An Optimized Spline-Based Registration of a 3D CT to a Set of C-Arm Images

Slavica Jonic, Philippe Thévenaz, Guoyan Zheng, Lutz-Peter Nolte, Michael UnserInt. J. Biomed. Imaging 2006 · DOI

TL;DR

This paper presents an optimized algorithm for registering 3D CT scans with 2D C-arm images, crucial for image-guided surgery. By leveraging a continuous B-spline model and a novel fast projection method, the algorithm achieves accurate and robust registration with fewer C-arm views, potentially reducing radiation exposure and improving surgical precision.

What Problem Does This Paper Address?

Image-guided navigation systems are vital in modern surgery, allowing surgeons to visualize instruments superimposed on pre-operative 3D CT data. However, registering the patient's current anatomy to the pre-operative CT at the time of surgery is a significant challenge. While C-arm fluoroscopy provides real-time 2D images, it lacks the full 3D information of CT. Existing methods for registering 3D CT to 2D C-arm images fall into two categories: those that match anatomical features (requiring accurate segmentation) and projection-based methods (requiring little to no segmentation). Projection-based methods are generally more robust but can be computationally intensive. This paper aims to bridge this gap by developing an improved, automatic 3D CT-based navigation system that uses a minimal number of intraoperative C-arm images for registration, thereby enhancing efficiency and potentially reducing radiation dose.

Methods

The core of the proposed method is a projection-based registration algorithm. It works by iteratively minimizing a least-squares measure of dissimilarity between the C-arm images and projections of the 3D CT volume.

Key novelties and techniques include:

  • Continuous B-spline Model: The CT volume and C-arm images are modeled using continuous cubic B-splines. This ensures high interpolation accuracy and provides a well-defined gradient of the dissimilarity measure, crucial for optimization.
  • Fast Projection Method (Shearing): A novel "shearing" method is introduced to compute volume projections efficiently. This technique replaces a computationally expensive 3D interpolation with a faster 2D interpolation by changing the integration variable along the projection ray.
  • Multiresolution Strategy (Coarse-to-Fine): To improve robustness and speed, the algorithm employs a multiresolution approach. It starts registration at a coarse resolution of the data and progressively refines it at finer resolutions, using the results from coarser levels as initial estimates. This is implemented using B-spline least-squares L2 pyramids.
  • Levenberg-Marquardt (LM) Optimizer: The optimization of the dissimilarity measure is performed using the Levenberg-Marquardt algorithm, which is well-suited for least-squares problems and offers a good balance between the robustness of steepest descent and the efficiency of Newton's method.
  • Gauss-Newton Approximation of the Hessian: To further improve the optimization process, a Gauss-Newton approximation of the Hessian matrix is used, which helps ensure positive-definiteness.

Datasets

The experiments were conducted using a human cadaver spine specimen.

  • CT Data: The specimen was CT-scanned, resulting in a 3D volume.
  • C-arm Images: A set of 2D C-arm images were acquired from the specimen.
  • Fiducial Markers: Five fiducial markers were implanted on the spine. These markers were used to establish a "ground-truth" registration, against which the algorithm's performance was evaluated. The coordinates of these markers were digitized in both the patient coordinate system (P-COS) and the reference coordinate system (R-COS).

The paper notes that the experiments were performed on both single vertebrae (by masking out other regions) and multiple vertebrae to simulate clinical scenarios and establish upper bounds on accuracy.

Key Insights and Ideas

  • B-splines are powerful for registration: The continuous nature of B-splines provides a smooth and well-behaved representation of the volume, leading to a well-defined gradient essential for accurate optimization. This is a significant advantage over discrete voxel-based methods.
  • Shearing for faster projections: The proposed shearing method for projection is a key algorithmic innovation. By cleverly changing the integration variable, it reduces the computational burden of generating projections, making the overall registration process faster.
  • Multiresolution is crucial for efficiency and robustness: The coarse-to-fine strategy significantly speeds up the registration process and improves its robustness by starting with a more global alignment before refining it. This is a common but effective technique in image registration.
  • Fewer views can suffice with good geometry: The study demonstrates that accurate registration can be achieved with as few as two or three C-arm views, provided their planes have a sufficient angle between them (e.g., >15° or >25°). This has direct implications for reducing patient and staff radiation exposure.
  • Joint registration of multiple vertebrae offers higher accuracy: Registering multiple vertebrae simultaneously, when possible, yields better accuracy than registering single vertebrae, highlighting the benefit of more anatomical context.

Key Results

  • Accuracy: The algorithm achieved registration accuracy of approximately 1.7 ± 0.5 mm when using pairs of C-arm images with angles >15° and registering all three vertebrae. For single-vertebra registration, an accuracy of 1.7 ± 0.3 mm was achieved with angles >25°. These accuracies are noted as being subvoxel with respect to the CT slice thickness of 2.5 mm.
  • Speed: The multiresolution strategy significantly reduced computation time. For example, stopping the registration at level K=2 (coarser levels) for three C-arm images took approximately two minutes on a Pentium III 700 MHz processor.
  • Influence of Views and Angle: Increasing the number of C-arm views generally improved accuracy. Enforcing a larger minimal angle between image planes (e.g., 25° vs. 15°) led to a reduction in the standard deviation of the registration accuracy, suggesting the elimination of outliers.
  • Tradeoff: A good tradeoff between radiation dosage, accuracy, and registration time was found with three C-arm orientations.

Limitations and Future Work

  • Ground Truth Accuracy: The accuracy of the ground truth itself was limited by the precision of the fiducial marker digitization (0.5 ± 0.5 mm navigation error) and CT index determination (up to 0.6 mm error). This means the reported residual misalignment is not solely attributable to the registration algorithm.
  • Data Masking: For single-vertebra experiments, large areas of the C-arm images were masked out to remove non-anatomical data (DRB, clamps) and fiducial markers. This reduced the available data for registration, particularly for the L4 vertebra.
  • Clinical Validation: The authors suggest that the algorithm should be validated using more realistic data, including soft tissues and actual patient data, to assess its performance in a clinical setting.
  • Perpendicular Angles: The paper mentions that allowing image planes to be closer to perpendicular might further improve accuracy, but the available data did not allow for this investigation. This suggests future work could explore optimal C-arm positioning.

Summary updated

BibTeX
@article{jonic2006optimized,
  author = {Slavica Jonic and Philippe Thévenaz and Guoyan Zheng and Lutz-Peter Nolte and Michael Unser},
  title = {{An Optimized Spline-Based Registration of a 3D CT to a Set of C-Arm Images}},
  journal = {Int. J. Biomed. Imaging},
  year = {2006},
  doi = {10.1155/ijbi/2006/47197},
}
Figure