Joint Segmentation and Quantification of Main Coronary Vessels Using Dual-Branch Multi-scale Attention Network

• Published in: Medical Image Computing and Computer Assisted Intervention - MICCAI 2021 Vol. 12901; pp. 369 - 378
• Main Authors: Zhang, Hongwei, Zhang, Dong, Gao, Zhifan, Zhang, Heye
• Format: Book Chapter
• Language: English
• Published: Switzerland Springer International Publishing AG 2021; Springer International Publishing
• Series: Lecture Notes in Computer Science
• ISBN: 3030871924; 9783030871925
• ISSN: 0302-9743; 1611-3349
• DOI: 10.1007/978-3-030-87193-2_35

Joint segmentation and quantification of main coronary vessels are important to the diagnosis and intraoperative treatment of coronary artery disease. They can help clinicians decide whether to carry out coronary revascularization and choose the interventional stent. However, joint segmentation and quantification in a framework is still challenging because of intrinsic distinction of optimization objects for these two tasks. In this paper, we propose a dual-branch multi-scale attention network (DMAN) to achieve synergistic optimization process in a framework. Our DMAN consists of a nested residual module and a attentive regression module. The nested residual module is used to extract and aggregate multi-level and multi-scale features. The attentive regression module introduces a two-phase attention block to express interactive correlation of separated regions and capture the informativeness of the important region in the image. Our DMAN is evaluated over 1893 X-ray coronary angiography images collected from 529 subjects. DMAN achieves the dice coefficient of 0.916 for segmentation and the MAE of 1.30 ± 0.62 mm for quantification.