow target (facet joint), but struggled with the deep target. Because the AI had to "guess" the shape of the deep bones, the map was slightly off, leading to bigger errors.
• Workload: The Ultrasound users felt more stressed and mentally tired. They had to work harder to interpret the "guess-work" map.

The Big Picture: Two Tools for Two Jobs

The study concludes that both systems work, but they have different superpowers:

• The CBCT (X-Ray) System is the Master Planner. It gives you a perfect, complete picture of the whole spine. It's great for getting your bearings, planning the route, and feeling confident. However, it's static (the picture doesn't change if the patient moves) and uses a tiny bit of radiation.
• The Ultrasound System is the Adaptive Scout. It's radiation-free and can update in real-time if the patient moves. However, because it relies on AI to "fill in the blanks," it can be less accurate in deep or complex areas.

The "Hybrid" Solution

The authors suggest the best future solution is a Hybrid GPS:
Imagine using the X-Ray map to plan the whole journey and get a clear view of the terrain, but then switching to the Ultrasound Scout during the actual drive to make real-time adjustments if the road shifts. This would give you the best of both worlds: the precision of a complete map with the safety and adaptability of real-time sound waves.

In short: If you need a perfect map to start, use the X-Ray. If you need to avoid radiation and adapt to movement, use the Ultrasound. But for the safest, most precise surgery, combining them is the ultimate goal.

Technical Summary

Here is a detailed technical summary of the paper "Comparative Study of Ultrasound Shape Completion and CBCT-Based AR Workflows for Spinal Needle Interventions."

1. Problem Statement

Accurate spinal interventions (e.g., facet joint injections, lumbar punctures) require precise anatomical localization. Current guidance methods face significant trade-offs:

• Fluoroscopy/CT: High accuracy but involve ionizing radiation and lack real-time adaptability to patient motion.
• Cone-Beam CT (CBCT): Offers 3D intraoperative imaging at lower doses but remains a static modality, insensitive to intraoperative deformation.
• Ultrasound (US): Radiation-free and real-time but suffers from acoustic shadowing and limited bone penetration, providing only partial 2D views that require clinicians to infer 3D structures.

While recent advances in ultrasound shape completion (using deep learning to reconstruct full vertebrae from partial US data) and Augmented Reality (AR) offer promising solutions, there is a lack of systematic, quantitative comparison between US-based shape completion workflows and CBCT-based workflows within a unified AR-guided framework.

2. Methodology

The study implemented a between-subject user study comparing two distinct AR-guided pipelines for spinal needle intervention on a phantom model.

System Architecture

Both workflows utilized a unified AR framework visualized via a Microsoft HoloLens 2 headset, integrating with:

• Robotic US: A KUKA LBR iiwa arm with a Siemens ACUSON Juniper ultrasound machine.
• Robotic CBCT: A Brainlab Loop-X system.
• Registration: A virtual-to-real registration process using point clouds captured by the HoloLens depth sensor and Iterative Closest Point (ICP) optimization to align the physical devices with the AR coordinate system.

Workflow A: Ultrasound-Based (US-AR)

1. Planning: Users defined a zig-zag robotic scanning trajectory in AR.
2. Acquisition: The robot performed an automated scan; US frames were synchronized with robot poses.
3. Reconstruction: A probabilistic deep-learning framework (Shape Completion Network) processed the partial US point clouds. It used learned shape priors from CT data to infer complete vertebral geometries, generating 3D meshes via Poisson surface reconstruction.
4. Visualization: The completed 3D model was overlaid in AR for trajectory planning.

Workflow B: CBCT-Based (CBCT-AR)

1. Planning: Users defined a 3D bounding box in AR to specify the scan volume.
2. Acquisition: The CBCT machine automatically adjusted its geometry and acquired a 3D volume.
3. Reconstruction: Bone segmentation was performed via intensity thresholding and morphological filtering in 3D Slicer, followed by mesh generation (Marching Cubes) and decimation.
4. Visualization: The segmented 3D model was overlaid in AR.

Experimental Design

• Participants: 20 users (biomedical engineering/medical imaging background).
• Tasks: Two phases: (1) Planning and Image Acquisition, and (2) Needle Insertion (Facet Joint injection and Lumbar Puncture).
• Metrics:
  • Quantitative: Planning time, insertion time, placement error (Euclidean distance to target).
  • Qualitative: NASA-TLX (workload), SUS (usability), SEQ (task ease), and a Trust in Automation questionnaire.

3. Key Contributions

• Unified Comparative Framework: First study to directly compare US-based shape completion and CBCT-based workflows within the same AR environment, controlling for interface and registration variables.
• Integration of Robotic US and Deep Learning: Demonstrated a fully automated pipeline where robotic scanning and probabilistic shape completion generate 3D anatomical models for AR guidance without manual segmentation.
• Quantitative Performance Benchmarking: Provided empirical data on how imaging modality choice affects user performance, accuracy, and trust in spinal interventions.
• Identification of Modality-Specific Limitations: Highlighted that while US offers radiation-free adaptability, its reliance on shape completion introduces specific error profiles in deeper spinal regions compared to the high-fidelity static models of CBCT.

4. Results

Planning Phase

• Time: CBCT planning was significantly faster (46.78s) compared to US (83.50s) ($p < 0.001$).
• Usability: No significant difference in SUS, NASA-TLX, or task ease scores between the two modalities. Both were rated as usable.
• Radiation: CBCT scans incurred an average dose of 24.84 mGy.

Needle Insertion Phase

• Time: CBCT showed marginally faster insertion times, though not statistically significant.
• Accuracy (Placement Error):
  • Facet Joint (FJ): No significant difference between US and CBCT.
  • Lumbar Puncture (LP): CBCT yielded significantly lower error (8.89 mm) compared to US (24.33 mm) ($p < 0.001$).
  • Combined Error: CBCT was significantly more accurate overall ($p < 0.001$).
• Subjective Ratings:
  • Usability & Task Ease: Significantly higher for CBCT ($p < 0.01$).
  • Workload (NASA-TLX): Significantly lower for CBCT ($p < 0.05$).
  • Trust: Users reported significantly higher trust in the CBCT-derived models ($p < 0.01$).

5. Significance and Conclusion

The study concludes that both workflows are viable, but they serve complementary roles:

• CBCT-based AR excels in efficiency, precision, and user confidence, particularly for deep or complex targets where anatomical completeness is critical. It provides a "gold standard" static context.
• US-based AR offers a radiation-free, adaptive alternative but is currently limited by the accuracy of shape completion algorithms in regions with poor US visibility (e.g., lower lumbar spine), leading to higher cognitive load and placement errors.

Clinical Implication: The findings motivate a hybrid AR guidance strategy. This approach would utilize CBCT for initial global anatomical context and high-precision planning, while integrating dynamic US updates for intraoperative adaptation to patient motion or deformation. Future work aims to develop real-time multimodal fusion to combine the safety of US with the fidelity of CBCT.