CSA-Net: A Transformer-based Polyp Segmentation Network

Abstract

It is well known that colorectal polyps are a precursor to colorectal cancer. Accurate segmentation of polyp images from colonoscopy can assist clinicians in localizing polyp regions and reduce the occurrence of misdiagnosis accurately. Many existing methods achieve good results in the polyp segmentation task, but their extraction of global and local features is often insufficient. In this paper, we propose a transformer-based polyp segmentation network (CSA-Net) that utilizes two types of attention modules- spatial attention and channel attention-to further adaptively fuse local features with their global features. The proposed network is validated on five polyp datasets. Experimental results show that our model outperforms previously proposed models.