SLSNet: Local Context Attention Transformer for Automatic Polyp Segmentation

Abstract

Cancer is a major global health issue, and colorectal cancer is the second leading cause of cancer death. It can develop from high-risk polyps, so accurate segmentation of polyps is essential for the diagnosis and treatment of this disease. However, in clinical practice, the examination of polyps remains challenging because their texture and color are similar to the surrounding tissues with low contrast, causing a large number of missed cases. In this paper, we proposed an automatic polyp segmentation method termed as SLSNet, in which Swin-Unet was treated as the backbone network and further improved by embedding the local context attention (LCA) module, channel transformation (CT) module and Squeeze-and-Excitation (SE) block. Specifically, in order to fully utilize the local context information of the polyp in the segmentation network and incorporate hard sample mining, the LCA module was introduced into the encoder to enhance the attention for uncertain and complex regions. At the same time, the CT module was utilized to extract more features from the decoder and send it to the LCA module to enhance the fine segmentation ability of the model. The SE block was introduced into the decoder to adaptively recalibrate channel-wise feature responses. To assess the performance of the presented SLSNet, experiments were conducted on Kvasir-SEG and CVC-ClinicDB public data sets. Extensive experimental results show that the presented SLSNet could better extract polyp features, and achieved superior performance for the automatic polyp segmentation compared with the existing methods.