Type of Document Dissertation Author Tamburo, Robert Joseph URN etd-05052006-134756 Title Feature-Based Correspondences to Infer the Location of Anatomical Landmarks Degree Doctor of Philosophy Program Bioengineering School School of Engineering Advisory Committee
Advisor Name Title George Stetten Committee Chair Ching-Chung Li Committee Member Fernando Boada Committee Member J. Robert Boston Committee Member Keywords
- feature-based correspondences
- anatomical landmark
Date of Defense 2006-04-24 Availability unrestricted Abstract A methodology has been developed for automatically determining inter-image correspondences between cliques of features extracted from a reference and a query image. Cliques consist of up to threefeatures and correspondences between them are determined via a hierarchy of similarity metrics based on the inherent properties of the features and geometric relationships between those features. As opposed to approaches that determine correspondences solely by voxel intensity, features that also include shape description are used. Specifically, medial-based features are
employed because they are sparse compared to the number of image voxels and can be automatically extracted from the image.
The correspondence framework has been extended to automatically estimate the location of anatomical landmarks in the query image by adding landmarks to the cliques. Anatomical landmark locations
are then inferred from the reference image by maximizing landmark correspondences. The ability to infer landmark locations has provided a means to validate the correspondence framework in the
presence of structural variation between images. Moreover, automated landmark estimation imparts the user with anatomical information and can hypothetically be used to initialize and
constrain the search space of segmentation and registration methods.
Methods developed in this dissertation were applied to simulated MRI brain images, synthetic images, and images constructed from several variations of a parametric model. Results indicate that the methods are invariant to global translation and rotation and can operate in the presence of structure variation between images.
The automated landmark placement method was shown to be accurate as compared to ground-truth that was established both parametrically and manually. It is envisioned that these automated methods could prove useful for alleviating time-consuming and tedious tasks in applications that currently require manual input, and eliminate intra-user subjectivity.
Files
Filename Size Approximate Download Time (Hours:Minutes:Seconds)
28.8 Modem 56K Modem ISDN (64 Kb) ISDN (128 Kb) Higher-speed Access tamburo_robert20060614.pdf 27.35 Mb 02:06:37 01:05:07 00:56:59 00:28:29 00:02:25 If you have questions or comments please send mail to ETD-Feedback.