SARM-Net: A Spatial Attention-Based Residual M-Net for Polyp Segmentation

Abstract

Polyp segmentation is a crucial step for the early diagnosis of colorectal cancer. However, the heterogeneous nature of polyps poses a significant challenge in the segmentation task, and it is still an unsolved problem. So in this study, we have proposed a deep learning network, namely “a spatial attention-based residual M-Net for polyp segmentation” (SARM-Net). The network is inspired by the M-Net architecture, where some additional modules are added to the existing architecture to improve the segmentation performance. We have employed residual connections in the M-Net architecture to preserve gradient information during backpropagation, facilitating optimal gradient flow. Meanwhile, unlike M-Net, where the contextual information is directly fed from the encoder to the decoder by skip connections, a spatial attention block (SAB) is introduced in our proposed network to focus on the relevant significant features and ignore the redundant features in the spatial dimension prior to the concatenation which will facilitate better optimization of the network. The segmentation performance was evaluated on the “Kvasir-SEG” database. The experimental results reflect the segmentation performance improvement compared to the traditional deep learning models and a recent state-of-the-art method.