Doctoral Dissertations

Date of Award

12-1998

Degree Type

Dissertation

Degree Name

Doctor of Philosophy

Major

Engineering Science

Major Professor

Laurence Miller

Committee Members

Keith Eckerman, Peter Groer, Jerry Carley

Abstract

In vivo targeted radiotherapy has the potential to be an effective treatment for many types of cancer. Agents which show preferred uptake by cancerous tissue are labeled with radio-nuclides and administered to the patient. The preferred uptake by the cancerous tissue allows for the delivery of therapeutically effective radiation absorbed doses to tumors, while sparing normal tissue. Accurate absorbed dose estimation for targeted radiotherapy would be of great clinical value in a patient's treatment planning. One of the problems with calculating absorbed dose involves the use of geometric mathematical models of the human body for the simulation of the radiation transport. Since many patients differ markedly from these models, errors in the absorbed dose estimation procedure result from using these models. Patient specific models developed using individual patient's anatomical structure would greatly enhance the accuracy of dosimetry calculations. Patient specific anatomy data is available from CT or MRI images, but the very time consuming process of manual organ and tissue identification limits its practicality for routine clinical use. This study uses a statistical classifier to automatically identify organs and tissues from CT image data. In this study, image "slices" from thirty five different subjects at approximately the same anatomical position are used to "train" the statistical classifier. Multi-dimensional probability distributions of image characteristics, such as location and intensity, are generated from the training images. Statistical classification rules are then used to identify organs and tissues in five previously unseen images. A variety of pre-processing and post-processing techniques are then employed to enhance the classification procedure. This study demonstrates the promise of statistical classifiers for solving segmentation problems involving human anatomy where there is an underlying pattern of structure. Despite the poor quality of the image data set, and the fact that the identification is done in the highly variable abdominal region, the overall results are quite good. The two classifiers developed in this study have overall identification results of greater than 88% over the five test images. For both classifiers, three of the five test images have greater than 90% correct pixel classification.

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

Share

COinS