Doctor of Philosophy (PhD)


Computer Science

Document Type



Magnetic resonance imaging (MRI) is a technique used primarily in medical settings to produce high quality images of the human body’s internal anatomy. Each image is of a thin slice through the body, with the typical distance between slices being a few millimeters. Brain segmentation is the delineation of one or more anatomical structures within images of the brain. It promotes greater understanding of spatial relationships to aid in such tasks as surgical planning and clinical diagnoses, particularly when the segmented outlines from each image slice are displayed together as a surface in three-dimensions. A review of the literature indicates that current brain segmentation methods require a trained human expert to inspect the images and decide appropriate parameters, thresholds, or regions of interest to achieve the proper segmentation. This is a tedious time-consuming task because of the large number of images involved. A truly automatic method is needed to transform brain segmentation into a practical clinical tool. This dissertation describes a novel pattern classification approach to the problem of automatically segmenting magnetic resonance images of the brain. Based on this approach, algorithms were designed and implemented to automatically segment a number of anatomical structures. These algorithms were applied to several standard image data sets of human subjects obtained from the Internet Brain Segmentation Repository (IBSR). The resulting segmentations of the lateral ventricles and the caudate nuclei were compared to reference manual segmentations done by expert radiologists. The Tanimoto similarity coefficient was very good for the lateral ventricles (0.81) and good for the caudate nuclei (0.67).



Document Availability at the Time of Submission

Release the entire work immediately for access worldwide.

Committee Chair

John Tyler