Application of a Selective Attention Network Based on a Mixed Attention Mechanism in a U-Net Decoder

The traditional Selective Kernel Network cannot deeply explore the multidimensional relationship between features.To handle the problem, this paper design a novel attention mechanism, which fuses the spatial attention and channel attention.And we design a new average attention mechanism to ensure balanced attention to obtain global features.The proposed attention module can learn the features adaptively, named as SANet(Selective Attention Network), which is successfully applied to the decoding part of the U-Net model.The experimental results show that our proposed SANet can further improve the accuracy and sensitivity of medical image segmentation by strengthening the feature selection and fusion ability while retaining the original advantages.