Joint Multi-Scale and Dual Attention Gate Network for Pulmonary Vessel Segmentation

Abstract

Pulmonary vessel segmentation is an important precondition for diagnosis, quantitative analysis and treatment planning of pulmonary diseases. Recently, Convolutional Neural Networks (CNNs) has achieved great success in biomedical image processing. However, due to the complex structure, inter-vessel differences and clinical lesions, pulmonary vessel segmentation still faces great challenges. In this work, we propose a novel end-to-end pulmonary vessel segmentation network named Joint Multi-Scale and Dual Attention Gate Network (JMSDAGNet). The JMSDAGNet aggregate feature maps of different scales, which can effectively improve the segmentation accuracy of thin vessels. To focus on vascular areas and suppress the noise caused by diseases, we introduce attention mechanisms that can adaptively learn local and global information. We trained and validated the proposed JMSDAGNet in a custom dataset. Both quantitative and qualitative results of the comprehensive experiments prove the superiority of our proposed method in comparison with the state-of-the-art methods.