Unsupervised segmentation of colonic polyps in narrow-band imaging data based on manifold representation of images and Wasserstein distance

摘要

Colorectal cancer (CRC) is one of the most common cancers worldwide and after a certain age (>= 50) regular colonoscopy examination for CRC screening is highly recommended. One of the most prominent precursors of CRC are abnormal growths known as polyps. If a polyp is detected during colonoscopy examination the endoscopist needs to decide whether the polyp should be discarded, removed, or biopsied for further examination. However, the last two options involve some risks for the patient, while not all the polyps are precancerous. On the other hand, discarding a polyp has the risk of failing to detect CRC. We propose an automatic and unsupervised method for the segmentation of colonic polyps for in vivo Narrow-Band-Imaging (NBI) data. Polyp segmentation is a crucial step towards an automatic real-time polyp classification system, that could help the endoscopist in the diagnosis of CRC. The proposed method is a histogram based two-phase segmentation model, involving the Wasserstein distance. These histograms incorporate fused information about suitable image descriptors, namely semi-local texture, geometry and color. To test the proposed segmentation methodology we use a dataset consisting of 86 NBI polyp frames: the 83% sensitivity, 95% specificity, and 93% accuracy suggest a better performance compared to the results obtained with other methods. (C) 2019 Elsevier Ltd. All rights reserved.