*2.5. Registration of the Pre- and Post-Operative Images*

To evaluate the electrode array placement within the cochlea, we need to be able to fuse the segmentation of the pre-operative scan and the electrode contacts of the postoperative scan to the same reference coordinate system. Although the post-operative scan is deteriorated by the metallic artifacts generated by the electrode contact, it still represents the bony structures somewhat similarly to what is seen in the pre-operative image. Rigid transformation is therefore possible for aligning pre- and post-operative images. We first pre-align the pre- and post-op image pair into their canonical poses with the previously estimated landmarks and fine-tune the final transform using the Elastix package [52,53]. We have observed that even for CT or CBCT images in Hounsfield units, the Advanced Mattes mutual information [54] with 64 histogram bins performs adequately. Invalid voxels

(usually found at the boundaries of the image) and metallic artifacts in all voxels with *HU* > 2500 are masked out and not used for computing the similarity.

**Figure 4.** Angular insertion depth estimation based on the number of electrodes inserted inside the cochlea. The comparison graph shows (**left**), contrarily to our approach, how the performance of state-of-the-art approaches decreases exponentially with insertion depth. On the (**right**) is an example of a predicted trajectory (blue) and inserted electrode.
