Multi-modal image registration, such as MRI to CT, is an important but often challenging aspect for clinical image analysis. It has applications in treatment planning requiring image fusion, or inter-subject, atlas-based analyses, as well as longitudinal analyses. Multi-modal image registration in the cervical spine presents extra challenges because of the variability in the field of view (FoV) of magnetic resonance imaging (MRI) of the spinal column between different image series, with cervical vertebrae having a similar appearance leading to many local registration minima.

This thesis explores methods to detect, localize, and label cervical vertebrae and the spinal cord in anatomical MRI, focusing on the application of deep learning techniques. Specifically, we generated a custom annotated dataset of the cervical spine MRI, based on the Spine Generic dataset [1], resulting in 149 T1w and 100 T2w labelled images. We then successfully trained a Mask R-CNN model [2] and utilized a weighted directed acyclic graph (DAG) to leverage the sequential hierarchy of the vertebrae to filter detections for the cervical vertebrae and spinal cord detection task. This resulted in state-of-the-art performance, where the model was robust to varying cervical spine FoVs. Lastly, we integrated the detector model into a multi-sequence and multi-modal deformable image registration pipeline, where the inference results were used to crop images to an appropriate FoV and seed initial alignment, prior to deformable registration. The multi-sequence pipeline utilized the generated custom dataset and successfully demonstrated the use of a trained cervical vertebrae detector for FoV cropping prior to affine and deformable registration. The multi-modal pipeline was created, adopting from the multi-sequence pipeline, but with an additional CT vertebrae detection branch and utilized affine registration prior to FoV cropping. It was evaluated utilizing a prospectively maintained clinical dataset (SpineMets) provided by collaborators at Sunnybrook Research Institute, containing treatment planning CT and MRI scans. The pipeline demonstrated the ability to successfully perform deformable registration on multi-modal imaging. However, performance on this clinical dataset was limited by the Mask R-CNN model.

The methodology developed in this thesis demonstrates the potential of deep learning object detection models combined with leveraging vertebrae hierarchy to provide robust detections of vertebrae in the cervical spine. In addition, we illustrated the application of these models to determine vertebrae FoVs towards improving multi-modal deformable image registration. By leveraging these techniques, this research can be applied to automate and increase efficiency of cervical spine image registration for clinical needs, resulting in reduced cost and burden on clinical experts.