Improving brain registration and segmentation using anatomical guidance

Neurodegenerative diseases afflict a significant portion of the world's aging population, with Alzheimer's disease alone affecting close to 500,000 Canadians. Non-invasive magnetic resonance imaging has shown potential to detect structural changes at an earlier stage than neuropsychological or cognitive evaluations, however, the high degree of anatomical variability among our brains has proven to be a challenge in developing general computational methods. Neuroanatomical registration and segmentation are fundamental components of many structural and functional analysis techniques, thus advances here can lead to the development of biomarkers for early disease progression. This dissertation focuses on methods for improved accuracy and robustness using context-specific anatomical guidance and locally-adaptive methods in registration, segmentation and brain morphometry. We first describe a subcortical brain mapping algorithm using initial segmentations for context-specific guidance, and apply it in a multi-atlas segmentation framework where locally-adaptive weights are computed based on a combination of supervised learning and dynamic registration accuracy estimates. Then, we propose a whole brain diffeomorphic registration algorithm, concurrently driven with subcortical and cortical segmentations, and utilize it in a locally-adaptive morphometry framework that is aided by local estimates of registration confidence. Finally, we turn our attention to time-series image analysis and describe a longitudinal growth model for a trajectory-based description of image sequences and a method for spatio-temporal normalization to a central template, then analyze longitudinal shape changes of the hippocampus in this framework using deformation-based approaches. These contributions allow for greater accuracy and robustness in brain registration and segmentation and thus potentially more powerful studies of brain morphometry and structure.