Key Laboratory for Biomedical Engineering of Ministry of Education, Institute of Biomedical Engineering, Zhejiang University, Hangzhou 310027, China.
Department of Radiology, School of Medicine, The Second Affiliated Hospital, Zhejiang University, Hangzhou 310009, China.
Math Biosci Eng. 2023 Jan 6;20(3):4988-5003. doi: 10.3934/mbe.2023231.
Coronary artery centerline extraction in cardiac computed tomography angiography (CTA) is an effectively non-invasive method to diagnose and evaluate coronary artery disease (CAD). The traditional method of manual centerline extraction is time-consuming and tedious. In this study, we propose a deep learning algorithm that continuously extracts coronary artery centerlines from CTA images using a regression method. In the proposed method, a CNN module is trained to extract the features of CTA images, and then the branch classifier and direction predictor are designed to predict the most possible direction and lumen radius at the given centerline point. Besides, a new loss function is developed for associating the direction vector with the lumen radius. The whole process starts from a point manually placed at the coronary artery ostia, and terminates until tracking the vessel endpoint. The network was trained using a training set consisting of 12 CTA images and the evaluation was performed using a testing set consisting of 6 CTA images. The extracted centerlines had an average overlap (OV) of 89.19%, overlap until first error (OF) of 82.30%, and overlap with clinically relevant vessel (OT) of 91.42% with manually annotated reference. Our proposed method can efficiently deal with multi-branch problems and accurately detect distal coronary arteries, thereby providing potential help in assisting CAD diagnosis.
冠状动脉中心线提取在心脏 CT 血管造影（CTA）中是一种有效的非侵入性方法，用于诊断和评估冠状动脉疾病（CAD）。传统的手动中心线提取方法既耗时又乏味。在这项研究中，我们提出了一种深度学习算法，该算法使用回归方法从 CTA 图像中连续提取冠状动脉中心线。在提出的方法中，使用 CNN 模块来提取 CTA 图像的特征，然后设计分支分类器和方向预测器来预测给定中心线点的最可能方向和管腔半径。此外，开发了一种新的损失函数来将方向向量与管腔半径相关联。整个过程从手动放置在冠状动脉口的一个点开始，直到跟踪到血管末端。网络使用包含 12 张 CTA 图像的训练集进行训练，并使用包含 6 张 CTA 图像的测试集进行评估。提取的中心线具有 89.19%的平均重叠（OV）、82.30%的首次错误重叠（OF）和 91.42%的与临床相关血管重叠（OT），与手动标注参考值相对应。我们提出的方法可以有效地处理多分支问题，并准确地检测到远端冠状动脉，从而为 CAD 诊断提供潜在的帮助。
