基于深度神经网络的冠状动脉造影图像血管分割和节段识别  

Establishment of an Artificial Intelligence Model of Vessel Segmentation and Segment Recognition on Coronary Angiography Using Deep Neural Network Algorithm

作　　者：谢丽华[1] 史晓彤 王筱斐 黄云飞 赵森祥 杜天明[2] 管常东 张洪刚[2] 徐波[1] XIE Lihua;SHI Xiaotong;WANG Xiaofei;HUANG Yunfei;ZHAO Senxiang;DU Tianming;GUAN Changdong;ZHANG Honggang;XU Bo(Catheterization Laboratory,National Center for Cardiovascular Diseases and Fuwai Hospital,CAMS and PUMC,Beijing(100037),China)

机构地区：[1]中国医学科学院北京协和医学院国家心血管病中心阜外医院介入导管室,北京市100037 [2]北京邮电大学人工智能学院 [3]北京红云智胜科技有限公司技术部

出　　处：《中国循环杂志》2020年第11期1064-1071,共8页Chinese Circulation Journal

基　　金：北京市科委医药协同创新研究基金资助项目(Z18110700190000);中国医学科学院医学与健康科技创新工程基金资助项目(2018-I2M-AI-007)。

摘　　要：目的:应用深度学习技术相关神经网络算法,创建冠状动脉造影图像血管分割和血管节段识别的人工智能模型。方法:纳入2018年7月于中国医学科学院阜外医院行冠状动脉造影患者2834例,共12900张冠状动脉造影图像。患者的冠状动脉造影图像由中国医学科学院阜外医院心血管介入诊疗影像分析核心实验室影像分析师标注。搭建一种创新的深度神经网络(DNN),分别进行冠状动脉造影图像血管分割和节段识别任务。在数据集中,11900张标注图像用于网络训练,1000张用于网络测试。以真实精标注图片为“金标准”,评价DNN对冠状动脉造影图像血管分割及节段识别的能力。结果:DNN对冠状动脉造影图像血管自动分割的平均准确度达99.2%(95%CI:99.1%~99.2%),F1分数为0.91±0.03,且对冠状动脉主支血管的分割结果优于一级分支血管。DNN对冠状动脉造影图像血管节段识别的平均准确度为98.6%(95%CI:98.6%~98.7%),F1分数为0.80±0.05,对冠状动脉主支血管段的识别效果优于一级分支血管,对血管近段的识别准确度优于血管远段。随着训练数据量增加,DNN对冠状动脉造影图像血管分割的性能明显提升。结论:该研究显示了DNN用于冠状动脉造影图像血管分割及节段识别的可行性,并获得了较高的准确度,为将来实现客观、高效的冠状动脉造影人工智能病变诊断提供了基础。Objectives:To establish an artificial intelligence model for coronary artery segmentation and vascular segment recognition on coronary angiographic images by using neural network algorithm related deep learning techniques.Methods:A total of 2834 patients,who underwent coronary angiography at Fuwai Hospital of the Chinese Academy of Medical Sciences in July 2018,were included,and 12900 angiographic pictures were collected.These data were annotated by certified angiogram analysts in angiographic core laboratory of Fuwai Hospital.Based on this dataset,a novel deep neural network(DNN)architecture was constructed,which was used to perform the segmentation and segment recognition on angiographic images.11900 annotated images out of this dataset were inputted to the segmentation network for model training,and 1000 images were used for network testing.The ability of DNN’s on correctly defining coronary artery segmentation and vascular segment recognition on angiographic images was evaluated by taking the fine-labeled pictures as the"gold standard".Results:The average accuracy of automatic segmentation of coronary vessels in angiographic images by the DNN was 99.2%(95%CI:99.1%-99.2%),and the F1 score was 0.91±0.03,and the evaluation result of coronary artery segmentation for the main vessel is superior to that of primary branch vessel.The average accuracy of vascular segment recognition in angiographic images was 98.6%(95%CI:98.6%-98.7%),and F1 score was 0.80±0.05.The result of vascular segment recognition on the main vessel is superior to that of the primary branch vessel,and the result of segment recognition on the proximal segments is also superior to that of the distal segments.In addition,the performance of the DNN improved in proportion with the increase of the amount of training data.Conclusions:Our results demonstrate the satisfactory capabilities of DNN based deep learning techniques in the field of coronary angiography segmentation.Our experience on this technique lays the foundation for subsequent computer-

关 键 词：深度神经网络 冠心病 冠状动脉造影 血管分割 血管节段识别 

分 类 号：R54[医药卫生—心血管疾病]