Fourier Channel Attention Powered Lightweight Network for Image Segmentation

The accuracy of image segmentation is critical for quantitative analysis. We report a lightweight network FRUNet based on the U-Net, which combines the advantages of Fourier channel attention (FCA Block) and Residual unit to improve the accuracy. FCA Block automatically assigns the weight of the learned frequency information to the spatial domain, paying more attention to the precise high-frequency information of diverse biomedical images. While FCA is widely used in image super-resolution with residual network backbones, its role in semantic segmentation is less explored. Here we study the combination of FCA and U-Net, the skip connection of which can fuse the encoder information with the decoder. Extensive experimental results of FRUNet on three public datasets show that the method outperforms other advanced medical image segmentation methods in terms of using fewer network parameters and improved accuracy. It excels in pathological section segmentation of nuclei and glands.