DOCTORAL DISSERTATION DEFENSE
DEPARTMENT OF COMPUTER SCIENCE
Candidate: Alan David Kalvin
Advisor: Robert Hummel
SEGMENTATION AND SURFACE-BASED MODELING OBJECTS
IN THREE-DIMENSIONAL BIOMEDICAL IMAGES
10:00 a.m., Tuesday, June 18, 1991
12th fl. conference room
719 Broadway

The rapid development of technologies for imaging the human body has led to a growing interest in the extraction and analysis of objects in 3D biomedical images for applications in fields such as clinical medicine, biomedical research, and physical anthropology.

This dissertation examines the problem of creating surface-based geometric models of biomedical objects that are suitable for analysis through visualization, mensuration, and manipulation. This is a two-stage problem. First the objects are identified by segmenting the 3D image into regions of interest, and then surface-based models of the objects are created.

We discuss the issues of segmentation and surface construction and introduce the following new methods for solving these problems.

First, we present the MLO algorithm, a general-purpose, domain-independent segmentation algorithm that has been applied successfully to identify skulls in CT images, the ventricle walls of the heart in MR images, brain ventricles in CT images, and carotid arteries in MR angiography images. It uses an iterative, cooperative procedure to segment an image by optimizing a cost function. To achieve a fast segmentation, a coarse-to-fine strategy is employed, using a multiresolution pyramid.

The GRG algorithm is a model-driven, special-purpose algorithm for identifying thin bone in CT head images. The algorithm, developed specifically for craniofacial surgical planning, uses anatomical knowledge in the segmentation process, and can handle the abnormal anatomy of craniofacial patients. It successfully finds most of the thin bone that can not be found using previous methods.

ALLIGATOR is a surface construction algorithm that creates models using the ``winged-edge'' data structure of Baumgart, enabling efficient access to the topological and geometric information of the surfaces, and permitting efficient, topologically consistent modifications to the representations.

Unlike previous surface construction algorithms, ALLIGATOR is suitable not just for visualizing biomedical objects, but for measuring and manipulating them as well. Another important feature of ALLIGATOR is that it uses an adaptive face-merging process to create surface models that are significantly more concise, in terms of vertices, edges, and faces, than the models produced by other surface construction algorithms.