基于深度学习YOLOv5网络的机器人辅助单孔腹腔镜子宫切除术实时解剖标志指示系统

Real⁃time anatomical landmark indication system for robot⁃assisted single⁃port laparoscopic hysterectomy based on deep learning YOLOv5 network

目的:开发和评估基于深度学习的人工智能模型,将其应用于临床教学,帮助临床医生学习识别解剖标志,提升学习兴趣,增强手术能力。方法:开发和训练深度学习模型以识别以下解剖标志,包括子宫、输卵管、圆形韧带、卵巢、子宫-卵巢韧带、膀胱、子宫-膀胱空间(膀胱分离安全区)、子宫骶韧带和子宫切除手术期间KOH阴道切除器的边缘。采用P(精度)、mAP(平均精度)、F1-置信曲线、R(召回率)来评估模型的准确性,采用临床分组教学应用的方式来验证该模型是否能对临床教学提供帮助。结果:在达芬奇机器人系统构建的手术视频库上训练人工智能模型。该模型P值为94.80%,mAP值为99.10%,F1得分为96.00%,R为99.00%。将其应用于临床教学后,试验组学员对机器人腹腔镜下妇科解剖结构辨认熟练程度,在8个方面均有显著提高(P<0.05),尤其是子宫-卵巢韧带、膀胱、子宫-膀胱空间(膀胱分离安全区),有了明显提升,培训后两组医师妇科专业理论成绩的提升有统计学意义(均P<0.05),试验组成绩优于对照组。结论:人工智能可用于识别手术领域的解剖结构,帮助临床医生更快更好地学习解剖标志识别,加强临床医生的手术能力,为降低术中不良事件的风险提供帮助。

Objectives:To develop and evaluate artificial intelligence models based on deep learning,and apply them to clinical teaching to help clinicians learn to recognize anatomical landmarks,improve learning interest,and enhance their surgical capabilities.Methods:A deep learning model was developed and trained to recognize the following anatomical landmarks,including the uterus,fallopian tubes,round ligament,ovary,utero‐ovarian ligament,bladder,utero‐bladder space(bladder separation safe zone),uterosacral ligament,and the edge of the KOH colpotomizer system during the hysterectomy.P(precision),mAP(average precision),F1‐confidence curve,and R(recall rate)were used to evaluate the accuracy of the model,and the application of clinical group teaching was used to verify whether the model can help clinical teaching.Results:We trained the artificial intelligence model on the library of surgical videos built by Da Vinci robotic system in our hospital.The P value of this model is 94.80%,the mAP value is 99.10%,the F1 score is 96.00%,and the R is 99.00%.After it was applied to clinical teaching,the students in the experimental group had significantly improved their proficiency in identifying gynecological anatomical structures under robotic laparoscopy in 8 aspects(P<0.05),especially the uterine‐ovarian ligament,bladder,and uterine‐bladder space(safety zone for bladder separation),there has been a significant improvement.After training,the theoretical scores of the gynecology specialty of the two groups of clinicians were significantly increased(both P<0.05),and the scores of the experimental group were better than those of the control group.Con⁃clusion:Artificial intelligence can be used to identify anatomical structures in the surgical field,help clinicians learn anatomical landmark recognition faster and better,strengthen the surgical ability of clinicians,and help reduce the risk of intraoperative adverse events.