The presence of uninformative frames in colonoscopy videos is a major factor that reduces the accuracy and efficiency of various video analysis applications. To address this issue, research on informative frame classification has been conducted, but the lack of a publicly available dataset has made reproducibility difficult. In this study, we propose a novel dataset, InfoColon, which integrates video data collected from multiple medical institutions with major public colonoscopy datasets. All colonoscopy frames were labeled as either an informative frame or one of six types of uninformative frames. We also propose an active learning method to efficiently label large amounts of data with a small initial labeled dataset. Using the constructed InfoColon, we demonstrate the potential for its application in consecutive informative frame classification and 3D reconstruction. We expect that the proposed InfoColon will be valuable for various applications involving colonoscopy video analysis.
Choi et al. (Thu,) studied this question.