人工智能自动评估腹腔镜肝切除术手术技能的概念验证研究
本研究回顾性分析146例日本内镜外科学会技能认证系统提交的腹腔镜肝切除术视频,使用DeepLabv3+分割出血区域并计算血像素计数(BPC),评估其与ESSQS手术熟练度的关联及对低分和未通过视频的区分能力。结果显示BPC与ESSQS评分相关,对低分组和未通过组有一定区分度,高特异性阈值可用于筛选需专家复核的手术。
要点
- BPC与ESSQS手术熟练度相关,从高分组到低分组单调增加(P<0.001)。
- BPC识别低分组的AUC为0.708,识别未通过组的AUC为0.648。
- 以BPC阈值4,794,565像素识别低分组特异性90.2%、敏感性29.2%。
核心结论
- 基于AI的BPC与腹腔镜肝切除术的ESSQS手术熟练度相关,可作为高特异性指标标记需专家复核的手术。
研究设计与样本
- 回顾性分析2020-2021年提交至日本内镜外科学会ESSQS的146例腹腔镜肝切除术视频,使用DeepLabv3+分割出血区域并计算BPC。
主要发现
- BPC中位数为1,741,749像素(范围7,944-10,507,218)。
- BPC从高分组到低分组单调增加(P<0.001),未通过组BPC高于通过组(P=0.001)。
- BPC识别低分组的AUC为0.708(95%CI 0.599-0.817),识别未通过组的AUC为0.648(95%CI 0.546-0.751)。
- 在低难度病例中,BPC识别低分组的AUC为0.784。
- 以BPC阈值4,794,565像素识别低分组,特异性90.2%、敏感性29.2%;应用于通过/未通过分类时,无通过视频BPC达到或超过该阈值。
关键数值与推荐
- ESSQS评分70分为通过/未通过 cutoff;低分组定义为ESSQS评分≤均值-1个标准差。
- BPC阈值选择标准为特异性≥90%,最终阈值为4,794,565像素。
临床怎么用
- BPC可作为客观、可扩展的高特异性指标,帮助标记需专家复核的手术视频,尤其在低难度病例中。
局限性
- 需外部验证和前瞻性实施研究。
以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。
原文
BACKGROUND: Maintaining a dry surgical field is fundamental to safe laparoscopic hepatectomy (LH); however, formal certification systems rely on labor-intensive expert review of unedited surgical videos. This study evaluated the association between AI-derived blood pixel count (BPC), which quantifies redness of the hepatic transection plane, and Endoscopic Surgical Skill Qualification System (ESSQS)-based surgical proficiency, and assessed its discriminatory performance for Low-score and Failed videos. METHODS: We retrospectively analyzed 146 LH videos submitted to the ESSQS of the Japan Society for Endoscopic Surgery in 2020-2021. Bleeding regions were segmented using DeepLabv3+. BPC was calculated as the total number of pixels classified as bleeding across all sampled frames. Outcomes were Pass/Fail classification using the 70-point cutoff and the Low-score group, defined as ESSQS scores ≤ mean - 1 standard deviation. ROC analysis assessed discrimination. A BPC threshold with specificity ≥90% was selected. RESULTS: The median BPC was 1,741,749 pixels (range, 7,944-10,507,218). BPC increased monotonically from the High-score group to the Low-score group (P < 0.001), and the Failed group had higher BPC than the Passed group (P = 0.001). The AUC of BPC was 0.708 (95% CI, 0.599-0.817) for identifying the Low-score group and 0.648 (95% CI, 0.546-0.751) for identifying the Failed group. Using a BPC threshold of 4,794,565 pixels, specificity and sensitivity for Low-score identification were 90.2 and 29.2%, respectively. In low-difficulty cases, the AUC for identifying the Low-score group was 0.784. When the same threshold was applied to Pass/Fail classification, no Passed video had a BPC at or above the threshold. CONCLUSIONS: AI-based BPC was associated with ESSQS-based surgical proficiency in LH. This objective, scalable, high-specificity metric may help flag procedures warranting expert review and timely feedback, particularly in low-difficulty cases. External validation and prospective implementation studies are warranted.
引用
- Aoki K, Kobayashi S, Sasaki K, et al. Proof-of-concept study toward the automated evaluation of surgical skills in laparoscopic liver resection using artificial intelligence. Surg Endosc. 2026. PMID: 42934424. doi:10.1007/s00464-026-13388-y
供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。