Section: New Results

Computational Physiology

Deep Learning based Metal Artifacts Reduction in post-operative Cochlear Implant CT Imaging

Participants : Zihao Wang [Correspondant] , Clair Vandersteen, Thomas Demarcy, Dan Gnansia, Charles Raffaelli, Nicolas Guevara, Hervé Delingette.

This work is funded by the Provence-Alpes-Côte-d'Azur region, the Université Côte d’Azur and Oticon Medical through CIMPLE https://team.inria.fr/epione/en/research/cimple/ research project.

Generative Adversarial Network, Metal Artifacts Reduction, Cochlea Implantation

We propose a 3D metal artifact reduction method using convolutional neural networks for post-operative cochlear implant imaging.[44]

  • Learn metal artifacts reduction by using pre-operative images and metal artifacts simulation to create image pairs for training GANs.

  • Metal artifacts simulation starts from a cochlea implantation fusion image and ends with the simulated post-operative image.(Fig. 15)

  • A 3D generative adversarial network (MARGANs) to create an image with a reduction of metal artifacts.

  • Evaluations on ten patients show the effectiveness of artefact reduction compared to two classical methods.

Figure 15. CI metal artifacts simulation workflow starting from a pre-operative image and ending with the simulated post-operative image after 9 processing steps.

Kinematic Spiral Shape Recognition in the Human Cochlea

Participants : Wilhelm Wimmer [Correspondant] , Clair Vandersteen, Nicolas Guevara, Marco Caversaccio, Hervé Delingette.

Supported by the Swiss National Science Foundation (no. P400P2_180822) and the French government (UCA JEDI - ANR-15-IDEX-01).

Approximate maximum likelihood, kinematic surface recognition, natural growth

To improve therapies for hearing loss and deafness, e.g., with auditory neuroprostheses, we developed a reliable detection algorithm for the cochlear modiolar axis in CT images (Fig. 16). The algorithm was tested in an experimental study with 4 experts in 23 human cochlea CT data sets [45] [27]. Our experiments showed that the algorithm reduces the alignment error providing more reliable modiolar axis detection for clinical and research applications.

Figure 16. Visualization of the bony labyrinth with reference modiolar axis (dashed line). Modiolar axes after manual landmark-based (left), PCA-based (middle), and robust kinematic detection (right) in CT data are shown for comparison.