基于人工智能的内镜虚拟标尺测量食管静脉曲张直径的初步研究（含视频）

Artificial intelligence‑based endoscopic virtual ruler to measure the diameter of esophageal varices (with video)

摘要:

目的 初步探究基于人工智能的无创测量技术——内镜虚拟标尺（endoscopic virtual ruler，EVR）测量食管静脉曲张（esophageal varices，EV）直径的应用效果。方法 前瞻性纳入2022年10月至2023年5月在安徽医科大学第一附属医院住院的肝硬化伴EV的患者，3名操作者均分别采用视觉评估、食管静脉曲张测压仪（esophageal varix manometer， EVM）和EVR测量患者EV直径，并记录测量花费的时间。使用组内相关系数（intraclass correlation coefficient，ICC）分析3种方法测量EV直径的一致性，使用重复测量方差分析比较3种方法测量时间的差异。结果 研究共纳入41例肝硬化伴EV患者。3名医师视觉评估EV直径之间的ICC值为0.594，3名医师使用EVM测量EV直径之间的ICC值为0.840，3名医师使用EVR测量EV直径之间的ICC值为0.884。使用EVR和EVM测量EV直径的ICC值明显高于视觉评估。EVR和EVM测量EV直径之间的ICC值为0.991。视觉评估EV直径花费的时间为（18.6±2.2）s（14.7~23.3 s），使用EVR测量EV直径花费的时间为（41.5±4.1）s（31.7~50.3 s），使用EVM测量EV直径花费的时间为（170.8±26.4）s（129.3~229.3 s）。重复测量方差分析（经Greenhouse‑Geisser校正）显示，3种测量方法花费的时间差异有统计学意义［F（1.033，41.313）=1 233.800，P<0.001］。Bonferroni事后检验显示，3种方法花费时间两两比较差异均有统计学意义（P<0.001）。结论 EVR和EVM测量EV直径高度一致且EVR花费时间短，EVR可以无创、精确地测量EV直径，操作简单便捷。

Abstract:

Objective To evaluate the performance of an artificial intelligence-based endoscopic virtual ruler (EVR) for non-invasive measurement of esophageal varices (EV) diameter. Methods Patients with liver cirrhosis and EV hospitalized at the First Affiliated Hospital of Anhui Medical University between October 2022 and May 2023 were prospectively enrolled. EV diameter was measured using visual estimation, esophageal varix manometer (EVM), and EVR, with procedure times recorded. The intraclass correlation coefficient (ICC) was used to assess the consistency of EV diameter measurement obtained from the three methods, and repeated-measures ANOVA was used to compare differences in time measurements across three methods. Results The study included 41 patients with liver cirrhosis and EV. Inter-observer ICC for visual estimation was 0.594, versus 0.840 for EVM and 0.884 for EVR. The ICC value between the EV diameters measured by EVR and EVM was higher than that of the visual assessment. The ICC value between EV diameter measurement by EVM and EVR was 0.991. Measurement times differed significantly across methods (P<0.001): visual estimation 18.6±2.2 s (14.7-23.3 s), EVR 41.5±4.1 s (31.7-50.3 s), and EVM 170.8±26.4 s (129.3-229.3 s). Repeated measures analysis of variance (corrected using Greenhouse-Geisser) revealed significant differences in time across the three measurement methods [F(1.033, 41.313)=1 233.800, P<0.001]. Subsequent Bonferroni post-hoc tests revealed significant differences in time between all method pairs (P<0.001). Conclusion EVR provides rapid, non-invasive EV diameter measurements with excellent agreement to EVM assessment, offering an efficient alternative to conventional techniques.