MANet: multi-attention network for polyp segmentation

• A new multi-attention based polyp automatic segmentation network is proposed. • Mixed attention and self-attention mechanisms are exploited to improve the accuracy. • A camouflage identification method is designed to distinguish the foreground polyp and background. • Experiments verify that the MANet produces superior performance for polyp segmentation . Currently, colonoscopy stands as the most efficient approach for detecting colorectal polyps. In clinical diagnosis, colorectal cancer is closely related to colorectal polyps. Therefore, precise segmentation of polyps holds paramount importance for the early detection and clinical diagnosis of colorectal cancer. Among conventional segmentation methods, multi-layer feature extraction is prone to ignore shallow features, while the segmentation of diminutive polyps perpetually depends on shallow features. Meanwhile, some polyps are frequently hide confusingly in the background due to their own characteristics, resulting in high similarity and low contrast in the anterior and posterior views, which causes an aggravation of distinguishing colorectal polyps during segmentation. In this work, we depict a multi-attention built upon polyp automatic segmentation network, which is called multi-attention network (MANet). In detail, we first implement the shallow feature extraction module (SFEM) to augment the representation ability of diminutive polyps. Then, to conquer the visual confusion of background similarity in the polyp region, a camouflage identification module (CIM) is devised to better identify the polyp region and assisted in segmentation of polyps. Finally, CIM is combined with the extracted shallow features to ameliorate the accuracy of polyp segmentation. Qualitative evaluation on five challenging datasets shows that our proposed multi-attention-based network model shows promising segmentation accuracy, especially in detecting small polyps with low contrast.