Dual-Path U-Net for Polyp Segmentation

Polyp segmentation aims to detect the polyp on a given image. A high accurcy of polyp detection helps to increase the survival rate of colorectal cancer. This paper presents a new deep learning based network, namely Dual-Path U-Net to perform polyp segmentation. It combines two U-Net in dual-path architecture with shortcut connections from both encoders to each decoder. The proposed method adopts the dilated convolution to learn feature maps with larger receptive fields. It also incorporates the squeeze-and-excitation to further improve the quality of feature maps. As documented in Experimental Results, the proposed schemes produces better mean F1-Score and mean IoU compared to the other competing methods.