Automatic Segmentation of Small Pulmonary Nodules in Computed Tomography Data Using a Radial Basis Function Neural Network with Application to Volume Estimation

Date of Award

12-2008

Document Type

Dissertation

Degree Name

Doctor of Philosophy (Ph.D.)

Institution Granting Degree

University of Dayton

Cedarville University School or Department

Engineering and Computer Science

First Advisor

Dr. Russell C. Hardie

Abstract

Lung cancer continues to be the leading cause of cancer death in the United States. The automatic detection and characterization of this deadly form of cancer is an area of ongoing research. This dissertation focuses primarily on the characterization of pulmonary nodules once they have been detected. This characterization includes the accurate segmentation of nodules within three dimensional (3D) computed tomography (CT) data as well as developing accurate volume estimates from these segmentations. In this dissertation a novel approach to the segmentation of pulmonary nodules from CT data is presented in which we compute a set of candidate segmentations which are characterized by a set of measured features. We then apply a trained artificial neural network to attempt to estimate the quality of the candidate segmentations. The highest quality candidate segmentation is kept as the winner. In addition, we propose a couple of techniques for reducing the computational complexity of the segmentation algorithm. We apply both simulated annealing and the golden section test as intelligent ways of searching the solution space. Finally, techniques for the accurate estimation of nodule volume are discussed. We discuss existing volume estimation approaches as well as introduce a new variation. We provide experimental results for segmentation and volume estimation algorithms that we present including a comparison of our segmentation algorithm to segmentations created by board certified radiologists.

Share

COinS