CrossTransFuse:A Colorectal Polyp Segmentation Approach Using Multi-Scale Cross-Attention and GiFusion

This paper introduces a polyp segmentation method named CrossTransFuse, designed to tackle the challenges posed by the varying sizes, shapes, and blurred boundaries of colorectal polyps, which increase the risk of misdiagnosis. By implementing a multi-scale interaction strategy, CrossTransFuse captures the subtle variations and complex characteristics of polyps. The proposed GiFusion module integrates features from both the Mask R-CNN and Transformer branches and uses residual blocks for connection. These integrated features effectively capture both global and local contexts at the current resolution, enabling more precise segmentation of polyp boundaries and accurate target localization. Evaluations on five common polyp segmentation datasets demonstrate that CrossTransFuse outperforms advanced networks, achieving more accurate segmentation of polyp boundary regions.