Pedicle Drilling Planning Transfer for Spine Surgery Using Functional Map Correspondences

Lilyan Leblanc, Raphaël Vialle, C. De Farias, E. Saghbiny, Naresh Marturi, Brahim TamadazteIROS 2024 · DOI

TL;DR

This paper introduces a novel method for accurately transferring planned pedicle screw trajectories from pre-operative CT scans to intra-operative spine models, even when the latter are noisy, incomplete, or deformed. By leveraging the Functional Maps (FM) framework, the approach significantly improves the reliability of surgical planning in spine surgery, potentially reducing complications and radiation exposure.

What Problem Does This Paper Address?

Precise placement of pedicle screws is critical in spine surgery to ensure stability and avoid severe complications like spinal cord injury or vascular damage. While navigation systems have improved accuracy, they struggle with the dynamic nature of the spine during surgery. Patient movement, muscle relaxation, and the inherent flexibility between vertebrae can cause the spine to deform, making it challenging to accurately map pre-operative plans to the real-time surgical field. Repeated intra-operative imaging to compensate for these changes increases radiation exposure for both patients and surgeons. Existing methods often fail when faced with noisy, incomplete, or deformed intra-operative data, necessitating a more robust approach for trajectory transfer.

Methods

The core of this work is the application of the Functional Maps (FM) framework to address the challenge of transferring surgical trajectories. The FM framework reframes the problem of finding point-to-point correspondences between shapes into a more computationally efficient problem of matching functions defined on those shapes.

Here's a breakdown of their pipeline:

  1. Compute Functional Bases: They use Laplace-Beltrami eigenfunctions to create orthonormal bases for both the source (pre-operative) and target (intra-operative) spine models. These functions are invariant to rigid motions and provide a low-pass filtering effect, smoothing out correspondences.
  2. Compute Descriptors: Global descriptors (Wave Kernel Signature - WKS) and local descriptors (Heat Kernel Signature - HKS) are computed for each model. These descriptors are robust to isometric deformations and noise.
  3. Estimate Functional Map (C): The FM, represented by a matrix 'C', is optimized to minimize an energy function. This step incorporates constraints to ensure approximate isometry and orientation preservation, and a mask to handle partial data.
  4. Refine and Convert to Point-wise Map: The estimated FM is refined using iterative alignment (Bijective and Continuous ICP - BCICP) and spectral up-sampling (ZoomOut algorithm). This refined FM is then converted into a point-wise mapping (Tp) that relates points on the source model to points on the target model.
  5. Transfer Drilling Trajectories: Using the point-wise mapping, the entry and exit points of the planned pedicle drilling trajectories are transferred from the source to the target spine model. To mitigate potential outliers from local discontinuities in the map, they segment individual vertebrae and compute rigid transforms for each, ensuring accurate transfer even with non-rigid deformations.

Datasets

The paper uses a simulated spine model derived from CT scans, consisting of five lumbar vertebrae. This model is then subjected to various simulated deformations (rigid, non-rigid, partiality) and noise levels to create different challenging scenarios for evaluation. No real patient data or new benchmark datasets are introduced; rather, the authors create a comprehensive set of test cases based on a single model.

Key Insights and Ideas

  • Functional Maps for Surgical Trajectory Transfer: The paper demonstrates that the FM framework, typically used in computer graphics for shape matching, is highly effective for transferring precise surgical trajectories, even in the presence of significant deformations and noise. This opens up new avenues for applying advanced geometric processing techniques to medical imaging.
  • Leveraging Vertebral Rigidity: By segmenting individual vertebrae and computing rigid transforms for each, the method effectively handles non-rigid spine deformations. This insight is crucial because while the spine as a whole can deform, individual vertebrae are largely rigid, allowing for more accurate local registration.
  • Robustness to Noise and Partiality: The FM approach, with its iterative refinement and specific constraints, shows remarkable resilience to noisy and incomplete data, which are common issues in intra-operative surgical environments.
  • Efficiency: The method achieves accurate results in just a few minutes per spine model using non-optimized Matlab code, suggesting its potential for real-time clinical application.

Key Results

The paper reports excellent performance across various challenging scenarios:

  • Overall Accuracy: In over 90 scenarios and 900 trajectory transfers, the method achieved an average linear RMSE of 0.94 mm and an average angular error of 1.32°.
  • Clinical Validity: 100% of the transferred drilling trajectories were found to be clinically valid, meaning they avoided medial or lateral breaches. This is a critical finding for surgical safety.
  • Best-Case Scenarios: Errors were negligible (linear RMSE ~0.006 mm, angular error ~0.007°), with 100% accuracy.
  • Worst-Case Scenarios: Even with noisy and partial data, the maximum linear RMSE was 2.34 mm and the maximum angular RMSE was 2.54°, still resulting in 100% clinically valid trajectories.
  • Deformation Handling: The method successfully transferred trajectories under rigid deformation, non-rigid deformation (up to 25° per vertebra), and partial data scenarios, maintaining high accuracy and validity.

Limitations and Future Work

  • Limitations:
    • The paper acknowledges that performance deteriorates in the presence of significant noise (e.g., standard deviation of Gaussian noise reaching 2 mm) when dealing with partial data. Isolated pairs of vertebrae might exhibit high errors.
  • The current evaluation is based on simulated data.
  • Future Work:
    • Expand the spine database to evaluate intra-spine and inter-spine registration more comprehensively.
  • Evaluate the method using multi-modal data (CT, ultrasound, RGB-D images) from both ex-vivo and in-vivo cases to better assess real-world surgical applications.
  • Explore finer spine model meshes to potentially improve point-wise correspondence.
  • Investigate noise filtering techniques and methods for detecting highly discontinuous registration transforms to enhance performance in noisy scenarios.
  • Consider incorporating partial mesh completion techniques like Gaussian Process Implicit Surfaces for more fragmented data.

Summary updated

BibTeX
@inproceedings{leblanc2024pedicle,
  author = {Lilyan Leblanc and Raphaël Vialle and C. De Farias and E. Saghbiny and Naresh Marturi and Brahim Tamadazte},
  title = {{Pedicle Drilling Planning Transfer for Spine Surgery Using Functional Map Correspondences}},
  booktitle = {IEEE/RSJ International Conference on Intelligent Robots and Systems},
  year = {2024},
  doi = {10.1109/iros58592.2024.10802159},
}
Figure