Growth Simulation Network for Polyp Segmentation

Colonoscopy is a gold standard, while automated polyp segmentation can minimize missed rates and timely treatment of colon cancer at an early stage. But most existing polyp segmentation methods have borrowed techniques related to image semantic segmentation, and the main idea is to extract and fuse feature information of images more effectively. As we know, polyps naturally grow from small to large, thus they have strong rules. In view of this trait, we propose a Growth Simulation Network (GSNet) to segment polyps from colonoscopy images. First, the completeness map (i.e., ground-truth mask) is decoupled to generate Gaussian map and body map. Among them, Gaussian map is mainly used to locate polyps, while body map expresses the intermediate stages, which helps filter redundant information. GSNet has three forward branches, which are supervised by Gaussian map, body map and completeness map, respectively. What’s more, we design a dynamic attention guidance (DAG) module to effectively fuse the information from different branches. Extensive experiments on five benchmark datasets demonstrate that our GSNet performs favorably against most state-of-the-art methods under different evaluation metrics. The source code will be publicly available at https://github.com/wei-hongbin/GSNet