| Magnetic Resonance Imaging (MRI) provides a non-invasive technique for studying brain structure and function. For the fragile newborn infant, it provides a low-risk means for collecting information about possible injury to the developing brain. In order to analyze this data quantitatively, structures or tissues of interest within the MRI image must first be identified. While this can be achieved through time-consuming, interactive hand-drawing, this method is subject to individual bias, and leads to intra- and inter-rater variability. This dissertation presents novel algorithms for segmentation and atlas construction, which together form a unique method for fully-automated labeling of cortical and subcortical gray matter, myelinated and unmyelinated white matter, and cerebrospinal fluid in Mill images of newborn brain. The key contributions of the segmentation algorithm are a novel method for training a supervised classifier from example images and a new machine learning technique based on classifier fusion and stochastic editing of training data. The atlas construction algorithm introduces a data-driven notion of group typical anatomy and a new solution, based on the Expectation-Maximization algorithm, which provides both an estimate of the group typical anatomy and transforms relating each subject to the group. This new technique is naturally resistant to outliers, converges to a stable result with fewer samples from a population than previous methods, and provides a label-specific measure of match useful for determining how well subjects match a group under a given transform model. Together these methods form the first fully-automatic newborn segmentation algorithm capable of distinguishing myelinated from unmyelinated white matter and cortical from sub-cortical gray matter. This new method makes possible reproducible, large scale studies of newborn brain with MRI and has generated preliminary results in a study of methadone exposure in utero, as well as a study of preterm infants receiving a novel intervention for complications of prematurity. |
