MAUNet: Polyp segmentation network based on multiscale feature fusion of attention U‐shaped network structure

Abstract

AbstractColorectal cancer is a prevalent malignant tumor affecting the digestive tract. Although colonoscopy remains the most effective method for colon examination, it may occasionally fail to detect polyps. In an effort to enhance the detection rate of intestinal polyps during colonoscopy, we propose MAUNet, a polyp segmentation network based on a multi‐scale feature fusion of an Attention U‐shaped network structure. Our model incorporates advanced components, including the Receptive Field Block, Reverse Attention Block, and Residual Refinement Module, mirroring the analytical process performed by medical imaging professionals. We evaluated the performance of MAUNet on five challenging datasets and conducted a comparative analysis against five state‐of‐the‐art models using six evaluation metrics. The experimental results demonstrate that MAUNet achieves varying levels of performance improvement across the five datasets. Particularly on the Kvasir dataset, the Mean Dice and Mean IOU metrics reached 91.6% and 84.3%, respectively, confirming the model's outstanding performance in polyp segmentation.