摘要
目的探讨基于计算机深度学习经皮椎板间脊柱内镜下手术视野的多元素识别网络模型的研究及应用价值。方法回顾性队列研究。纳入2021年9月—2022年3月徐州中心医院脊柱外科行经皮椎板间脊柱内镜下腰椎间盘切除术的腰椎间盘突出患者62例,其中男34例、女28例,年龄27~77(50.0±14.7)岁。收集患者内镜手术视频,选取4840张经皮脊柱内镜手术视野图片(包含各种组织结构及手术器械)建立图片数据集,按照2∶1∶2分为训练集、验证集和测试集,开发8种基于实例分割的卷积神经网络模型(模型的分割头部分别为Solov2、CondInst、Mask R-CNN及Yolact,主干网络分别设置为ResNet101、ResNet50)。采用边框检测、轮廓分割的均值平均精度(mAP)及图像实时识别的每秒帧数(FPS)来衡量各模型对(神经、黄韧带、髓核等)解剖结构,以及(内镜钳、高速金刚石磨钻等)手术器械的分类、定位及图像实时识别的性能。结果(1)8种卷积神经网络模型在图像边界框检测的精度方面由高到低依次为Mask R-CNN(ResNet101)、CondInst(ResNet101)、CondInst(ResNet50)、Mask R-CNN(ResNet50)、Yolact(ResNet101)、Yolact(ResNet50),其中,Mask R-CNN(ResNet101)模型精度最高(mAP=68.7%),Yolact(ResNet50)精度最低(mAP=49.7%)。(2)8种卷积神经网络模型在图像轮廓分割的精度方面由高到低依次Solov2(ResNet101)、Solov2(ResNet50)、Mask R-CNN(ResNet101)、CondInst(ResNet101)、Mask R-CNN(ResNet50)、CondInst(ResNet50)、Yolact(ResNet101)、Yolact(ResNet50)。其中,Solov2(ResNet101)精度最高(mAP=70.1%),Yolact(ResNet50)精度最低(mAP=55.2%)。(3)在图像实时识别方面,Yolact模型速度最快,其次为Solov2模型、Mask R-CNN模型,CondInst(ResNet101)速度最慢。结论基于计算机深度学习的经皮椎板间脊柱内镜手术视野多元素识别模型可以实时识别和跟踪解剖组织及手术器械。其中,Mask R-CNN(ResNet101)模型可用作脊柱内镜操作虚拟教育工�
Objective This study aimed to investigate the research and application value of multi-element identification network models for the visual field of endoscopic interlaminar spine surgery through computer deep learning.Methods We conducted a retrospective study on 62 patients diagnosed with lumbar disc herniation who underwent percutaneous interlaminar spinal endoscopic lumbar discectomy at the Xuzhou Central Hospital from September 2021 to March 2022.Among them,34 were males and 28 were females.Their ages ranged from 27 to 77(50.0±14.7)years.We established an image database by collecting endoscopic surgical videos of the patients,which was labeled by two spinal surgeons.We selected 4,840 images of the visual field of percutaneous endoscopic spine surgery(including various tissue structures and surgical instruments),divided into training data,validation data,and test data according to 2∶1∶2.We developed eight convolutional neural network models based on instance segmentation(the segmentation heads of the models were Solov 2,CondInst,Mask R-CNN,and Yolact,whereas the backbone of the networks were set to ResNet 101 and ResNet 50,respectively).Mean average precision(mAP)of image bourdary box detection and contour segmentation and frames per second(FPS)were used to measure the performance of each model for classification,localization,and recognition in real time.Results(1)The accuracy of 8 kinds of convolutional neural network models ranged from high to low in image boundary box detection was Mask R-CNN(ResNet 101),CondInst(ResNet 101),CondInst(ResNet 50),Mask R-CNN(ResNet 50),Yolact(ResNet 101),Yolact(ResNet 50).The accuracy of Mask R-CNN(ResNet 101)model was the highest(mAP=68.7%),and the accuracy of Yolact(ResNet 50)model was the lowest(mAP=49.7%).(2)The accuracy of 8 convolutional neural network models ranged from high to low in image contour segmentation was Solov2(ResNet 101),Solov2(ResNet 50),Mask R-CNN(ResNet 101),CondInst(ResNet 101),and Mask R-CNN(ResNet 50),CondInst(ResNet 50),Yolact(ResNet 101),Yolact(
作者
卜晋辉
王亚日
何博
付傲
赵佳琦
黄森
梁军
王振飞
许龙
雷雁
董明会
刘光普
牛茹
马超
刘光旺
Bu Jinhui;Wang Yari;He Bo;Fu Ao;Zhao Jiaqi;Huang Sen;Liang Jun;Wang Zhenfei;Xu Long;Lei Yan;Dong Minghui;Liu Guangpu;Niu Ru;Ma Chao;Liu Guangwang(Affiliated Xuzhou Clinical College of Xuzhou Medical University,Xuzhou 221009,China;School of Computer Science,China University of Mining and Technology,Xuzhou 221116,China;Department of Orthopedic Surgery,Xuzhou Central Hospital,Xuzhou Central Hospital Affiliated to Nanjing University of Chinese Medicine,Xuzhou School of Clinical Medicine of Nanjing Medical University,Xuzhou Central Hospital Affiliated to Medical School of Southeast University,Xuzhou 221009,China)
出处
《中华解剖与临床杂志》
2024年第5期289-295,共7页
Chinese Journal of Anatomy and Clinics
基金
江苏省卫生健康科研项目(M2022048)
江苏省青年医学重点人才培养计划(青苗工程)(QNRC2016392)
徐州市医学领军人才培养项目(XWRCHT20210035)
徐州市引进临床医学专家团队项目(2019TD002)
徐州市科技项目(KC22161)。
关键词
人工智能
卷积神经网络模型
腰椎间盘突出症
脊柱内镜手术
Artificial intelligence
Convolutional neural network model
Lumbar disc herniation
Spinal endoscopic surgery
