Scholarlink

Purpose <p>Tracheal intubation skills are essential for medical professionals. In case of TI difficulties, video laryngoscopes are recommended. Clinical practice is shifting from traditional direct laryngoscopes to video laryngoscopes. However, compatibility of difficulty assessment systems between these methods remains unverified. This paper trains a model using VLS images to identify crucial tracheal intubation targets.</p> Methods <p>The dataset includes images from 7 clinical attending physicians performing intubations on a mannequin. Seven key intubation targets were labeled. YOLOv7 was used for model training with Leave-One-Out Cross-Validation for evaluation.</p> Results <p>The model achieved Precision, Recall, and F1-score of 92.19%±3.88%, 92.16%±2.31%, and 92.17%±3.01%, respectively.</p> Conclusion <p>Three model training approaches were tested. Removing underexposed images improved performance. Marking unlabeled regions as "nonTarget" enhanced crucial target identification but caused confusion between intubation and exit tube appearances, decreasing recall. Future research using continuous image sequences could resolve this through temporal ordering.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Application of YOLO-Based for Precise Identification of Critical Anatomical Targets in Tracheal Intubation

  • 期刊论文

  • 发表日期：

    2025年11月18日

查看全文
  • Kun-Yi Huang,
  • Yu-Hwa Wu,
  • Chia-Chih Tseng,
  • Qian-Bei Hong

摘要

Purpose

Tracheal intubation skills are essential for medical professionals. In case of TI difficulties, video laryngoscopes are recommended. Clinical practice is shifting from traditional direct laryngoscopes to video laryngoscopes. However, compatibility of difficulty assessment systems between these methods remains unverified. This paper trains a model using VLS images to identify crucial tracheal intubation targets.

Methods

The dataset includes images from 7 clinical attending physicians performing intubations on a mannequin. Seven key intubation targets were labeled. YOLOv7 was used for model training with Leave-One-Out Cross-Validation for evaluation.

Results

The model achieved Precision, Recall, and F1-score of 92.19%±3.88%, 92.16%±2.31%, and 92.17%±3.01%, respectively.

Conclusion

Three model training approaches were tested. Removing underexposed images improved performance. Marking unlabeled regions as "nonTarget" enhanced crucial target identification but caused confusion between intubation and exit tube appearances, decreasing recall. Future research using continuous image sequences could resolve this through temporal ordering.