Localization of Cochlear Implant Electrodes from Cone Beam Computed Tomography using Particle Belief Propagation

Hendrik Hachmann, Benjamin Krüger, Bodo Rosenhahn, Waldo NogueiraISBI 2021 · DOI

Introduction

Cochlear implants (CIs) are sophisticated neural prostheses designed to restore hearing by providing direct electrical stimulation to the auditory nerve. The clinical efficacy of these devices is intrinsically linked to the precise anatomical placement of the electrode array within the spiral-shaped cochlea. Accurate post-operative localization of electrode contacts is essential for optimizing device programming, such as deactivating electrodes that are poorly positioned or causing cross-talk.

The Localization Challenge

Clinicians typically rely on Cone Beam Computed Tomography (CBCT) to assess electrode placement. However, CBCT imaging presents significant hurdles:

  1. Resolution Constraints: The resolution is often low relative to the sub-millimeter scale of the electrodes.
  2. Metal Artifacts: The metallic components of the implant generate high-intensity streaks and "blobs" in the reconstructed volumes, obscuring the true positions of the contacts.
  3. Sequential Errors: Existing automated tools often process electrodes sequentially, making them prone to "drift" or "folding" errors where the algorithm loses track of the array's path in noisy regions.

Methodology: Markov Random Fields and Particle Belief Propagation

To overcome these limitations, researchers from Leibniz University Hannover and Hannover Medical School developed MRF-A, an automated localization algorithm based on Markov Random Fields (MRF).

Global Optimization via MRF

Unlike sequential methods, the MRF approach treats the entire electrode array as a globally connected system. Each electrode's position is modeled as a node in a graph, where edges represent physical and geometric constraints between neighboring electrodes.

Geometric constraints in the MRF model, including distances (d) and angles (α) between electrode candidates.

Particle Belief Propagation (PBP)

The optimal configuration of the array is found using Particle Belief Propagation. This stochastic inference method utilizes several key strategies:

  • Physical Priors: The model incorporates known physical properties of the implant, such as inter-electrode distances and the overall helical trajectory.
  • Heuristic Sampling: Instead of random sampling, "smart particles" are generated based on expected helix geometry and local intensity peaks in the CBCT data.
  • Energy Minimization: The algorithm performs multiple runs, selecting the configuration with the lowest mathematical "energy," representing the best fit to both the image data and the geometric priors.

Experimental Validation

The MRF-A algorithm was evaluated against two primary datasets:

  1. Synthetic Data: A custom generator produced simulated CBCT volumes with known ground-truth coordinates to test robustness against noise.
  2. Clinical Data: Ten post-operative CBCT scans from CI patients (using MidScala and SlimJ arrays) were manually labeled by experts to serve as a benchmark.

Key Results

The MRF-A method demonstrated a significant performance leap over existing Graph-Based (GB-A) and Correlation-Based (CB-A) automated techniques.

Comparison of localization error (mm) across electrode numbers for different algorithms. MRF-A (green) shows lower and more stable error rates.

On real clinical datasets, MRF-A achieved a 31.5% improvement in mean precision and a 48.6% improvement in median precision compared to the next best method. The mean localization error was reduced to approximately 0.50 mm, providing a level of accuracy suitable for clinical decision-making. The joint optimization approach effectively prevented the "drift" seen in other methods, as the positions of well-defined electrodes helped constrain the search for those obscured by artifacts.

Conclusion

While the MRF-A algorithm requires more computation time (approximately 21 minutes per scan) than simpler methods, its superior accuracy makes it a valuable tool for post-operative clinical assessment. By leveraging global geometric constraints and robust probabilistic inference, the method provides a reliable way to map the "spark" of electrical stimulation within the cochlea, ultimately paving the way for more personalized and effective hearing restoration.

Summary updated

BibTeX
@inproceedings{hachmann2021localization,
  author = {Hendrik Hachmann and Benjamin Krüger and Bodo Rosenhahn and Waldo Nogueira},
  title = {{Localization of Cochlear Implant Electrodes from Cone Beam Computed Tomography using Particle Belief Propagation}},
  booktitle = {ISBI},
  year = {2021},
  doi = {10.1109/isbi48211.2021.9433845},
}
Figure