Yen Hung-Hsuan, Hsiao Yi-Hsiang, Yang Meng-Han, Huang Jia-Yuan, Lin Hsu-Ting, Huang Chun-Chieh, Blue Jakey, Ho Ming-Chih
Department of Surgery, National Taiwan University Hospital Hsin-Chu Branch, No. 2, Sec. 1, Shengyi Rd., Zhubei City, Hsinchu County, 302058, Taiwan.
Department of Surgery, National Taiwan University Hospital, Taipei, Taiwan.
Surg Endosc. 2025 May;39(5):3006-3016. doi: 10.1007/s00464-025-11663-y. Epub 2025 Mar 21.
Laparoscopic cholecystectomy (LC) is a common procedure with standardized steps and validated assessment tools. However, the role of surgical actions in competency assessment remains underexplored, and automated models for surgical action recognition are lacking.
The Cholec80 dataset of 80 LC videos was analyzed for the Calot's Triangle Dissection (CTD) phase. Strasberg's critical view of safety (CVS) score and second-by-second annotations of surgical actions were evaluated. Videos were categorized into high_simple, low_simple, and high_complex groups based on competency levels and cholecystitis grade. The dataset was randomly divided into training (66 videos) and testing (14 videos) sets based on subgrouping. Surgical metrics were compared between subgroups, and a Random Forest model was constructed to predict competency levels using these metrics. In addition, a Video-Masked Autoencoders (VideoMAE) model was developed for surgical action recognition.
The high_simple group had significantly shorter CTD duration, fewer action transitions, and lower percentages of suctioning/irrigating, coagulating, and idle actions, but higher CVS scores and dissecting percentages. The Random Forest model achieved 93% accuracy (AUC: 0.96) in competency prediction, with CVS score, CTD duration, and percentages of dissecting, coagulating, and exposing as the top five important features. The VideoMAE model attained 89.11% overall accuracy in recognizing surgical actions, with the highest recall (0.97) for dissecting and the lowest (0.51) for suctioning/irrigating.
This study highlights the importance of surgical actions in competency assessment and presents automated models for evaluation and action recognition. These tools have potential to transform surgical education by providing objective and data-driven feedback for skill improvement.
腹腔镜胆囊切除术（LC）是一种具有标准化步骤和经过验证的评估工具的常见手术。然而，手术操作在能力评估中的作用仍未得到充分探索，且缺乏用于手术动作识别的自动化模型。
对80个LC视频的Cholec80数据集进行胆囊三角解剖（CTD）阶段分析。评估了斯特拉斯伯格安全关键视图（CVS）评分和手术操作的逐秒注释。根据能力水平和胆囊炎分级将视频分为高简单、低简单和高复杂组。基于子分组将数据集随机分为训练集（66个视频）和测试集（14个视频）。比较各亚组之间的手术指标，并构建随机森林模型以使用这些指标预测能力水平。此外，还开发了一种视频掩码自动编码器（VideoMAE）模型用于手术动作识别。
高简单组的CTD持续时间显著更短，动作转换更少，吸引/冲洗、凝固和空闲动作的百分比更低，但CVS评分和解剖百分比更高。随机森林模型在能力预测中达到了93%的准确率（AUC：0.96），其中CVS评分、CTD持续时间以及解剖、凝固和暴露的百分比是最重要的五个特征。VideoMAE模型在识别手术动作方面的总体准确率达到89.11% 在解剖方面召回率最高（0.97），在吸引/冲洗方面最低（0.51）。
本研究强调了手术操作在能力评估中的重要性，并提出了用于评估和动作识别的自动化模型。这些工具有可能通过提供客观的、数据驱动的反馈以促进技能提升，从而改变外科手术教育。
