Pulmonary Vessel Segmentation via Stage-Wise Convolutional Networks With Orientation-Based Region Growing Optimization

Abstract

For large-scale medical images, automatic pulmonary vessel segmentation is a fundamental and essential research for various pulmonary disease diagnoses. Most existing approaches have made much progress in vessel segmentation, but the accuracy still remains unsatisfactory due to the absence of discriminative features extracted from the images. To address this problem, we propose novel stagewise convolutional networks followed by an orientation-based region growing method. The stage-wise convolutional networks aim to learn discriminative features of pulmonary vessels automatically in a stageby-stage manner, where stage I is the lung segmentation module and stage II is the main vessel segmentation module. In stage I, the lung segmentation module extracts pulmonary regions based on the convolutional neural networks to preprocess computed tomography (CT) scans and provides a good initial value for subsequent work. In stage II, the main vessel segmentation module exploits refined fully convolutional networks to hierarchically learn rich representations for pulmonary vessels, which enables accurate vessel segmentation. In addition, we further propose an optimization module that refines the results from the previous module based on the orientation of vessels in 3D space. Extensive experiments demonstrate that the proposed method has achieved the best performance in pulmonary vessel segmentation compared with the state of the arts.

Keywords

• Automatic
• pulmonary vessel segmentation
• convolutional networks
• region growing