Dual Attention Cascade Transformer for Polyp Segmentation

Yuan Ju, Bingxin Wei, Haewoon Nam

Abstract: Polyp segmentation is of great importance for the prevention of colon cancer. Such a task remains highly challenging due to the high similarity of polyps to the background. Models based on convolutional neural networks and transformer have achieved promising results in this task, but their ability to combine local and global information remains limited. In this paper, we propose a novel network Dual Attention Cascade Transformer (DACFormer) that effectively combines local and global contextual information to suppress the effect of context on target recognition. The proposed method adopt the cascade structure of feature reuse, which effectively combines the semantic information of features at all levels and further exploits the potential of transformer and improves the generalization ability of the model is effectively improved. We conducted tests on four public datasets, CVC-ClinicDB, Kvasir, CVC-300, CVC-ColonDB. The results show that our network outperforms the current mainstream networks on the four benchmark datasets.