Colorectal cancer is the second leading cause of cancer-related deaths, claiming close to 50,000 lives annually in the United States alone. Colonoscopy is an important screening tool that has contributed to a significant decline in colorectal cancer-related deaths. During colonoscopy, a tiny video camera at the tip of the endoscope generates a video signal of the internal mucosa of the human colon. The video data is displayed on a monitor for real-time diagnosis by the endoscopist. Despite the success of colonoscopy in lowering cancer-related deaths, a significant miss rate for detection of both large polyps and cancers is estimated around 4-12%. As a result, in recent years, many computer-aided object detection techniques have been developed with the ultimate goal to assist the endoscopist in lowering the polyp miss rate. Automatic object detection in recorded video data during colonoscopy is challenging due to the noisy nature of endoscopic images caused by camera motion, strong light reflections, the wide angle lens that cannot be automatically focused, and the location and appearance variations of objects within the colon. The unique characteristics of colonoscopy video require new image/video analysis techniques.

The dissertation presents our investigation on edge cross-section profile (ECSP), a local appearance model, for colonoscopic object detection. We propose several methods to derive new features on ECSP from its surrounding region pixels, its first-order derivative profile, and its second-order derivative profile. These ECSP features describe discriminative patterns for different types of objects in colonoscopy. The new algorithms and software using the ECSP features can effectively detect three representative types of objects and extract their corresponding semantic unit in terms of both accuracy and analysis time.

The main contributions of dissertation are summarized as follows. The dissertation presents 1) a new ECSP calculation method and feature-based ECSP method that extracts features on ECSP for object detection, 2) edgeless ECSP method that calculates ECSP without using edges, 3) part-based multi-derivative ECSP algorithm that segments ECSP, its 1st - order and its 2nd - order derivative functions into parts and models each part using the method that is suitable to that part, 4) ECSP based algorithms for detecting three representative types of colonoscopic objects including appendiceal orifices, endoscopes during retroflexion operations, and polyps and extracting videos or segmented shots containing these objects as semantic units, and 5) a software package that implements these techniques and provides meaningful visual feedback of the detected results to the endoscopist. Ideally, we would like the software to provide feedback to the endoscopist before the next video frame becomes available and to process video data at the rate in which the data are captured (typically at about 30 frames per second (fps)). This real-time requirement is difficult to achieve using today's affordable off-the-shelf workstations. We aim for achieving near real-time performance where the analysis and feedback complete at the rate of at least 1 fps.

The dissertation has the following broad impacts. Firstly, the performance study shows that our proposed ECSP based techniques are promising both in terms of the detection rate and execution time for detecting the appearance of the three aforementioned types of objects in colonoscopy video. Our ECSP based techniques can be extended to both detect other types of colonoscopic objects such as diverticula, lumen and vessel, and analyze other endoscopy procedures, such as laparoscopy, upper gastrointestinal endoscopy, wireless capsule endoscopy and EGD. Secondly, to our best knowledge, our polyp detection system is the only computer-aided system that can warn the endoscopist the appearance of polyps in near real time. Our retroflexion detection system is also the first computer-aided system that can detect retroflexion in near real-time. Retroflexion is a maneuver used by the endoscopist to inspect the colon area that is hard to reach. The use of our system in future clinical trials may contribute to the decline in the polyp miss rate during live colonoscopy. Our system may be used as a training platform for novice endoscopists. Lastly, the automatic documentation of detected semantic units of colonoscopic objects can be helpful to discover unknown patterns of colorectal cancers or new diseases and used as educational resources for endoscopic research.