LDPolypVideo Benchmark: A Large-Scale Colonoscopy Video Dataset of Diverse Polyps

Abstract

Computer-Aided Diagnosis (CAD) systems for polyp detection provide essential support for colorectal cancer screening and prevention. Recently, deep learning technology has made breakthrough progress in medical image computation and computer-aided diagnosis. However, the deficiency of training data seriously impedes the development of polyp detection techniques. Existing fully-annotated databases, including CVC-ClinicDB, ETIS-Larib, CVC-Colon dataset, Kvasir-Seg, and CVC-ClinicVideoDB, are very limited in polyp size and shape diversity, which is far from the significant complexity in the actual clinical situation. In this paper, we propose LDPolypVideo, a large-scale colonoscopy video database that contains a variety of polyps and more complex bowel environments. Our database contains 160 colonoscopy videos and 40,266 frames in total with polyp annotations, which are four times the size of the largest existing colonoscopy video database CVC-ClinicVideoDB. In order to improve the efficiency of polyp annotation, we design an intelligent annotation tool based on object tracking. Extensive experiments have been conducted to evaluate state-of-the-art object detection approaches on our LDPolypVideo dataset. The average drops of Recall and Precision of four SOTA approaches on this dataset are 26% and 15%, respectively. The great performance drop demonstrates the significant challenges but also the great value of our large-scale and diverse polyp video dataset to facilitate the research on polyp detection. Our dataset is available at https://github.com/dashishi/LDPolypVideo-Benchmark.