ACA U-Net: Asymmetric Channel Attention Mechanism for Semantic Segmentation of Breast Ultrasound Images

According to the latest cancer data released by the World Health Organization, the number of new cases of breast cancer has surpassed lung cancer as the most prevalent cancer in the world. Early diagnosis and treatment is an important tool in the fight against breast cancer and is a data-proven viable way to reduce the risk of death. Deep learning-based medical image segmentation methods that can automatically segment lesions have been a hot topic for researchers in recent years. Although many studies have utilized local or non-local attention to improve segmentation performance, they tend to compute attention only on feature maps of the same level. In this paper, we propose a method for feature map recombination in the channel dimension, which uses Asymmetric Channel Attention with cross-level feature fusion to achieve better segmentation performance on ultrasound breast images. We evaluated on the BUSI dataset, our proposed Asymmetric Channel Attention U-Net (ACA UNet) achieves the best performance compared to other representative networks with little increase in computational cost.