BSCA-Net: Bit Slicing Context Attention network for polyp segmentation

• We propose a novel attention mechanism for polyp segmentation. • We design a novel SSBU block to extract rich contextual and geometric information. • We propose an effective encoder and decoder to capture geometric information. • Experimental results show the proposed network is superior to state-of-the-art networks. In this paper, we propose a novel Bit-Slicing Context Attention Network (BSCA-Net), an end-to-end network, to improve the extraction ability of boundary information for polyp segmentation. The core of BSCA-Net is a new Bit Slice Context Attention (BSCA) module, which exploits the bit-plane slicing information to effectively extract the boundary information between polyps and the surrounding tissue. In addition, we design a novel Split-Squeeze-Bottleneck-Union (SSBU) module, to exploit the geometrical information from different aspects. Also, based on SSBU, we propose an multipath concat attention decoder (MCAD) and an multipath attention concat encoder (MACE), to further improve the network performance for polyp segmentation. Finally, by combining BSCA, SSBU, MCAD and MACE, the proposed BSCA-Net is able to effectively suppress noises in feature maps, and simultaneously improve the ability of feature expression in different levels, for polyp segmentation. Empirical experiments on five benchmark datasets (Kvasir, CVC-ClinicDB, ETIS, CVC-ColonDB and CVC-300) demonstrate the superior of the proposed BSCA-Net over existing cutting-edge methods.