MSB-Net: Multi-Scale Boundary Net for Polyp Segmentation

Abstract

Polyp of intestinal tract is the precursor of colorectal cancer. Accurate computer-aided polyp location and segmentation in colonoscopy is of great importance since it provides valuable information for endoscopists. However, polyps are arduous to be segmented due to their high inter-class similarity, high intra-class variation, and low contrast with surrounding mucosa. To address these challenges, we propose a multi-scale boundary network (MSB-Net) for polyp segmentation. We first focus on the multi-scale feature representation and propose a novel architectural unit to extract intra-stage and contextual information, which is named ResU-Block (RUB). RUBs are connected by the proposed multi-squeeze-and-excitation (Multi-SE) units which can recalibrate the feature information from a multi-scale perspective. We then generate a coarse prediction using the partial decoder, of which the boundary is further refined by a shallow-level attention (SA) module. In addition, we exploit the boundary details using a set of reverse attention (RA) modules, which can progressively establish relationships between regions and boundaries from deep-level features. Comprehensive experiments on five public datasets across five metrics elucidate that our architecture outperforms other SOTA methods by a large margin while maintaining comparable model complexity and inference speed.