人工智能技术在心脏超声常规参数测量及左室舒张性慢性心力衰竭诊断中的应用  

Application of artificial intelligence technology in measurement of conventional echocardiographic parameters and diagnosis of left ventricular diastolic chronic heart failure

作　　者：于立恒 林锡祥 陈煦 何昆仑[1] YU Liheng;LIN Xixiang;CHEN Xu;HE Kunlun(Chinese PLA General Hospital,Beijing 100853,China)

机构地区：[1]解放军总医院医学大数据研究中心,北京100853 [2]解放军总医院研究生院,北京100853

出　　处：《陕西医学杂志》2023年第7期826-830,共5页Shaanxi Medical Journal

基　　金：中国2020医疗健康智能筛查诊断项目(2020-0103-3-1)。

摘　　要：目的:探究人工智能算法模型在心脏超声常规参数测量及左室舒张性慢性心力衰竭(CHF)诊断中的应用效果。方法:收集410例左室舒张性CHF疑似病例者心脏超声图像为研究对象。由一组高年资超声医师完成纳入病例者的心脏超声参数测量;同时,采用人工智能深度学习模型,即图像分割卷积神经网络,对纳入者心脏超声图像进行自动智能分割,并设计专门算法计算基于心脏超声影像的心脏结构和功能参数。通过比较两种方法获得心脏超声参数的一致性、偏差情况及受试者工作特征曲线(ROC)来探究人工智能深度学习技术在左室舒张性CHF诊断中的价值。结果:人工智能与医师所测心脏超声参数左室舒张末期容积(LVEDV)、左室收缩末期容积(LVESV)、左室射血分数(LVEF)间一致性一般(0.60≤ICC<0.80),E/A、E/e’、主动脉瓣峰值流速(AVPW)一致性极好(ICC≥0.80)。Bland-Altman分析表明,人工智能与医师所测心脏超声参数LVEDV、LVESV、LVEF间偏差较大,平均偏差分别为-16.9%、-7.0%和1.0%,E/A、E/e’、AVPW间偏差较低,平均偏差分别为0.0%、-0.4%和-0.0%。ROC曲线显示,人工智能所测E/A、E/e’、AVPW均对左心室舒张性CHF具有一定诊断价值(AUC=0.853、0.777、0.770,P<0.05)。结论:人工智能可快速识别并分割处理心脏超声图像,自动计算常规心脏参数,且关键参数与高年资超声医师结果一致性较好,并能用于左室舒张性CHF的临床辅助诊断。Objective:To explore the application effect of artificial intelligence algorithm model in the measurement of conventional echocardiographic parameters and the diagnosis of left ventricular diastolic chronic heart failure(CHF).Methods:The ultrasound images of 410 suspected cases of left ventricular diastolic CHF were collected.Cardiac ultrasound parameters were measured by the same group of senior sonographers.At the same time,artificial intelligence model(image segmentation convolutional neural network)was used to segment and process the input cardiac ultrasound images,and the parameters of cardiac ultrasound were calculated and output.The consistency and deviation of cardiac ultrasound parameters obtained by two methods were compared.The value of artificial intelligence parameters in the diagnosis of left ventricular diastolic CHF was explored by receiver operating characteristic curve(ROC).Results:The artificial intelligence was consistent with the measured cardiac ultrasound parameters left ventricular end-diastolic volume(LVEDV),left ventricular end-systolic volume(LVESV)and left ventricular ejection fraction(LVEF)(0.60≤ICC<0.80),and the consistency of E/A,E/e’and AVPW was very good(ICC≥0.80).Blan-altman analysis showed that the deviation between artificial intelligence and the measured cardiac ultrasound parameters LVEDV,LVESV and LVEF was large,the average deviation was-16.9%,-7.0%and 1.0%respectively,and the deviation between E/A,E/e’and AVPW was low,the mean deviations were 0.0%,-0.4%and-0.0%respectively.ROC curve showed that E/A,E/e’and AVPW measured by artificial intelligence had certain diagnostic value for left ventricular diastolic CHF(AUC=0.853,0.777,0.770,P<0.05).Conclusion:Artificial intelligence can quickly identify and segment cardiac ultrasound images,automatically calculate conventional ultrasound parameters,and some conventional parameters can be highly consistent with the results of senior sonographers,and can be used for the clinical diagnosis of left ventricular diastolic CHF.

关 键 词：心脏超声 人工智能 深度学习 图像分割卷积神经网络 慢性心力衰竭 左室舒张功能 

分 类 号：R540.4[医药卫生—心血管疾病]