Contextual Attention Network: Transformer Meets U-Net

Convolutional neural networks (CNN) (e.g., UNet) have become the de facto standard and attained immense success in medical image segmentation. However, CNN based methods fail to build long-range dependencies and global context connections due to the limited receptive field of the convolution operation. Therefore, Transformer variants have been proposed for medical image segmentation tasks due to their innate capability of capturing long-range correlations through the attention mechanism. However, since Transformers are not designed to capture local information, object boundaries are not well preserved, especially in difficult segmentation scenarios with partly overlapping objects. To address this issue, we propose a contextual attention network that includes a boundary representation on top of the CNN and Transformer features. It utilizes an CNN encoder to capture local semantic information and includes a Transformer module to model the long-range contextual dependency. The object-level representation is included by extracting hierarchical features that are then fed to the contextual attention module to adaptively recalibrate the representation space using local information. In this way, informative regions are emphasized while taking into account the long-range contextual dependency derived by the Transformer module. The results show that our approach is amongst the top performing methods on the skin lesion segmentation benchmark, and specifically shows its strength on the SegPC challenge benchmark which also includes overlapping objects. Implementation code in github.