A perception network for improved segmentation of endoscopic polyp images with enhanced detail

Colorectal cancer is a prevalent and highly lethal malignant tumor disease of the digestive tract worldwide. The automatic and accurate segmentation of polyp regions in colorectal tissue pathology images is vital in the early diagnosis of CRC. However, polyp segmentation poses significant challenges due to the diverse sizes, colors, and textures of polyps and indistinct boundaries with surrounding tissues. To tackle these problems, we designed a perception network to improve polyp segmentation by enhancing detail representation. The network aggregates fine grained features through a perception module to generate a shape distribution map containing edge information. The shape distribution map supplements the spatial information to the shape distribution guidance module for polyp segmentation. Hidden polyps with low contrast are detected using content guided attention in the jump joining phase to enhance polyp lesion feature representation. Moreover, convolution with enhanced detail and an attention feature fusion mechanism is used in the perception module to improve the network’s ability to perceive polyp lesions with diverse shapes, further enhancing colonic tissue pathology image segmentation accuracy. Extensive experiments on five public polyp benchmark datasets demonstrate that our approach outperforms existing methods while maintaining low complexity.