Nonlinear Activation-Free Contextual Attention Network for Polyp Segmentation

The accurate segmentation of colorectal polyps is great significance for the diagnosis and treatment cancer. However, faces complex problems such as low contrast in peripheral region salient images, blurred borders, diverse shapes. In addition, number traditional UNet network parameters large effect average. To overcome these problems, an innovative nonlinear activation-free uncertainty contextual attention proposed this paper. Based on network, encoder a decoder are added to predict saliency map each module bottom-up flow pass it next module. We use Res2Net backbone extract image features, enhance features through simple parallel axial channel attention, obtain high-level with global semantics low-level edge details. At same time, n on-activation introduced, which can reduce complexity between blocks, thereby further enhancing feature extraction. This work conducted experiments five commonly used polyp datasets, experimental evaluation metrics from mean intersection over union, Dice coefficient, absolute error were all improved, show that our method has certain advantages existing methods terms performance generalization performance.