开发用于微创手术中安全直视光学穿刺器置入的深度学习模型:一种预防穿刺器损伤的创新方法
本研究基于6层腹壁解剖结构标注3600张图像,训练YOLOv8深度学习模型并与报警系统结合,在腹腔镜视频中实时检测穿刺器置入时的腹壁层次。测试集平均精度均值95.8%、精确率89.8%、召回率91.7%,报警系统经本机构有经验外科医生验证并接受。
要点
- 基于6层腹壁解剖标注图像训练YOLOv8模型,用于穿刺器置入时实时识别腹壁层次。
- 测试集平均精度均值95.8%、精确率89.8%、召回率91.7%。
- 模型与报警系统结合,识别腹膜和腹腔层时触发报警,提醒术者调整镜头方向。
核心结论
- 深度学习有潜力在直视光学穿刺器置入过程中辅助外科医生,模型可实时检测精确的解剖标志参考。
- 模型与报警系统整合可及时提醒术者相应倾斜镜头,该框架的实施有潜力减少与直视光学穿刺器置入相关的并发症,从而提升手术安全性和结局。
研究设计与样本
- 基于6层腹壁解剖结构(皮下、腹直肌前鞘、腹直肌、腹直肌后鞘、腹膜和腹腔)标注边界框并分配类别,采用YOLOv8模型结合深度学习检测器训练数据集。
- 模型在静态图像上训练,并在腹腔镜视频上推理,以确保手术室实时检测。
- 共从89例腹腔镜视频病例中采集3600张图像,其中3000张用于训练、200张用于验证、400张用于测试。
主要发现
- 测试集结果为平均精度均值95.8%、精确率89.8%、召回率91.7%。
- 报警系统在本机构经有经验的外科医生验证并接受。
关键数值与推荐
- 报警系统在识别腹膜和腹腔层时激活。
- 测试集平均精度均值95.8%、精确率89.8%、召回率91.7%。
临床怎么用
- 原文未据此给出推荐。
局限性
- 原文未报告局限性。
以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。
原文
BACKGROUND: Direct optical trocar insertion is a common procedure in laparoscopic minimally invasive surgery. However, misinterpretations of the abdominal wall anatomy can lead to severe complications. Artificial intelligence has shown promise in surgical endoscopy, particularly in the employment of deep learning models for anatomical landmark identification. This study aimed to integrate a deep learning model with an alarm system algorithm for the precise detection of abdominal wall layers during trocar placement. METHOD: Annotated bounding boxes and assigned classes were based on the six layers of the abdominal wall: subcutaneous, anterior rectus sheath, rectus muscle, posterior rectus sheath, peritoneum, and abdominal cavity. The cutting-edge YOLOv8 model was combined with a deep learning detector to train the dataset. The model was trained on still images and inferenced on laparoscopic videos to ensure real-time detection in the operating room. The alarm system was activated upon recognizing the peritoneum and abdominal cavity layers. We assessed the model's performance using mean average precision (mAP), precision, and recall metrics. RESULTS: A total of 3600 images were captured from 89 laparoscopic video cases. The proposed model was trained on 3000 images, validated with a set of 200 images, and tested on a separate set of 400 images. The results from the test set were 95.8% mAP, 89.8% precision, and 91.7% recall. The alarm system was validated and accepted by experienced surgeons at our institute. CONCLUSION: We demonstrated that deep learning has the potential to assist surgeons during direct optical trocar insertion. During trocar insertion, the proposed model promptly detects precise landmark references in real-time. The integration of this model with the alarm system enables timely reminders for surgeons to tilt the scope accordingly. Consequently, the implementation of the framework provides the potential to mitigate complications associated with direct optical trocar placement, thereby enhancing surgical safety and outcomes.
引用
- Jearanai S, Wangkulangkul P, Sae-Lim W, et al. Development of a deep learning model for safe direct optical trocar insertion in minimally invasive surgery: an innovative method to prevent trocar injuries. Surg Endosc. 2023. PMID: 37558826. doi:10.1007/s00464-023-10309-1
供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。