Sveric Krunoslav M, Botan Roxana, Winkler Anna, Dindane Zouhir, Alothman Ghatafan, Cansiz Baris, Fassl Jens, Kaliske Michael, Linke Axel
Department for Internal Medicine and Cardiology, Herzzentrum Dresden, Faculty of Medicine and University Hospital Carl Gustav Carus, TUD Dresden University of Technology, Fetscherstr. 76, 01307 Dresden, Germany.
Institute for Structural Analysis, TUD Dresden University of Technology, 01062 Dresden, Germany.
Eur Heart J Imaging Methods Pract. 2024 Dec 6;2(4):qyae130. doi: 10.1093/ehjimp/qyae130. eCollection 2024 Oct.
To evaluate the accuracy and feasibility of artificial intelligence (AI) in left ventricular global longitudinal strain (GLS) analysis as compared to conventional (Manual) and semi-automated (SemiAuto) method in echocardiography (Echo).
GLS validation was performed on 550 standard Echo exams by expert cardiologists. The performance of a beginner cardiologist without experience of GLS analysis was assessed on a subset of 90 exams. The AI employs fully automated view selection, classification, endocardial border tracing, and calculation of GLS from an entire Echo exam, while SemiAuto requires manual chamber view selection, and Manual involves full user input. Interobserver agreement was assessed using the intraclass correlation coefficient (ICC) for all three methods. Agreement of measures included Pearson's correlation (R) and Bland-Altman analysis [median bias; limits of agreement (LOA)]. With an 89% feasibility the AI showed good agreement with Manual (R = 0.92, bias = 0.7% and LOA: -3.5 to 4.8%) and with SemiAuto (r = 0.90, bias = 0.10% and LOA: -4.5 to 4%). ICCs for GLS were 1.0 for AI, 0.93 for SemiAuto, and 0.80 for Manual. After the 55th analysis, the beginner showed stable time performance with Manual (171 s), contrasting with the consistent performance of SemiAuto (85-69 s) from the beginning. The highest agreement between beginner and expert readers was achieved with AI (R = 1.00), followed by SemiAuto (R = 0.85) and Manual (R = 0.74).
Automated GLS analysis enhances efficiency and accuracy in cardiac diagnostics, particularly for novice users. Integration of automated solutions into routine clinical practice could yield more standardized results.
评估在超声心动图（Echo）中，与传统（手动）和半自动（SemiAuto）方法相比，人工智能（AI）在左心室整体纵向应变（GLS）分析中的准确性和可行性。
由心脏科专家对550例标准超声心动图检查进行GLS验证。在90例检查的子集中评估了一名没有GLS分析经验的初级心脏科医生的表现。AI采用全自动视图选择、分类、心内膜边界追踪，并从整个超声心动图检查中计算GLS，而SemiAuto需要手动选择腔室视图，手动方法则需要用户全程输入。使用组内相关系数（ICC）评估所有三种方法的观察者间一致性。测量一致性包括Pearson相关性（R）和Bland-Altman分析[中位数偏差；一致性界限（LOA）]。AI的可行性为89%，与手动方法显示出良好的一致性（R = 0.92，偏差 = 0.7%，LOA：-3.5至4.8%），与SemiAuto也显示出良好的一致性（r = 0.90，偏差 = 0.10%，LOA：-4.5至4%）。GLS的ICC，AI为1.0，SemiAuto为0.93，手动方法为0.80。在第55次分析后，初级医生使用手动方法时表现出稳定的时间（171秒），这与SemiAuto从一开始就保持一致的表现（85 - 69秒）形成对比。初级医生和专家读者之间的一致性最高的是AI（R = 1.00），其次是SemiAuto（R = 0.85）和手动方法（R = 0.74）。
自动GLS分析提高了心脏诊断的效率和准确性，特别是对于新手用户。将自动化解决方案整合到常规临床实践中可以产生更标准化的结果。
