Polyp Segmentation Achieved Through Multi-Scale Feature Fusion Based on Attention Mechanism

Given the high prevalence of colon diseases, it is crucial to utilize computer vision technology for precise polyp segmentation during colonoscopy. However, manual detection becomes increasingly challenging due to the similarity in texture and color between polyp tissue and surrounding normal tissue, leading to higher risks of false positives and missed detections. This study introduces a novel network called CabaNet that prioritizes boundary awareness through coordinate attention. The network consists of four main modules: an encoder utilizing pyramid vision transformer PVT, a cascade part decoder employing Parallel hierarchical aggregation decoder (PHAD) for feature fusion and obtaining a comprehensive semantic global map. Additionally, the refinement pyramid module (RP) enhances feature fusion through a lightweight multi-scale learning strategy. The reverse coordinate attention module (Coo-A) combines features from PHAD and RP to achieve precise and complete edge prediction mapping. Experimental results on five commonly used datasets demonstrate significant improvement in image segmentation accuracy by our proposed model.