應用於手術影片評估之動作片段自動擷取與分析框架

Abstract

隨著腹腔鏡微創手術技術的廣泛應用，膽囊摘除手術（Laparoscopic Cholecystectomy, LC）已成為臨床外科醫師必須熟練的基本手術之一。然而，現行手術品質評估方式多仰賴資深醫師的主觀回顧與判斷，不僅耗時費力，更易受評估者經驗、觀察角度與主觀偏見所影響，限制了其於大規模臨床訓練與術後監控中的可行性與一致性。因此，如何建立一套具臨床意義、標準化且可自動化執行的品質評估機制，成為醫學影像分析與智慧手術發展中亟待解決的挑戰。 本研究針對上述問題，提出一套應用於腹腔鏡手術影片之動作片段自動擷取與分析框架，整合動作序列分析、平移相似度衡量、非監督式分群與機器學習等技術，以期發展具結構化與量化能力之手術品質評估系統。研究以公開資料集 Cholec80 為實證基礎，設計三大核心模組：首先，建立動作片段抽樣與階段化切割機制，將完整手術影片轉換為具有臨床語意的多段動作序列；其次，設計可容忍時間平移的相似度計算方法，搭配階層式分群法，自高品質手術中萃取具代表性的標竿動作片段；最後，透過與代表片段之距離計算，建立品質指標並導入迴歸模型（如SVR與XGBoost），進行測試影片之品質分數預測。 實驗結果顯示，所提方法在多種距離指標與品質評分間均具顯著相關性，其中加權Lag Distance對LCATCalot Sum評分具備穩定預測能力，模型可有效分辨不同品質層次之手術影片，展現出良好的實務應用潛力。本研究之貢獻不僅在於提出一套具延展性與通用性的品質評估框架，更奠定自動化手術訓練評量與智慧手術品質控制之重要基礎。未來可進一步擴充至其他手術類型或結合即時預測技術，推進手術AI評估系統的臨床落地。

With the widespread adoption of minimally invasive laparoscopic techniques, laparoscopic cholecystectomy (LC) has become a fundamental surgical procedure that all general surgeons must master. However, current approaches to surgical quality assessment largely rely on retrospective evaluations by senior surgeons, which are time-consuming, labor-intensive, and susceptible to observer experience, viewing perspective, and subjective bias. These limitations hinder the scalability and consistency of quality monitoring and surgical training in large-scale clinical settings. Addressing this critical challenge, there is an urgent need to establish a clinically meaningful, standardized, and automated assessment framework for surgical performance evaluation. This study proposes a novel framework for automated extraction and analysis of action segments in laparoscopic surgical videos. The framework integrates action sequence analysis, lag-tolerant similarity metrics, unsupervised clustering, and machine learning techniques to develop a structured and quantifiable approach to surgical quality assessment. Using the publicly available Cholec80 dataset, the framework is structured into three main modules: (1) a segment sampling and temporal slicing mechanism to convert entire surgical videos into clinically meaningful action sequences; (2) a lag-distance-based similarity metric to tolerate temporal misalignment, combined with hierarchical clustering to extract representative benchmark segments from high-quality surgeries; and (3) a quality prediction model that calculates distances between test video segments and benchmark segments, and subsequently predicts surgical quality scores via regression models such as SVR and XGBoost. Experimental results demonstrate significant correlations between the proposed distance metrics and clinically annotated quality scores. In particular, the weighted lag distance shows robust predictive performance for LCATCalot Sum scores, enabling effective differentiation of surgical quality levels. The proposed framework not only contributes to a scalable and generalizable system for automated quality assessment but also establishes a foundational methodology for intelligent surgical training evaluation and quality control. Future extensions may include applications to other surgical procedures or real-time inference, further advancing the clinical deployment of AI-based surgical assessment systems.