Dissertation/Thesis Abstract

Image analysis of Optical Coherence Tomography images of the urinary bladder for the recognition of bladder cancer
by Lingley-Papadopoulos, Colleen A., D.Sc., The George Washington University, 2008, 214; 3336756
Abstract (Summary)

The vast majority of bladder cancers originate within 600 μm of the tissue surface, making Optical Coherence Tomography (OCT) a potentially powerful tool for recognizing cancers that are not easily visible with current techniques. OCT is a new technology, however, and surgeons are not familiar with the resulting images. Technology able to analyze and provide diagnoses based on OCT images would improve the clinical utility of OCT systems. In addition, to avoid the need for gathering training data sets each time a new OCT system is used, it is important that the technology developed be system-independent.

In this dissertation, I present an automated, system-independent algorithm that uses texture and image analysis to detect bladder cancer from OCT images. The algorithm was developed on and applied to 245 OCT images of bladder tissue (taken from three different systems and 35 patients) to classify the images as non-cancerous or cancerous. The results, when compared with the corresponding pathology, indicated that the algorithm was effective at differentiating cancerous from non-cancerous tissue with a sensitivity of 91%, and specificity of 78%.

However, further testing of the algorithm on data received after algorithm development indicated that the previously designed algorithm was too dependent on the training data. The developed algorithm worked very well on the specific images used for development, but had been made so specific to the development set, that the algorithm did not work well on other data sets. While a complex algorithm is probably necessary to differentiate cancerous tissue from the wide range of non-cancerous pathologies, a simpler algorithm is necessary to avoid over-training the algorithm given a training set of limited size.

A simpler algorithm was designed on a subset of the images available, and tested on the remaining images. By developing and training the algorithm on a subset of the images available, it was possible to test the algorithm on images that were completely independent from the test set. By comparing the results of the algorithm on the training set and the results of the algorithm on the test set, it was possible to recognize if the algorithm had been over-trained. The simpler algorithm had a sensitivity of 83% and a specificity of 64%, when tested on images not used for algorithm development and training, and a sensitivity of 90% and specificity of 63% when tested on the training data. These values were very similar, so it was concluded that the algorithm was not over-trained. Furthermore, when the simpler algorithm was trained using a data set taken with one imaging system, and trained on images taken using a different imaging system, the sensitivity was 76%, and the specificity 66%. Since these values were also similar to the values mentioned above, it was concluded that the algorithm made it possible to compare images taken with two different imaging systems.

The work described in this document offers a preliminary algorithm and design methodology. In order for the algorithm to be improved and verified, much more data taken from different imaging systems is required. However, with additional data, and further development, it should be possible to combine an algorithm similar to the one described in this dissertation with any OCT system for the purpose of guiding endoscopic biopsies toward tissue likely to contain cancer.

Indexing (document details)
Advisor: Zara, Jason M.
Commitee: Doroslovacki, Milos, Eom, Kie-Bum, Loew, Murray H., Manyak, Michael J.
School: The George Washington University
Department: Electrical Engineering
School Location: United States -- District of Columbia
Source: DAI-B 69/12, Dissertation Abstracts International
Subjects: Biomedical engineering
Keywords: Bladder cancer, Image analyis, Optical coherence tomography, Texture analysis, Urinary bladder, Wavelet analysis
Publication Number: 3336756
ISBN: 9780549906377
Copyright © 2019 ProQuest LLC. All rights reserved. Terms and Conditions Privacy Policy Cookie Policy