Automated Vessel Segmentation in CT and CTA scans of the Lung Via Deep Neural Networks

Background: The distribution of pulmonary vessels in computed tomography images is important for diagnosing disease, formulating surgical plans and pulmonary research. However, there are many challenges of pulmonary vascular segmentation due to its characteristics of narrow and long pipes, discrete distribution and tree-like structure. With the development of deep learning and medical image processing, an automatic, accurate and fast segmentation algorithm of pulmonary blood vessels becomes possible. Methods: Based on the International Symposium on Image Computing and Digital Medicine 2020 challenge pulmonary vascular segmentation task, this paper objectively evaluates the performance of 12 different algorithms in chest computed tomography and computed tomography angiography. First, we present the annotated reference dataset including computed tomography and computed tomography angiography. Second, by analyzing the advantages and disadvantages of each team’s algorithm from 12 different institution, the reasons for some defects and improvements are summarized. Finally, we discuss the ways and methods to improve the results. Results: These methods were compared with the ground truth by the numerical results and the intuitive results from computed tomography and computed tomography angiography images. Most methods do an admirable job in pulmonary vascular extraction, with dice coefficients ranging from 0.70 to 0.85, and the dice coefficient for the top three methods are about 0.80. Conclusions: These results show that the methods which consider spatial information, fuse multi-scale feature map, or have an excellent post-processing are significant for further improving the accuracy of pulmonary vascular segmentation. Keywords: segmentation; pulmonary; vessel; U-Net; network; CT images; CTA