GAU-Net: U-Net Variant for Segmentation of Skin Lesion

Image segmentation of skin lesions is a critical component in the auxiliary diagnosis of medical-related diseases. However, traditional U-Net network methods struggle to achieve accurate segmentation results when presented with high-resolution skin lesion images, especially under small batch size training conditions. In this study, we propose a novel GAU-Net network that leverages the attention mechanism of mobile networks. Based on U-Net, the method employs Group Normalization and LeakyReLU activation functions in the convolution block to obtain normalized and reliable features, thereby enhancing the network’s ability to represent high-level semantic information. Additionally, the proposed method introduces Coordinate Attention to embed effective position information into the channel Attention, further improving the model’s segmentation results. We evaluated the segmentation performance of GAU-Net on the ISIC2016 skin disease dataset, showing that it outperformed traditional segmentation indicators such as mIoU and mAP. These results demonstrate the superiority of GAU-Net as an effective solution for skin lesion image segmentation.