深度学习技术识别喉返神经在经胸乳入路腔镜甲状腺手术中的探索  

Application of deep learning to identify recurrent laryngeal nerve in endoscopic thyroidectomy via breast approach

作　　者：花苏榕[1] 王智弘 李佳颐 高俊义[1] 王晶 贺光琳 帕拉沙特·叶尔肯别克 韩显林[1] 陈革[1] 廖泉[1] Hua Surong;Wang Zhihong;Li Jiayi;Gao Junyi;Wang Jing;He Guanglin;Yeerkenbieke Palashate;Han Xianlin;Chen Ge;Liao Quan(Department of General Surgery,Peking Union Medical College Hospital,Chinese Academy of Medical Sciences,Beijing 100730,China;Peking Union Medical College Hospital,Beijing 100730,China;Hangzhou Hikvision Digital Technology Co.,Ltd.,Hangzhou 310052,China;Hangzhou Hikimaging Technology Co.,Ltd.,Hangzhou 310052,China;Department of Second General Surgery,Ili Kazakh Autonomous Prefecture Friendship Hospital,Ili,835800,China)

机构地区：[1]中国医学科学院北京协和医院基本外科,北京100730 [2]中国医学科学院北京协和医学院,北京100730 [3]杭州海康威视数字技术股份有限公司,杭州310052 [4]杭州海康慧影科技有限公司,杭州310052 [5]新疆伊犁哈萨克自治州友谊医院普外二科,伊犁835800

出　　处：《中华内分泌外科杂志》2022年第3期287-292,共6页Chinese Journal of Endocrine Surgery

基　　金：中央高校基本科研业务费专项资金资助(3332019020);清华大学-北京协和医院自主科研联合资助项目(PTQH201911015)。

摘　　要：目的探索深度学习技术识别喉返神经(recurrent laryngeal nerve,RLN)在经胸乳入路腔镜甲状腺手术(endoscopic thyroidectomy,ETE)中的应用价值。方法收集2020年2月至2021年8月北京协和医院基本外科进行的经胸乳入路ETE视频。经2名甲状腺医师的筛选后,符合条件的手术视频根据一位高年资医师的意见分为低辨识度及高辨识度组,经过抽帧、标注、审核与校对后,按照随机数方法以5:1的比例分为训练集及测试集,统一输送至D-Linknet模型进行训练。根据交并比计算测试集中的灵敏度、精确率及平均Dice系数。结果46个视频共153520帧图片纳入了本研究。其中训练集共39个视频131039帧,测试集共计7个视频22481帧。交并比阈值为0.1及0.5时,高辨识度组中灵敏度及精确率分别为92.9%/72.8%及85.8%/67.2%,而在低辨识度组中则分别为47.6%/54.9%及37.6%/43.5%,平均Dice系数在两组中分别为0.781及0.663,证实了该模型对RLN具有较好的识别能力。结论基于深度学习的人工智能RLN识别在经胸乳入路ETE视频中可行,有可能帮助外科医生降低手术中误损伤风险,提高手术安全性。Objective To explore whether deep learning could apply to recognize the recurrent laryngeal nerve(RLN)in videos of endoscopic thyroidectomy(ETE)via breast approach.Methods Videos of ETE via breast approach in Peking Union Medical College Hospital from Feb.2020 to Aug.2021 were collected.Videos containing RLN were selected,and the outline of RLN was marked by two thyroid surgeons.Then data were divided into a training set and a test set in a ratio of 5:1 and classified into the high and low difficulty group according to a senior thyroid surgeon’s opinion.Those pictures were input to D-LinkNet model.Precision,sensitivity and mean dice index was calculated.Results A total of 46 videos including 153,520 frames of pictures were included in this study.131,039 frames of 39 videos were in the training set and 22,481 frames of 7 videos were in the test set.When the intersection over union threshold was 0.1,the sensitivity and precision was 92.9%/72.8%and 47.6%/54.9%in high and low recognition group,respectively.When the intersection over union threshold was 0.5,the sensitivity and precision turned to 85.8%/67.2%and 37.6%/43.5%in high and low difficulty group,respectively.Mean Dice index was 0.781 and 0.663 in high and low difficulty group,respectively.Conclusions RLN recognition based on deep learning is feasible and has potential application value in ETE,which may help surgeons reduce the risk of accidental injury of RLN and improve the safety of thyroidectomy.

关 键 词：喉返神经 腔镜甲状腺手术 人工智能 深度学习 

分 类 号：R653[医药卫生—外科学] R736.1[医药卫生—临床医学]