Robust and Reliable Deep-Learning-Based Techniques for Image-Guided Cochlear Implant Procedures

Abstract

Cochlear implants (CIs) are neuroprosthetic devices that restore hearing in patients with severe-to-profound sensorineural hearing loss by bypassing damaged auditory pathways. Despite their success, CI outcomes vary considerably, highlighting the need for image-guided methods to optimize surgical planning and CI programming. This dissertation develops robust and reliable deep learning-based techniques for three critical aspects of image-guided cochlear implant procedures. First, it introduces a hybrid active shape model and deep learning (DL) method for robust and accurate segmentation of the intracochlear anatomy in clinical CT images. Second, it presents a unified DL framework for electrode array (EA) localization that can handle both closely- and distantly-spaced EAs, complemented by an automated image quality assessment network to ensure reliable results. Finally, it addresses the challenge of MR-only preoperative planning by developing methods to synthesize CT images from multi-sequence MRI, demonstrating robustness across multiple imaging sites and scenarios with missing modalities. As a whole, the methods in this dissertation achieve high accuracy while ensuring robustness and reliability, marking a significant advancement in improving image-guided cochlear implant procedures through automated image analysis techniques.