Annu Int Conf IEEE Eng Med Biol Soc. 2021 Nov;2021:3846-3849. doi: 10.1109/EMBC46164.2021.9629655.
Coronary artery extraction in cardiac CT angiography (CCTA) image volume is a necessary step for any quantitative assessment of stenoses and atherosclerotic plaque. In this work, we propose a fully automated workflow that depends on convolutional networks to extract the centerlines of the coronary arteries from CCTA image volumes, starting from identifying the ostium points and then tracking the vessel till its end based on its radius and direction. First, a regression U-Net is employed to identify the ostium points in the image volume, then these points are fed to an orientation and radius predictor CNN model to track and extract each artery till its end point. Our results show that an average of 96% of the ostium points were identified and located within less than 5mm from their true location. The coronary arteries centerlines extraction was performed with high accuracy and lower number of training parameters making it suitable for real clinical applications and continuous learning.
在心脏 CT 血管造影 (CCTA) 图像体积中提取冠状动脉是任何狭窄和动脉粥样硬化斑块定量评估的必要步骤。在这项工作中，我们提出了一种完全自动化的工作流程，该流程依赖于卷积网络从 CCTA 图像体积中提取冠状动脉的中心线，从识别开口点开始，然后根据其半径和方向跟踪血管直到其末端。首先，使用回归 U-Net 识别图像体积中的开口点，然后将这些点输入到方向和半径预测器 CNN 模型中，以跟踪和提取每条动脉，直到其端点。我们的结果表明，平均有 96%的开口点被识别出来，并且它们的位置距离真实位置不到 5mm。冠状动脉中心线提取具有很高的准确性和较少的训练参数，使其适合实际的临床应用和持续学习。
