摘要: |
为提高基于超声心动图Simpson法的左心室射血分数(LVEF)测量的效率,提出了一种深度学习自动测量LVEF的方法。首先,建立卷积神经网络(CNN),利用收集的38153张标记的数据对网络进行训练测试和验证,将采集到的超声心动图数据自动分成五类,从而获取到心尖二腔切面(A2C)和心尖四腔切面(A4C);其次,建立全卷积网络,以VGG-19为主干架构,利用收集的3871张 A2C 和 4679张A4C数据进行训练验证,对自动获得的A2C和A4C的左心室进行自动分割,并计算得出LVEF。测试结果显示,该方法在得到A2C和A4C的准确率达96.8%,而分割真阳性率达到88.8%,所得LVEF误差率为0.038947。由于所提出的方法是完全利用机器去完成,较传统的方法精度和效率更高。 |
关键词: 左心室射血分数 超声心动图 深度学习 卷积神经网络 全卷积神经网络 |
DOI: |
投稿时间:2018-11-05修订日期:2018-11-13 |
基金项目:本项目受到国家重点研发计划(2017YFA0104302)和国家自然科学基金(61871126)的支持。 |
|
AUTOMATIC MEASUREMENT STUDY OF LEFT VENTRICULAR EJECTION FRACTION VIA DEEP LEARNING IN ULTRASOUND IMAGES |
jiangjianhui,luoshouhua |
(School of Biological Science and Medical Engineering, Southeast University) |
Abstract: |
To improve the efficiency of measuring left ventricular ejection fraction(LVEF)based on echocardiography, This paper proposes a new method for automatically calculating LVEF based on deep learning. Firstly, The convolution neural network (CNN), trained verified and tested by 38,153 marked echocardiography data, is used to divide the collected echocardiographic data into five categories and apical four-chamber view(A4C) and apical two-chamber view(A2C) are obtained. Then, the fully convolutional networks (FCN) ,using VGG-19 as the backbone architecture, trained and verified by collected 3871 A2C and 4679 A4C data, is used for segmenting the left ventricle of the two obtained views and the LVEF is obtained. Finally, The test result shows that the accuracy of obtaining the A4C and A2C is 96.8% and the true positive rate of segmentation is over 88.8%. A error is 0.038947 between the automatic and manual. The proposed method is calculated by the machine, which is more accurate and efficient than the conventional method. |
Key words: left ventricular ejection fraction, echocardiography, deep learning, convolution neural network, fully convolutional networks |