Semi-Supervised Domain-Adaptive Pulmonary Artery Segmentation Via Uncertainty Guidance And Shape Strengthening

Artery segmentation is often required for visualization and quantitative analysis of pulmonary diseases during surgical planning. By intravenous injection of contrast agent and carefully adjusting the trigger delay time, arteries or veins could be enhanced in computed tomography angiography (CTA). On the other hand, contrast between vessels and background is relatively low in non-contrast CT (NCCT), and it is difficult to distinguish arteries and veins. Additionally, manual or semi-automatic annotation of pulmonary vessels is time-consuming and labor-intensive. In this paper, we propose a novel semi-supervised domain-adaptive pulmonary artery segmentation framework for NCCT by using annotated CTA and a limited number of annotated NCCT images as training samples. Specifically, an uncertainty-driven Bayesian convolutional neural network-based adaptation module is proposed to predict the uncertainty map to guide pulmonary artery segmentation in NCCT. To explicitly learn tubular structures in different domains (i.e., CTA and NCCT), we propose a shape strengthening module (SSM), as a discriminator to strengthen vessel shape and boundaries in both CTA and NCCT domains. Experiments show that the proposed method performs better than state-of-the-art methods for pulmonary artery segmentation on NCCT images.