Multichannel Fully Convolutional Network for Coronary Artery Segmentation in X-Ray Angiograms

Abstract

Accurate segmentation of coronary arteries in X-ray angiograms is an important step for the quantitative study of coronary artery disease. However, accurate segmentation is a challenging task because coronary arteries are thin tubular structures with relatively low contrast and the presence of artifacts. In this paper, a novel deep-learning-based method is proposed to automatically segment the coronary artery from angiograms by using multichannel fully convolutional networks. Since the artifacts appear in both live images (after the injection of contrast material) and mask images (before the injection of contrast material) and the blood vessels appear only in live images, we take the mask images into consideration to distinguish real blood vessel structures from artifacts. Therefore, both live images and mask images are used as multichannel inputs to provide enhanced vascular structure information. The hierarchical features are then automatically learned to characterize the spatial associations between vessel and background and are further used to achieve the final segmentation. In addition, a dense matching between the live image and mask image is processed for a precise initial alignment. The experimental results demonstrate that our method is effective and robust for coronary artery segmentation, compared with several state-of-the-art methods.

Keywords

• Coronary artery
• fully convolutional network
• dense matching
• U-net