Journal of Frontiers of Computer Science and Technology ›› 2021, Vol. 15 ›› Issue (5): 958-970.DOI: 10.3778/j.issn.1673-9418.2005007

• Graphics and Image • Previous Articles    

Global Feature and Multi-level Feature Aggregation Segmentation Algorithm for Coronary

GU Jia, FANG Zhijun, TIAN Fangzheng   

  1. School of Electronic and Electrical Engineering, Shanghai University of Engineering Science, Shanghai 201600, China
  • Online:2021-05-01 Published:2021-04-30

全局特征及多层次特征聚合的冠脉分割算法

顾佳方志军田方正   

  1. 上海工程技术大学 电子电气工程学院,上海 201600

Abstract:

Coronary computed tomographic angiography (CTA) image segmentation plays an important role in many practical applications, such as assisting doctors to judge vascular occlusion, vascular disease diagnosis, etc. In view of the large amount of noise in CTA images and not delicate segmentation results of traditional deep learning algorithms (including FCN, U-Net, V-Net, etc.), this paper proposes a global feature and multi-level feature aggregation network, which includes three network models, global feature module, feature fusion and refined V-shape multi-level feature aggregation module, and deep supervision module. The global feature module can filter the original CTA image and generate the basic features by integrating the early and later feature information and integrating the rich details and semantic information. The refined V-shape module generates refined feature maps of different levels on the basis of basic features, and obtains accurate coronary segmentation images by aggregating the refined feature maps of different levels. In addition, a deep supervision mechanism is added after each refined V-shape module to avoid the problem of gradient disappearing. The results show that the proposed method is superior to the mainstream baseline intuitively and quantitively. The ablation experiments also prove the effectiveness of each module.

Key words: coronary computed tomographic angiography (CTA) segmentation, global feature, refined V-shape, multi-level feature aggregation, deep supervision

摘要:

冠脉计算机断层扫描血管造影(CTA)图像分割在辅助医生判断血管堵塞、血管疾病诊断等许多实际应用中发挥重要作用。针对CTA图像中存在大量噪声和FCN、U-Net、V-Net等经典深度学习算法分割结果不细腻的问题,提出了全局特征及多层次特征聚合网络。这种新型的网络由全局特征模块、特征融合与V形细化多层次特征聚合模块以及深度监督三部分组成。全局特征模块综合早期和后期特征信息,在融合丰富的细节和语义信息基础上实现对原始CTA图像过滤操作,生成基础特征。细化V形模块在基础特征的基础上生成不同层次的细化特征图,通过聚合不同层次的细化特征图,得到精准冠脉分割图像。此外,在每一个细化V形模块之后加入深度监督机制来避免梯度消失的问题。对提出的方法进行了定量与定性的分析,结果表明,该方法优于主流基线。消融实验也证明了每个模块的有效性。

关键词: 冠脉计算机断层扫描血管造影(CTA)分割, 全局特征, 细化V形, 多层次特征聚合, 深度监督