Automatic assessment of informative frames in endoscopic video
N. C. van Dongen, Fons van der Sommen, Svitlana Zinger, E. J. Sekoon, Peter H. N. de With · 2016
Recently, studies have proven the feasibility of Computer-Aided Detection (CAD) systems for the early detection of esophageal cancer, exploiting the high-quality visual information of HD endoscopy. However, many frames in an endoscopic video are not suited for the detection of cancer, as they are out-of-focus or contaminated by intestinal juices and/or specular reflections. The analysis of useless low-quality frames should be avoided and may also lead to spurious (false) detections. In this paper, we present a novel automated method for the assessment and selection of informative frames in endoscopic videos. This method can serve as a pre-processing step for a CAD system. The proposed method distinguishes informative and noninformative frames with an area under the curve of 97.4%, which significantly outperforms state-of-the-art algorithms.