Abstract: Owning to the complex morphological structure of the coronary artery and the variations in the acquisition parameters of Computed Tomography (CT) angiography images, the influence of image quality issues such as uneven distribution of image gray scale, motion artifacts and noise results in inaccurate segmentation of the coronary artery structure, including missed segments and misjudgments. Segmentation network of coronary artery structure from CT angiography images based on multi-scale spatial features named 3D Multi Scale Parallel Net (MSP-Net) was proposed. Firstly, in view of the characteristics of large spatial span and small local proportion of coronary artery, a multi-scale parallel fusion network was used to extract global features and local features from CT angiography(CTA) for fusion to ensure the extraction of coronary artery structure features. Then, the coronary artery structure was reconstructed using the fusion method of different scale features to enhance the accuracy of segmentation results and reduce missed and misjudgments. Finally, supervision signal was adopted at different network depths to improve training efficiency. The experimental results show that the average Dice Similarity Coefficient(DSC) of the proposed method in the coronary artery automatic segmentation task reaches 87.16%, which is 4.04 percentage points higher than that of nnU-Net and 2.31 percentage points higher than that of Swin UNETR, and the average 95% Hausdorff Distance (HD95) reaches 3.69mm. It is 14.43mm lower than nnU-Net and 13.75mm lower than Swin UNETR. It has been demonstrated that the proposed network can effectively improve the segmentation accuracy of coronary artery structure, and help clinicians to more accurately interpret the coronary artery structure of patients, so as to evaluate the disease more effectively.

Key words: coronary artery segmentation, Computed Tomography Angiography(CTA), computer-aided diagnosis, multi-scale fusion, semantic segmentation

摘要： 针对冠状动脉血管形态结构的复杂性以及由于计算机断层扫描（CT）血管造影图像的采集条件差异，导致图像灰度分布不均匀、运动伪影和噪声等图像质量问题，进而产生冠状动脉结构分割时的漏判、误判的问题，提出一种基于多尺度空间特征的冠状动脉CT血管造影图像分割网络——三维多尺度并行网络（3D Multi Scale Parallel Net, MSP-Net）。首先，针对冠状动脉在空间跨度大、局部占比小的特点，采用多尺度网络分别提取冠状动脉CT血管造影图像的全局特征和局部特征进行融合，保证冠状动脉结构特征完整性的提取。其次，冠状动脉重建采用由粗到细的思维，增强图像特征冗余性，确保冠脉边界分明，后利用不同尺度特征度融合方法重建冠状动脉结构以加强分割结果的准确率，减少漏判和误判；最后，为加快网络的训练过程，采用深监督策略，在不同网络层级引入监督信号，从而提高训练效率。实验结果表明，所提网络在冠状动脉自动分割任务中的平均Dice相似系数（DSC）达到87.16%，比nnU-Net提高4.04个百分点，比Swin UNETR提高了2.31个百分点，平均95%的豪斯多夫距离（HD95）达到3.69mm，比nnU-Net降低了14.43mm，比Swin UNETR降低了13.75mm。该网络能有效提高冠状动脉结构分割精度，有助于临床医生更准确地解患者冠状动脉结构，从而更有效地评估病情。

关键词: 冠状动脉分割, 计算机断层扫描血管造影, 计算机辅助诊断, 多尺度融合, 语义分割