基于yolox框架整合左心室分割和关键点检测的人工智能模型评估慢性肾衰竭患者左心室收缩功能  

Diagnostic efficacy of artificial intelligence model based on yolox framework integrating left ventricular segmentation and key point detection to automatically measure left ventricular ejection function in patients with chronic renal failure

作　　者：李寒笑 姬强 赵阳 贾创 纪淑姣[4] 袁建军[2] 邢雨 曾甜 朱好辉[2] Li Hanxiao;Ji Qiang;Zhao Yang;Jia Chuang;Ji Shujiao;Yuan Jianjun;Xing Yu;Zeng Tian;Zhu Haohui(Xinxiang Medical University,Xinxiang 453000,China;Department of Ultrasound,Henan Provincial People′s Hospital,Zhengzhou 450003,China;University of Chinese Academy of Sciences,Beijing 100049,China;Department of Ultrasound,Henan Provincial Chest Hospital,Zhengzhou 450003,China)

机构地区：[1]新乡医学院,新乡453000 [2]河南省人民医院超声科,郑州450003 [3]中国科学院大学,北京100049 [4]河南省胸科医院超声科,郑州450003

出　　处：《中华超声影像学杂志》2024年第5期407-414,共8页Chinese Journal of Ultrasonography

基　　金：国家自然科学基金面上项目(82371980);2021年度河南省卫生健康中青年学科带头人培育项目;2021年河南省医学科技攻关计划项目(SBGJ202102013)。

摘　　要：目的评价基于yolox框架整合左心室分割和关键点检测的人工智能(AI)模型对慢性肾衰竭(CRF)患者左心室射血分数(LVEF)的检测性能。方法回顾性收集2019年1月至2023年6月于河南省人民医院就诊的2000例18~80岁无节段性室壁运动异常、结构性心脏病、心脏外科手术后及心肌病的成人4284幅超声心动图图像勾画心内膜,建立基于yolox框架整合左心室分割和关键点检测的AI模型图像,并以5∶1的比例分为训练集(1675例)及测试集(325例);收集2020年5月至2021年5月于河南省胸科医院就诊的正常成人志愿者100例228幅超声心动图图像作为外部测试集验证;连续入组2019年4月至2023年6月于河南省人民医院就诊的患者204例792幅超声心动图图像对AI模型测量效能进行评价。运用Spearman相关性统计方法分析AI模型测量与3名高年资超声心动图专业医师手动测量及TomTec软件测量方法一致性。将受检者分别分为图像清晰、图像不清晰组以及LVEF正常、LVEF减低组,比较亚组间一般资料的差异。计算组内相关系数(ICC)分析一致性,从而评估模型性能。结果AI模型测量的LVEF与高年资医师手动测量及TomTec模型测量均呈显著强相关(rs=0.834、0.826,均P<0.01)。图像清晰组、图像不清晰组3种测量方法ICC值分别为0.96、0.97;所有受检者、LVEF正常组和LVEF减低组3种测量方法ICC值分别为0.96、0.90、0.96。结论基于yolox框架整合左心室分割和关键点检测的AI模型自动测量CRF患者LVEF具有良好评估效能。Objective To evaluate the detection performance of left ventricular ejection fraction(LVEF)in patients with chronic renal failure(CRF)by an artificial intelligence(AI)model based on yolox framework integrating left ventricular segmentation and critical point detection.Methods From January 2019 to June 2023,a total of 4284 echocardiographic images of 2000 adults aged 18-80 years without segmental wall motion abnormalities,structural heart disease,cardiac surgery or cardiomyopathy were collected in Henan Provincial People′s Hospital to delineate the endocardial membrane,as a training set,an AI model based on yolox framework integrating left ventricular segmentation and critical point detection was established.The images were divided into the training set(n=1675)and the test set(n=325)in a ratio of about 5∶1.All 228 echocardiographic images of 100 normal adult volunteers who were treated in Henan Provincial Chest Hospital from May 2020 to May 2021 were collected as external test set validation.All 792 echocardiographic images of 204 patients treated in Henan Provincial People′s Hospital from April 2019 to June 2023 were continuously enrolled to evaluate the measurement efficiency of AI model.Spearman correlation statistical method was used to analyze the consistency of AI model measurement with manual measurement and TomTec software measurement methods of 3 senior echocardiographic professionals.Subjects were divided into clear image group,unclear image group,normal LVEF group and reduced LVEF group,the differences of general data between the two groups were compared.The correlation coefficient(ICC)within the group was calculated to analyze the consistency,so as to evaluate the model performance.Results LVEF measured by AI model was significantly correlated with both manual measurement and TomTec model measurement(rs=0.834,0.826;all P<0.01).ICC values of the clear image group and the unclear image group were 0.96 and 0.97,respectively.ICC values for all subjects,normal LVEF group and reduced LVEF group were 0.

关 键 词：超声心动描记术 人工智能 左心室收缩功能 TomTec模型 慢性肾衰竭 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程] R692.5[自动化与计算机技术—控制科学与工程] R540.45[医药卫生—泌尿科学]