Joint Attention for Medical Image Segmentation

Medical image segmentation is crucial for computer aided diagnosis. In recent years, spatial attention mechanisms have leaded to breakthroughs in the task of image segmentation. In this work, we firstly present a unified formula for spatial attention mechanisms. Within this framework, we find that point-wise attention has better localization while self-attention can learn more global features. Motivated by this observation, we then propose a new joint attention module, which jointly leverages the advantages of point-wise attention and self-attention. Moreover, by integrating joint attention with DenseUNet, we conduct image segmentation experiments on two public datasets. The proposed method outperforms recent state-of-the-art models, verifying the superiority of joint attention. Additionally, ablation studies demonstrate that our joint attention obtains more balanced results compared to the previous point-wise attention and self-attention. The design of joint attention provides a novel insight into understanding spatial attention mechanisms.