BAG-Net: A Novel Architecture for Enhanced Medical Image Segmentation with Global Context Attention and Boundary Self-Attention
Yuyang Lei, Shengxian Yan, Jing Zhang, Xiang Li, Penghui Wang, Xiao Zhi Gao, Hui Cao · Symmetry · 2025
U-Net and its various variants have been widely applied in medical image segmentation in recent years, and significant success has been achieved in addressing complex segmentation tasks. These networks excel in feature extraction and enable efficient identification of key structural features in diverse medical images. However, convolutional neural networks face limitations during feature extraction, especially when modeling long-range contextual dependencies. This limitation hinders their ability to capture global features and may cause a decline in performance for complex segmentation tasks. To address these challenges, a novel architecture called BAG-Net (Boundary And Global Attention Network) is proposed that integrates global contextual information with local features more effectively. The network includes a global context attention component, which helps model long-range contextual features. Furthermore, a U-Net variant is created by introducing SE-Net into the skip connections in order to extract local information. In addition, a boundary self-attention component is employed to capture boundary details. The combined effect of these three components enables BAG-Net to fully exploit both local and global information and achieve high-precision segmentation. Experimental results show that BAG-Net outperforms traditional methods across all performance metrics. Thus, new perspectives for the advancement of medical image segmentation techniques are offered, and this provides a valuable reference for clinical applications.