Polyp Instance Segmentation in Colonoscopy Images: A Benchmark

Abstract

Colonoscopy has become the most popular technique to detect abnormalities, which are polyps or adenomas in the colon, to prevent them from becoming cancerous. However, the risk of mistakes during clinical examination is significant. Therefore, it is necessary to have a support system that provides reliable predictions and helps doctors not to neglect abnormal signs of disease manifestations. While most current datasets and methods only focus on solving the polyp semantic segmentation problem, polyp shape is also one of the crucial factors that help classify and rank the risk of colorectal cancer. Therefore, in addition to surveying dataset benchmarks and polyp semantic segmentation methods for analyzing endoscopic images, this paper aims to construct a dataset of endoscopic images with segmentation and shape annotations for each independent polyp instance. The newly built dataset is then used for benchmarking current state-of-the-art instance segmentation methods. The results show the feasibility of using these methods to detect polyps by shape in endoscopic images that help identify signs of colorectal cancer.