Hybrid Encoded Attention Networks for Accurate Pulmonary Artery-Vein Segmentation in Noncontrast CT Images

Pulmonary artery-vein segmentation in computed tomography image is essential to lung disease diagnosis. It is still a challenge to segment small distal vessels, crossover and adhesion of arterioles due to complicated arteriovenous structures and limited computed tomography resolution. This work proposes a new U-shaped architecture of hybrid encoded attention networks that employ stacked hybrid units and 3D resample-free attention gates for automatic pulmonary artery-vein segmentation. Specifically, the hybrid unit uses normal operations, 3D hybrid easy operation, and channel shuffle to extract diverse features while the 3D resample-free attention gate can detect regions of interest such as pulmonary arteries and veins and suppress task-independent responses. We validated our method on 50 computed tomography volumes from LIDC-IDRI, with the experimental results demonstrating that it works more stable and effective than currently available approaches, improving the average dice similarity coefficients of the arteries and veins to (84.32%, 86.41%), respectively.