STAU-Net: A Spatial Structure Attention Network for 3D Coronary Artery Segmentation

Automated segmentation of coronary artery is critical yet challenging for the detection and quantification of cardiovascular diseases. Considering the limitation of computing power, most existing 3D coronary artery segmentation methods divide original data into patches or 2D slices for segmentation to support the limited GPU memory, thereby causing limited segmentation performance due to the loss of contextual information of coronary artery structure. To solve above issues, this paper proposes a novel model for 3D coronary artery segmentation by enhancing structural information of features. Specifically, the proposed framework consists of a structure attention fusion (STAF) block and up-sample fusion (UF) block. The STAF block utilizes channel attention and spatial attention to enhance the fused feature maps from the output of dilated convolution at adjacent scales, and the UF block offsets the loss contextual information by fusing the feature map of the upper decoder. Also, the framework first resamples the input to a fixed size to implement training and up-sample to original size by customized post-processing at output stage. Compared with other related segmentation networks, the results demonstrate that our method can segment more detailed information of coronary artery tree and achieve better performance.