Enhancing Informative Frame Filtering by Water and Bubble Detection in Colonoscopy Videos Ashok Dahal 1 , JungHwan Oh 1 , Wallapak Tavanapong 2 , Johnny Wong 2 , and Piet C. de Groen 3 1 Department of Computer Science and Engineering, University of North Texas, Denton, TX 76203, U.S.A. 2 Computer Science Department, Iowa State University, Ames, IA 50011, U.S.A. 3 Mayo Clinic College of Medicine, Rochester, MN 55905, U.S.A. Abstract - Colonoscopy has contributed to a marked decline in the number of colorectal cancer related deaths. However, recent data suggest that there is a significant (4-12%) miss-rate for the detection of even large polyps and cancers. To address this, we have been investigating an ‘automated feedback system’ which informs the endoscopist of possible sub-optimal inspection during colonoscopy. A fundamental step of this system is to distinguish non-informative frames from informative ones. Existing methods for this cannot classify water/bubble frames as non-informative even though they do not carry any useful visual information of the colon mucosa. In this paper, we propose a novel texture feature based on accumulation of pixel differences, which can detect water and bubble frames with very high accuracy with significantly less processing time. The experimental results show the proposed feature can achieve more than 93% overall accuracy in almost half of the processing time the existing methods take. Keywords: Colonoscopy; Clustering; Texture; Pixel Difference; Feature Extraction 1 Introduction Colonoscopy is an endoscopic technique that allows a physician to inspect the mucosa of the human colon. It has contributed to a marked decline in the number of colorectal cancer related deaths [1]. However, recent data suggest that there is a significant (4-12%) miss-rate for the detection of even large polyps and cancers [2]. To address this, we have been investigating an ‘automated feedback system’ which informs the endoscopist of possible sub-optimal inspection during colonoscopy in order to improve the quality of the actual procedure being performed [3, 4]. A fundamental step of this system is to distinguish non- informative frames from informative ones. An informative frame in a colonoscopy video can be broadly defined as a frame which is useful for convenient naked-eye analysis of the colon mucosa (Fig. 1). A non-informative frame has the opposite definition (Fig. 2). In general, non-informative frames can be considered out-of-focus frames. Informative and non- informative frames can be loosely termed as clear and blurry frames, respectively. We developed an accurate algorithm for this informative frame filtering (IFF) [5], which is firstly to detect the presence of such vivid lines, and secondly to measure the amount of curvaceous connectivity they possess. Then, with a carefully chosen threshold, we identify frames which exhibit more curvaceous connectivity and classify them as informative, and vice-versa. Fig. 1. Examples of Informative Frames. Fig. 2. Examples of Non-Informative Frames. Fig. 3 shows some frames having water and bubbles, which do not carry any useful visual information of mucosa. These frames need to be classified as non-informative. However, most IFF algorithms [5, 6] classify them as informative since they have clear edges and are in-focus. These types of frames are caused by water injection for cleaning purpose during the colonoscopy procedure, and need to be discarded from the further processing. We define a frame as water or bubble frame if more than 50% of the frame is covered with water or bubble. We call the frames in Fig. 3(a-b) ‘water’ frames, and the ones in Fig. 3(c-d) ‘bubble’ frames for convenience. Based on our observation with 100 colonoscopy videos, the percentage of these frames varies from 5.6% to 20.7% and 9.7% on average. Accurately detecting and discarding water and bubble frames can improve the performance of the ‘automated feedback system’ mentioned earlier [3, 4]. (a) (b) (c) (d) Fig. 3. Examples of Water/Bubble Frames: (a) and (b) Water frames, (c) and (d) Bubble frames. In this paper, we propose a novel method for water and bubble frame detection based on image texture focusing on accumulation of pixel value differences. We compare it with other existing texture based algorithms in terms of accuracy 24 Int'l Conf. Health Informatics and Medical Systems | HIMS'15 |