FCAU-Net: A Frequency Channel Attention Convolutional Neural Network for Medical Image Segmentation

Recently, U-Net and its subsequent extensions such as Attention U-Net have emerged as the leading medical image segmentation methods. However, many approaches that combine attention mechanisms with U-Net have overlooked a fundamental issue, that is, they fail to represent channel attention mechanisms using scalars. In this work, we propose a novel method called FCAU-Net, which extends the compression of channel attention mechanism to the frequency domain that combines the multi-spectral attention module with U-Net for medical image segmentation. FCAU-Net has a stronger ability to learn significant features specific to local regions than classic U-Net. In addition, we introduce a new inception block that decomposes the large kernel depth-wise convolution of the inception architecture with two parallel branches of deep convolutions, aiming to further enhance the model's feature representation and semantic information acquisition capabilities while reducing the number of parameters added by the large convolutional kernel. Experiments on multiple medical image segmentation datasets demonstrate that our method achieves better segmentation performance.