Machine Learning-Based Techniques for Medical Image Registration and Segmentation and a Technique for Patient-Customized Placement of Cochlear Implant Electrode Arrays

Abstract

This dissertation introduces several innovative machine learning-based techniques for medical image registration and segmentation and a technique for patient-customized placement of cochlear implant electrodes. We have made three major contributions: (1) Most deformable registration methods are dependent on a good preregistration initialization. We developed a learning-based method to automatically find a set of robust landmarks in MR images of the head to initialize non-rigid registration algorithms. To validate our method, we used it to initialize five well-established deformable registration algorithms. We compared our method to a standard approach that involves estimating an affine transformation with an intensity-based approach. We showed that for all five registration algorithms the final results are statistically better when they are initialized with our method than when the standard approach is used. (2) Cochlear implants (CIs) are surgically implanted neural prosthetic devices that are used to treat severe-to-profound hearing loss. CIs are programmed post-implantation and precise knowledge of the implant position to the intracochlear anatomy can help the programming audiologists. Over the years, we have developed algorithms that permit determining the position of implanted electrodes relative to the intracochlear anatomy using pre- and post-implantation CT image pairs. However, these do not extend to CI recipients for whom pre-implantation CT images are not available. We developed three deep learning-based methods to segment the intracochlear anatomy in the artifact-affected post-implantation CT images of CI recipients. Our methods can provide the audiologists with information that guides the post-programming process of the CIs for patients for whom the pre-implantation CT images are not available. (3) Pre-curved electrode arrays are commonly used in CIs. Modiolar placement of such arrays has been shown to lead to better hearing outcomes. We retrospectively evaluated the modiolar positioning of electrode arrays within a large cohort. We found that the average CI recipient with a pre-curved electrode array has a number of electrodes distant to the modiolus where they are not most effective. Our results also indicate the approach we propose for selecting patient-customized electrode array insertion depth would lead to better perimodiolar placement of pre-curved electrode arrays.