摘要
近年來，臨床主治醫師的插管手術正逐漸從傳統的直接喉鏡 (Dedo Laryngoscope, DL) 轉變為影像喉鏡 (Video Laryngoscope, VLS)，而兩者所使用的氣管插管難度評估系統或其他相關系統是否能互相兼容仍有待驗證。因此本論文希望透過VLS錄製的影像進行模型訓練，並使模型能夠辨識出氣管插管過程中所出現的氣管插管的關鍵目標物件。本論文資料集搜集7位臨床主治醫師實際對插管模型安妮進行氣管插管的影像，並且共標記7種氣管插管的關鍵目標，後續採用YOLOv7進行模型訓練，實驗以留一交叉驗證 (Leave-One-Out Cross-Validation) 的方式進行模型評估，模型Precision、Recall及F1-score的平均與標準差已達92.19%±3.88%、92.16%±2.31%、92.17%±3.01%。
關鍵詞：氣管插管、關鍵目標、物體偵測、影像喉鏡
Abstract
In recent years, clinical tracheal intubation has increasingly shifted from the use of traditional direct laryngoscopes to video laryngoscopes. However, it remains unclear whether existing difficulty assessment systems and related evaluation frameworks are equally applicable to both approaches. To address this gap, this study develops an object detection model trained on images captured using a video laryngoscope system (VLS) to identify critical anatomical structures encountered during tracheal intubation. The dataset comprises images collected from seven clinical attending physicians performing intubation procedures on a training mannequin, with seven key anatomical targets manually annotated. The YOLOv7 model was employed for object detection training, and performance was evaluated using a leave‑one‑out cross‑validation strategy. The model achieved an average precision of 92.19% (±3.88%), recall of 92.16% (±2.31%), and F1‑score of 92.17% (±3.01%), demonstrating robust and consistent recognition performance across evaluators. These results suggest that object detection techniques combined with video laryngoscope imagery can effectively support the automated recognition of essential anatomical structures during tracheal intubation, with potential applications in training, assessment, and clinical support systems.
Keywords: Tracheal intubation, Object detection, Video laryngoscope, Anatomical structure recognition