文献速递胃肠外科围术期管理与 ERAS

基于血液像素分析的机器学习自动评估腹腔镜远端胃切除术手术技能

Machine learning-driven automated evaluation of surgical skills during laparoscopic distal gastrectomy based on blood pixel analysis.
PubMed 文献 · Surg Endosc · 2026-06-15

本研究开发基于逻辑回归的机器学习模型,通过RGB值将术中视频像素分类为血液或非血液,使用28例腹腔镜远端胃切除术视频提取的1008张图像(含234699个血液像素和376579个非血液像素)进行训练验证测试。模型总体准确率94.0%,灵敏度99.8%,特异度90.4%。分析2016-2021年内镜手术技能资格系统视频发现,高分组每帧血液像素数显著低于新手组,尤其在右侧大弯侧和胰腺上缘淋巴结清扫阶段。

要点

  • 机器学习模型可通过血液像素定量客观评估腹腔镜远端胃切除术技能
  • 高分组每帧血液像素数显著低于新手组,差异在淋巴结清扫阶段尤为明显
  • 该方法为手术表现评估提供了客观高效的基础

核心结论

  • 基于机器学习的血液像素定量方法可作为腹腔镜远端胃切除术手术技能的客观指标,具有可行性。

研究设计与样本

  • 开发基于逻辑回归的机器学习模型,依据RGB值分类血液与非血液像素;使用28例腹腔镜远端胃切除术视频提取的1008张图像(234699个血液像素和376579个非血液像素)进行训练、验证和测试;分析2016-2021年内镜手术技能资格系统提交的腹腔镜远端胃切除术视频,比较高分组(n=42)、低分组(n=24)和新手组(n=22)的每帧血液像素数。

主要发现

  • 模型总体准确率94.0%,灵敏度99.8%,特异度90.4%。
  • 高分组每帧血液像素数显著低于新手组(561.8 [279.8-937.5] vs. 1662.9 [660-3076],p=0.003)。
  • 右侧大弯侧淋巴结清扫阶段高分组显著低于新手组(761 [244.9-1058.3] vs. 2140.1 [650.4-2986.4],p=0.006)。
  • 胰腺上缘淋巴结清扫阶段高分组显著低于新手组(634.4 [205-948.2] vs. 1867.4 [1011.9-3925.3],p<0.001)。

临床怎么用

  • 可利用术中视频血液像素定量作为腹腔镜远端胃切除术技能评估的客观辅助工具,尤其在淋巴结清扫阶段。

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: In laparoscopic distal gastrectomy (LDG), intraoperative bleeding directly reflects surgical proficiency, including tissue handling and hemostasis. Conventional assessments of surgical skills, such as global operative assessment of laparoscopic skills, are subjective and time-consuming. Bleeding events can serve as an objective indicator for automated skill evaluation in this procedure. This study aimed to assess the feasibility of using an automated machine learning approach to objectively evaluate surgical skills during LDG by quantifying blood pixels in intraoperative videos. METHODS: A logistic regression-based machine learning model was developed to classify pixels as blood or non-blood based on RGB values. A total of 1008 images extracted from 28 LDG videos, comprising 234,699 blood pixels and 376,579 non-blood pixels, were used for training, validation, and testing. Additionally, LDG videos submitted for the endoscopic surgery skills qualification system between 2016 and 2021 were analyzed. Blood pixel counts per frame were calculated for both overall and specific lymphadenectomy phases and were then compared among the high-score (n = 42), low-score (n = 24), and novice groups (n = 22). RESULTS: The model achieved an overall accuracy of 94.0%, with a sensitivity of 99.8% and specificity of 90.4%. Analysis of the endoscopic surgery skills qualification system videos revealed significantly lower blood pixel counts per frame in the high-score group compared with the novice group (high-score group 561.8 [279.8-937.5] vs. novice group 1662.9 [660-3076], p = 0.003), particularly during lymphadenectomy on the right greater curvature (high-score group 761 [244.9-1058.3] vs. novice group 2140.1 [650.4-2986.4], p = 0.006) and suprapancreatic lymphadenectomy (high-score group 634.4 [205-948.2] vs. novice group 1867.4 [1011.9-3925.3], p < 0.001). CONCLUSIONS: This study demonstrates the feasibility of an automated machine learning-based approach to quantify blood pixels during LDG as an objective indicator of surgical skill. This method may provide a foundation for a more objective and efficient assessment of surgical performance.

来源:Surg Endosc,查看原文。版权归原作者所有。

引用

  1. Arakaki S, Takenaka S, Kitaguchi D, et al. Machine learning-driven automated evaluation of surgical skills during laparoscopic distal gastrectomy based on blood pixel analysis. Surg Endosc. 2026. PMID: 42298036. doi:10.1007/s00464-026-12927-x

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。