基于时空注意力模型的腹腔镜手术缝合技能评估方法

Laparoscopic Suturing Skill Assessment Method Based on Spatiotemporal Attention

摘要:

对腹腔镜缝合技能的客观一致评估是外科培训的关键,但现有自动化评估方法在捕捉长时序依赖与操作细节方面存在不足,限制了评估的全面性与准确性。为解决该问题,提出了一种基于时空注意力机制的智能评估框架,采用 TimeSformer 作为骨干网络,提取更具判别性的时空特征以克服长时序动作依赖捕捉难点,并设计 AdaPool 模块,实现视频级特征的自适应汇聚,满足评估对关键操作片段的敏感度需求。基于 LapSuture 数据集的实验结果表明,该方法在技能分级任务中取得了83.0% 的准确率和 81.6% 的 F1 分数,优于传统方法。研究结果凸显了注意力驱动架构在实现稳健、可扩展且客观的手术技能评估方面的潜力,为提升外科教育与能力评价提供了新的思路。

Abstract:

Objective and consistent assessment of laparoscopic suturing skills is crucial for surgical training. However, existing automated assessment methods have limitations in capturing long-term temporal dependencies and operational details, which restricts the comprehensiveness and accuracy of assessment. To address this issue, an intelligent assessment framework based on spatiotemporal attention mechanism is proposed. It adopts TimeSformer as the backbone network to extract more discriminative spatiotemporal features, overcoming the difficulty in capturing long-term temporal action dependencies. Additionally, an AdaPool module is designed to realize adaptive aggregation of video-level features, satisfying the sensitivity requirement of assessment for key operational segments. Experimental results based on the LapSuture dataset demonstrate that this method achieves an accuracy of 83.0% and an F1-score of 81.6% in the skill grading task, outperforming traditional methods. The research results highlight the potential of attention-driven architectures in achieving robust, scalable, and objective surgical skill assessment, providing new insights for enhancing surgical education and competency evaluation.