实用医学杂志 ›› 2022, Vol. 38 ›› Issue (14): 1830-1833.doi: 10.3969/j.issn.1006⁃5725.2022.14.022

• 新技术新方法 • 上一篇    下一篇

深度学习技术在胎儿超声心动图图像自动识别中的应用

罗刚1 泮思林1 乔思波3 庞善臣3 陈涛涛2 孙玲玉2 董玉坤3   

  1. 青岛大学附属妇女儿童医院1 心脏中心，2 超声科（山东青岛 266034）；3 中国石油大学（华东）计算机科学与技术学院（山东青岛 266580）
  • 出版日期:2022-07-25 发布日期:2022-07-25
  • 通讯作者: 泮思林 E⁃mail: silinpan@126.com
  • 基金资助:
    国家自然科学基金（编号：81970249）；泰山学者工程资助（编号：2018）

Deep learning technology for automatic recognition of fetal echocardiography images 

LUO Gang*，PAN Si⁃lin，QIAO Sibo，PANG Shanchen，CHEN Taotao，SUN Lingyu，DONG Yukun.   

  1. Heart Center，Women and Chil⁃dren′s Hospital，Qingdao University，Qingdao 266034，China

  • Online:2022-07-25 Published:2022-07-25
  • Contact: PAN Silin E⁃mail：silinpan@126.com

摘要：

目的 探讨深度学习技术在胎儿超声心动图图像自动识别中应用的可行性。方法 对YOLOv4模型进行改进，引入多级残差混合注意力机制模块（MRHAM），建立 MRHAM⁃YOLOv4⁃Slim 模型。选取青岛大学附属妇女儿童医院收集的 2000 张标准胎儿超声心动图四腔心切面图片建立实验数据集，将MRHAM⁃YOLOv4⁃Slim 与多种学习模型进行图像识别的分析比较，验证该模型的有效性。结果 本研究建立的学习模型能够更精确识别图像中的心腔结构，准确率为 0.85，召回率为 0.92，F1 分数为 0.88，平均精度为 0.910。该模型具体识别左心房、右心房、左心室和右心室的准确度分别为 0.87、0.93、0.86 和 0.89。结论 本研究建立 MRHAM⁃YOLOv4⁃Slim 模型性能优越，可更准确的识别四腔心切面中心腔结构，接近超声医师识别水平，为人工智能在胎儿超声心动步图中的进一发展奠定基础。

关键词: 深度学习, 人工智能, 胎儿, 超声心动描记术

Abstract:

Objective To explore the feasibility of deep learning technology for automatic recognition of fetal echocardiography images. Methods The detection model of YOLOv4 was improved and the multi residual hybrid attention module（MRHAM）was introduced，which was named MRHAM⁃YOLOv4⁃Slim. A total of 2000 fetal echocardiographic four chamber cardiac views at the Women and Children′ s Hospital，Qingdao University were selected to establish the experimental datasets. MRHAM⁃YOLOv4⁃Slim was compared with various artificial intelligence models for image recognition. Results MRHAM⁃YOLOv4⁃Slimaccuratelyidentified the cardiac cavity structure in the four chamber view. The accuracy rate was 0.85，the recall rate was 0.92，the F1 score was 0.88，and the average accuracy was 0.91. The accuracy of the model in identifying left atrium，right atrium，left ventricle
and right ventricle was 0.87，0.93，0.86 and 0.89 respectively. Conclusion The performance of MRHAM ⁃YO⁃LOv4⁃Slim model was better than most of the models，which identified theheartstructures in the four chamber view more accurately. Its recognition level was close to that of ultrasound doctors. This study would contribute to the further development of artificial intelligence in fetal echocardiography.

Key words: deep learning, artificial intelligence, fetus, echocardiography

粤ICP备18046265号
版权所有 © 2019 《实用医学杂志》编辑部
地址：广州市越秀区惠福西路进步里2号之6　
邮编：510180　电话：020-81866302；81872080；81840509；81922330　
E-mail： syyxzz@syyxzz.com
本系统由北京玛格泰克科技发展有限公司设计开发