Polyp Segmentation with Multi-Scale Guidance and Multi-Level Supervision

Colorectal endoscopy is an effective method for detecting and treating colorectal polyps before they evolve into colorectal cancer. However, accurately segmenting polyps in endoscopic images is challenging due to their diverse appearance in terms of size, color, texture, and indistinct borders with their surroundings. Existing methods usually first employ a single complex encoder to handle all polyps, ignoring the differences in size and shape; each decoding step only uses output from adjacent layers, ignoring important global information; and they only provide supervision to the final output, which results in inefficient segmentation of challenging samples with obscure visual characteristics. These issues lead to poor performance as data-set diversity increases. In order to address these limitations, we propose a multi-scale attention detector that can handle various polyp types and sizes. The detector employs a semantic express module to capture crucial global information and a multi-level and multi-view supervision mechanism to segment polyps. Experiments are conducted on diverse datasets including complex SUN-SEG and smaller Kvasir and CVC, and the results demonstrate that our proposed model achieves state-of-the-art performance.