CSUnet: a dual attention and hybrid convolutional network for polyp segmentation

Medical polyp segmentation is a key task in the field of computer vision and medical image analysis, which aims to automatically detect and segment polyps in medical images for early cancer screening and treatment planning. Neural networks, in particular convolutional neural networks (CNNS), play a key role in this task. We present CSUNet, a traditional U-shaped encoder-decoder network architecture specifically designed for polyp segmentation. Firstly, the hybrid convolution mechanism was used to effectively capture the features of different scales and improve the segmentation performance. In addition, an attention module is added to the skip connection to facilitate efficient feature map extraction and fusion. We trained and verified the network with two public datasets: Kvasir-SEG and CVC-ClinicDB. In the polyp segmentation task, the Mean Intersection over Union value of CSUNet on these datasets reached 87.78% and 92.07%, respectively. The effectiveness of the proposed method enables the model to show strong segmentation performance and excellent visual results. Experiments on the collected public large-scale datasets are better than other state-of-the-art methods.