基于深度学习的超声心动图自动识别瓣膜反流  

Automatic Detection of Valvular Regurgitation by Echocardiography Based on Deep Learning

作　　者：郭马特 宋艳杰 石婵 孙诗敏 马佳 刘博罕 王秋霜[4] 张丽伟[2] 杨菲菲[2,5] GUO Mate;SONG Yanjie;SHI Chan;SUN Shimin;MA Jia;LIU Bohan;WANG Qiushuang;ZHANG Liwei;YANG Feifei(不详;Department of cardiology,Sixth Medical Center of PLA General Hospital,Beijing 100000,China;Big Data Center of PLA General Hospital,Beijing 100039,China)

机构地区：[1]解放军总医院第四医学中心心内科,北京100037 [2]解放军总医院第六医学中心心内科,北京100000 [3]解放军总医院第一医学中心心外科,北京100039 [4]解放军总医院第四医学中心健康医学科,北京100037 [5]解放军总医院大数据中心,北京100039

出　　处：《中国医学影像学杂志》2025年第2期147-151,共5页Chinese Journal of Medical Imaging

基　　金：国家自然科学基金(82202265)。

摘　　要：目的 探讨深度学习算法自动识别基于彩色多普勒动态视频瓣膜反流的可行性。资料与方法 回顾性收集2015年6月—2019年9月解放军总医院第四医学中心1 109例瓣膜反流的超声心动图影像作为训练集和验证集,前瞻性连续性收集2023年5月13日—6月13日解放军总医院第四医学中心1 562例超声心动图影像作为测试集,其中二尖瓣反流378例,主动脉瓣反流223例。利用深度学习算法建立切面分类模型和瓣膜反流识别模型,分析深度学习模型识别切面分类的效能。结果 本研究建立的深度学习切面分类模型可自动识别诊断二尖瓣反流和主动脉瓣反流的2个切面,对胸骨旁长轴彩色多普勒切面和心尖四腔二尖瓣彩色多普勒切面的识别准确度分别为1.00和0.93;深度学习模型诊断二尖瓣反流的敏感度为0.847,特异度为0.852,准确度为0.849,曲线下面积为0.930。深度学习模型诊断主动脉瓣反流的敏感度为0.857,特异度为0.861,准确度为0.859,曲线下面积为0.940。结论 深度学习算法可自动识别瓣膜反流,具有成为心脏瓣膜病筛查工具的潜力。Purpose To investigate the feasibility of a deep learning framework to automatically analyze echocardiographic color Doppler videos in detecting valvular regurgitation.Materials and Methods This study retrospectively collected echocardiographic images of 1109 patients with valvular regurgitation in the Fourth Medical Center of PLA General Hospital,from June 2015 to September 2019 as the training and validation sets.A prospective continuous collection of 1562 echocardiography images was used as the test set in the Fourth Medical Center of PLA General Hospital from May 13 to June 13,2023,including 378 cases of mitral regurgitation and 223 cases of aortic regurgitation.This study developed deep learning networks to establish view classification model and valvular regurgitation recognition model,including the efficiency of section classification of deep learning models.Results The deep learning view classification model in this study could automatically identify two views for diagnosing mitral regurgitation and aortic regurgitation.The recognition accuracy for the parasternal long axis color Doppler view and the apical four chamber mitral color Doppler view was 1.00 and 0.93,respectively.The sensitivity,specificity,accuracy and area under the curve of the deep learning model for diagnosing mitral regurgitation were 0.847,0.852,0.849 and 0.930,respectively.The sensitivity,specificity,accuracy and area under the curve of the deep learning model in diagnosing aortic regurgitation were 0.857,0.861,0.859 and 0.940,respectively.Conclusion Deep learning algorithms can automatically identify valvular regurgitation and have the potential to become a screening tool for valvular heart disease.

关 键 词：深度学习 超声心动描记术 心脏瓣膜疾病 二尖瓣闭锁不全 主动脉瓣关闭不全 诊断 

分 类 号：R445.1[医药卫生—影像医学与核医学] TP18[医药卫生—诊断学] R542.5[医药卫生—临床医学]