2023 , Vol. 20 >Issue 04: 424 - 429
DOI: https://doi.org/10.3877/cma.j.issn.1672-6448.2023.04.008
基于深度学习的超声心动图自动识别节段性室壁运动异常的研究
Copy editor: 汪荣
收稿日期: 2021-11-02
网络出版日期: 2023-08-07
基金资助
北京市自然科学基金(7202198)
国家自然科学基金(82202265)
版权
Automatic detection of regional wall motion abnormalities by echocardiography based on deep learning
Received date: 2021-11-02
Online published: 2023-08-07
Copyright
探讨基于深度学习算法的超声心动图自动识别节段性室壁运动异常的效能。
本研究回顾性收集了2015年6月至2019年9月在解放军总医院第四医学中心门诊及住院患者的超声心动图2274例作为训练集和验证集,其中包括心肌梗死患者1137例;另于2021年3月至2021年5月前瞻性收集1324例连续性超声心动图影像作为测试集,其中包括105例心肌梗死患者。本研究分为三个步骤,包括切面识别、左心室心肌分割以及室壁运动异常检测,并进一步比较了模型输入多个切面与输入单个切面对节段性室壁运动异常识别效能的差异。
本研究神经卷积网络模型,对心尖四腔心切面(A4C),心尖两腔心切面(A2C)和心尖三腔心切面(A3C)的识别准确性分别为95%、98%、94%。心尖三个切面对左心室内膜分割的准确性均优于对心外膜的分割,且对心尖四腔心切面的分割准确性最佳(89.16%)。无论在内部验证集,还是外部测试集中,模型输入心尖三个切面对节段性室壁运动异常的识别效能均优于仅输入心尖四腔心单切面(ROC曲线下面积:0.942 vs 0.897;0.937 vs 0.828)。
深度学习技术不仅可以自动识别超声心动图动态视频图像,并且可以识别节段性室壁运动异常,深度学习模型可以应用于临床实践,有助于提高超声的诊断效率。
杨菲菲 , 林锡祥 , 陈亦新 , 王秋霜 , 张丽伟 , 陈煦 , 张梅青 , 王淑华 , 何昆仑 . 基于深度学习的超声心动图自动识别节段性室壁运动异常的研究[J]. 中华医学超声杂志(电子版), 2023 , 20(04) : 424 -429 . DOI: 10.3877/cma.j.issn.1672-6448.2023.04.008
To investigate the efficiency of a deep learning (DL) framework to automatically analyze echocardiographic videos in detecting regional wall motion abnormalities.
A total of 2274 echocardiographic videos of outpatients and inpatients in the Fourth Medical Center of the PLA General Hospital from June 2015 to September 2019 were retrospectively collected as training and validation datasets, and 324 consecutive echocardiographic videos were prospectively collected as the test set, including 105 patients with myocardial infarction. We developed a three-stage DL framework, including 1) view classification, 2) left ventricular myocardial segmentation, and 3) regional wall motion abnormality detection. The difference in the recognition efficiency of segmental wall motion anomalies between the model based on videos in multiple views and those in a single view was then compared.
The classification accuracy of the neural convolutional network model was 95%, 98%, and 94% for apical four chamber view (A4C), apical two chamber view (A2C), and apical three chamber view (A3C) videos, respectively. The accuracy of left ventricular endocardial segmentation based on videos in the three apical views was better than that of epicardial segmentation, and the segmentation result based on apical four chamber view videos was the best (89.16%). In both the internal validation dataset and external test dataset, the performance of the model based on videos in the three apical views was significantly better than that based on the single apical four chamber view videos (area under the curve: 0.942 vs 0.897; 0.937 vs 0.828).
The DL algorithm can not only automatically analyze echocardiographic videos, but also detect regional wall motion abnormalities. DL models can be applied in clinical practice to improve the diagnostic efficiency of echocardiography.
表1 训练集和测试集心肌梗死患者一般临床资料 |
临床资料 | 训练集(n=1137) | 测试集(n=105) |
---|---|---|
年龄(岁, ±s) | 71.0±13.0 | 62.3±12.7 |
男性[例(%)] | 812(71.4) | 77(73.3) |
并发症 | ||
高血压[例(%)] | 385(33.9) | 63(60.0) |
高血脂[例(%)] | 543(47.8) | 30(28.6) |
糖尿病[例(%)] | 448(39.4) | 20(19.0) |
心肌梗死部位 | ||
前壁[例(%)] | 489(43.0) | 31(29.5) |
下壁[例(%)] | 300(26.4) | 36(34.3) |
侧壁[例(%)] | 17(1.5) | 4(3.8) |
多个壁[例(%)] | 331(29.1) | 34(32.4) |
常规超声参数 | ||
LVEF(%, ±s) | 44.0±8.6 | 47.7±9.0 |
LVEDV(ml, ±s) | 111.0±39.3 | 116.7±36.9 |
LVESV(ml, ±s) | 60.0±29.4 | 62.7±29.9 |
LV(mm, ±s) | 48.0±6.5 | 49.5±5.6 |
RV(mm, ±s) | 31.0±3.5 | 31.6±3.1 |
LA(mm, ±s) | 40.0±4.5 | 38.6±3.6 |
RA(mm, ±s) | 32.0±4.1 | 33.0±4.0 |
注:LVEF为左心室射血分数;LVEDV为左心室舒张末容积;LVESV为左心室收缩末容积;LV为左心室侧径;RV为右心室侧径;LA为左心房侧径;RA为右心房侧径 |
表3 模型输入心尖三个切面与输入心尖四腔心单切面识别节段性室壁运动异常的效能比较(内部验证集) |
超声心动图切面 | AUC(95%CI) | 敏感度(95%CI) | 特异度(95%CI) | 准确性(95%CI) |
---|---|---|---|---|
心尖三个切面 | 0.942(0.930~0.957) | 0.857(0.839~0.881) | 0.916(0.890~0.939) | 0.875(0.855~0.898) |
心尖四腔心切面 | 0.897(0.855~0.927) | 0.772(0.744~0.809) | 0.880(0.858~0.920) | 0.805(0.781~0.828) |
注:AUC为ROC曲线下面积;95%CI为95%可信区间 |
表4 模型输入心尖三个切面与输入心尖四腔心单切面识别节段性室壁运动异常的效能比较(外部测试集) |
超声心动图切面 | AUC(95%CI) | 敏感度(95%CI) | 特异度(95%CI) | 准确性(95%CI) |
---|---|---|---|---|
心尖三个切面 | 0.937(0.913~0.958) | 0.867(0.820~0.931) | 0.868(0.807~0.914) | 0.938(0.924~0.950) |
心尖四腔心切面 | 0.828(0.781~0.872) | 0.724(0.663~0.822) | 0.806(0.740~0.859) | 0.847(0.828~0.866) |
注:AUC为ROC曲线下面积;95%CI为95%可信区间 |
1 |
|
2 |
|
3 |
|
4 |
|
5 |
|
6 |
|
7 |
|
8 |
|
9 |
|
10 |
|
/
〈 |
|
〉 |