PAAN: Pyramid Attention Augmented Network for polyp segmentation

Sida Yi, Yuesheng Zhu, Guibo Luo

Keywords: medical image segmentation, polyp segmentation, spatial attention

Abstract: Polyp segmentation is a task of segmenting polyp lesion regions from normal tissues in medical images, which is crucial for medical diagnosis and treatment planning. However, existing methods still suffer from low accuracy in polyp boundary delineation and insufficient suppression of irrelevant background due to the blur boundaries and textures of polyps. To overcome these limitations, in this paper a Pyramid Attention Augmented Network (PAAN) is proposed, in which a pyramid feature diversion structure with spatial attention mechanism is developed so that good feature representation with low information loss can be achieved by conducting channel attention-based feature diversion and inter-layer fusion, while reducing computational complexity. Also, our framework includes an Enhanced Spatial Attention module (ESA), which can improve the quality of initial polyp segmentation predictions through spatial self-attention mechanism and multi-scale feature fusion. Our approach is evaluated on five challenging polyp datasets— Kvasir, CVC-ClinicDB, CVC-300, ETIS, and CVC-colonDB and achieves excellent results. In particular, we achieve 94.2% Dice and 89.7% IoU on Kvasir, outperforming other state-of-the-art methods.

Submission Number: 34