Deep Learning Methods for Inner Ear Modeling and Augmented-Reality-Guided Cochlear Implantation

Abstract

Cochlear implants (CIs) are neural prosthetics used to treat patients with severe-to-profound hearing loss. Factors such as postoperative programming and intraoperative electrode placement can significantly impact hearing outcomes for patients. This dissertation presents deep-learning-based methods to extend frameworks for inner ear modeling and augmented reality (AR) guided surgery. We propose a loss term inspired by the Mumford-Shah functional to increase accuracy for both weakly and self supervised applications. We present methods of accurately localizing the internal auditory canal, auditory nerve fibers, and epitympanum in preoperative CT images, as well as methods of registering preoperative segmentations to intraoperative surgery video frames. These contributions support more accurate electrode placement and patient-specific CI programming, which can improve hearing outcomes for patients with cochlear implants.