GCCSwin-UNet: Global Context and Cross-Shaped Windows Vision Transformer Network for Polyp Segmentation

Abstract

Accurate polyp segmentation is of great importance for the diagnosis and treatment of colon cancer. Convolutional neural networks (CNNs) have made significant strides in the processing of medical images in recent years. The limited structure of convolutional operations prevents CNNs from learning adequately about global and long-range semantic information interactions, despite the remarkable performance they have attained. Therefore, the GCCSwin-UNet framework is suggested in this study. Specifically, the model utilizes an encoder–decoder structure, using the patch-embedding layer for feature downsampling and the CSwin Transformer block as the encoder for contextual feature extraction. To restore the feature map’s spatial resolution during upsampling operations, a symmetric decoder and patch expansion layer are also created. In order to help the backbone module to do better feature learning, we also create a global context module (GCM) and a local position-enhanced module (LPEM). We conducted extensive experiments on the Kvasir-SEG and CVC-ClinicDB datasets, and compared them with existing methods. GCCSwin-UNet reached remarkable results with Dice and MIoU of 86.37% and 83.19% for Kvasir-SEG, respectively, and 91.26% and 84.65% for CVC-ClinicDB, respectively. Finally, quantitative analysis and statistical tests are applied to further demonstrate the validity and plausibility of our method.