The TGDouble-UNet Network for Polyp Segmentation

Accurate segmentation of polyps in colorectal endoscopic images enables the early detection of potential precancerous lesions in the colon and rectum. However, the similarity in texture and color between polyps and the background poses a significant challenge to precise segmentation. To improve the accuracy of polyp segmentation in colorectal endoscopic images, this paper proposes the TGDouble-UNet network model, which features a cascaded encoder-decoder architecture. To address the limitation of traditional CNN encoders, which focus only on local image features, a TG Block module composed of group convolution and Transformer is proposed. This allows the model to capture both global and local features of the polyps. To alleviate the issue of inaccurate decoding caused by information loss during the encoding process, the BranchUnet structure is introduced to extract second-level encoded features. Finally, the polyp segmentation result is obtained by combining the encoded features from both levels. To validate the effectiveness and advancements of the model, quantitative and result visualization experiments were conducted on the publicly available datasets CVC-ClinicDB and Kvasir-SEG. The experimental results show that the proposed TGDouble-UNet model achieved a Dice coefficient of 0.9267 and a mIoU of 0.8727 on CVC-ClinicDB, and a Dice coefficient of 0.9199 and an mIoU of 0.8647 on Kvasir-SEG. Compared to the currently well-performing Double-UNet, the TGDouble-UNet improved the Dice and mIoU by 0.28% and $\mathbf{1. 1 6 \%}$ on CVC-ClinicDB, and by $\mathbf{1. 3 5 \%}$ for both Dice and mIoU on Kvasir-SEG.