Doctoral Dissertations

Date of Award

5-2001

Degree Type

Dissertation

Degree Name

Doctor of Philosophy

Major

Electrical Engineering

Major Professor

Mongi A. Abidi

Abstract

This dissertation presents the development, implementation, and application of a unified probabilistic shape and appearance model (PSAM) algorithm for boundary-based segmentation and recognition of semirigid objects on complex backgrounds. The algorithm requires that the boundary be represented by a set of landmark points (LPs). These LPs are iteratively adjusted to fit the boundary of a new object based on a priori information gathered from a training set. PSAM is derived from compound Bayesian decision theory, and the formulation is general enough that it can be used as a starting point to derive a variety of other probabilistic boundary-finding techniques. The motivation for developing PSAM arose from a need to segment and recognize semirigid anatomic structures within medical images that have faint and/or missing edge information.

The starting point for this research was the active shape model (ASM). ASM is described, along with some practical improvements that were made to the published algorithm. These practical improvements are demonstrated on synthetic and real dataASM was tested on a a set of 2D medical images of kidneys within X-ray CT images of labora- tory mice. Although ASM performance was improved because of the practical improvements, some remaining fundamental problems led to poor segmentation accuracy in many cases. These fundamental problems inspired the development of the PSAM algorithm.

PSAM contains three specific model components: (1) a global shape model (GSM), (2) a local shape model (LSM), and (3) a gray-level model (GLM)The GLM formulation is based on gradient gray-level profiles normal to the object boundary through each LP. All three of the PSAM components are optimized simultaneously when boundary searches are performed within new images. PSAM is formulated so that the influence of each of these components on the final boundary position can be controlled by the system operator. This allows the same PSAM algorithm to be used in applications with predictable global shape and relatively poor object edge strength, as well as in other applications where global shape is unpredictable but object edges are prominent.

The new PSAM algorithm formulation provides confidence metrics for each of the three model components that give the operator feedback on the segmentation result. These confidence metrics indicate how well each PSAM component (GSM, LSM, and GLM) of the final boundary fits within the distribution of each component as derived from the training data. These confidence metrics can be monitored to alert an operator to any boundary results where one or more model components were found to be "out of bounds" relative to the training data. Furthermore, it was demonstrated that for some applications, the GLM confidence metric can be used as a predictor of segmentation accuracy.

The performance of the PSAM algorithm is summarized on both synthetic and real-world dataThe results of three cases of real medical image data segmentations are presented. These cases include X-ray tomographic images of anatomic structures within laboratory mice. Specifically, the skull, the heart and lungs, and the kidneys are segmented using PSAM and ASM; and the results of the two algorithms are directly compared. In all cases the PSAM algorithm performed well and in fact, outperformed ASM by a substantial margin. It is shown that PSAM has a much larger degree of success than ASM on the most difficult segmentation cases. The PSAM performance is summarized, and a variety of future research topics are suggested that could lead to improved performance and broader applicability.

Download
Files over 3MB may be slow to open. For best results, right-click and select "save as..."

Share

COinS