Segmentation and indexing of endoscopic videos

Over the last few years it has become common to archive video recordings of endoscopic surgeries. These videos are of high value for medical doctors, patients, hospital management and planning, but currently they are used rarely or not at all. One reason lies in the amount of data - each day tens to hundreds of hours of new videos are added to archives. Another reason is that no metadata are stored that would support content-based search. In order to fully utilize these videos it is necessary to analyze the content of the recordings. Endoscopic videos are in some aspects fundamentally different to other types of videos. Therefore, common content-based analysis methods must be adopted to this special kind of video. This thesis investigates segmentation and indexing of endoscopic videos. These techniques are the prerequisites for further applications like video search and browsing, keyframe extraction, video summaries, and (semi-)automatic documentation. The video footage, which is used for implementing, testing and evaluating segmentation and indexing algorithms, shows three different types of endoscopic interventions. Endoscopic video recordings are typically one-shot videos, consisting partly of poor quality due to lighting conditions, different kind of noise or camera work. Considering these characteristics two segmentation and indexing methods for endoscopic videos are discussed in this work. (I) Shots are the basic structuring elements in common videos and used as basis for numerous content analysis algorithms. In one-shot endoscopic videos basic motion patterns may serve as basis for similar purposes. The three basic motion patterns are (1) no motion, (2) camera motion and (3) instrument motion. Additionally, instrument motion (3) can be divided into insertion into the operation area, removing from the operation area and instrument specific motion patterns. In addition to detect small motion patterns the identification of instruments can help to segment endoscopic videos into much larger semantic meaningful scenes. (II) Many endoscopic surgeries are well structured procedures, some of them follow a gold standard like the laparoscopic resection of the gallbladder (cholecystectomy). Within these procedures the appearance of an instrument or a group of instruments, as well as the sequence of the appearances, indicates which step of the surgery is currently being executed. This valuable information can be used as basis for a semantic segmentation method. The second method described in this thesis uses instrument detection and classification as basis to detect certain scenes. Therefore, I evaluate the state-of-the-art classification toolchain (Bag of Visual Words (BoW) and Support Vector Machine (SVM)) with three different local features. The toolchain uses the whole frame for the classification of instruments, which negatively impacts the performance. An additional algorithm is developed that uses color information in the L*ab color space to exclude features that are not located onto instruments.Keine Zusammenfassung vorhandenDipl.-Ing. Manfred Jürgen Primus, Bakk.techn.Alpen Adria Universität Klagenfurt, Dissertation, 2015OeBB(VLID)240975