M.S. Thesis Defense - Kung-En (Dean) Lin

Date: 20 Nov, 2008
Time: 2:00 PM
Location: 223 Atanasoff Hall
Topic: A comparative study of techniques for detecting villi images of colonoscopy video
Major Professor(s): Johnny Wong and Wallapak Tavanapong


Abstract:

Colorectal cancer is the second leading cause of cancer-related deaths in the United States. Colonoscopy is the preferred screening method for colorectal cancer. The appearance of images showing villi (finger-like tissues) during colonoscopy indicates that the terminal ileum portion of the small intestine has been reached. This means that the entire colon has been intubated during colonoscopy, which is one of the recommendations for quality inspection by the American Society of Gastroenterology and American College of Gastroenterology.

For automatic measurement of quality of colonoscopy, we investigate techniques for detecting villi images using Support Vector Machine and existing texture features, namely Local Binary Pattern (LBP), rotation invariant uniform LBP, and Haar wavelet features. We evaluate these techniques on real colonoscopy images according to four traditional performance metrics: sensitivity, specificity, accuracy, and precision. In addition, we propose a pre-processing step before feature extraction. Our proposed pre-processing technique does not depend on colors or brand of endoscopes. Experimental results show that the classification using rotation invariant uniform LBP features outperforms the classification using the other features, achieving over 90% in all performance metrics. Furthermore, the pre-processing step increases sensitivity by 6.6% compared to when the pre-processing step is not used.