法医学杂志

法医学杂志 ›› 2020, Vol. 36 ›› Issue (2): 210-215.DOI: 10.12116/j.issn.1004-5619.2020.02.012

• 论著 • 上一篇    下一篇

卷积神经网络在识别等速运动配合程度中的应用

陈邵文1， 崔丹妮2， 夏晴3， 夏文涛3， 江洁清1， 沈忆文1   

  1. 1. 复旦大学基础医学院法医学系，上海 200032； 2. 复旦大学生物医学研究院，上海 200032； 3. 司法鉴定科学研究院 上海市法医学重点实验室 司法部司法鉴定重点实验室 上海市司法鉴定专业技术服务平台，上海 200063
  • 发布日期:2020-04-25 出版日期:2020-04-28
  • 通讯作者: 沈忆文，女，主任法医师，硕士研究生导师，主要从事法医临床学和法医病理学研究；E-mail：shenyiwen@fudan.edu.cn
  • 作者简介:陈邵文（1994—），女，硕士研究生，主要从事法医临床学研究；E-mail：17211010031@fudan.edu.cn
  • 基金资助:
    “十三五”国家科技支撑计划资助项目（2016YFC0800701）

Application of Convolutional Neural Network in Identifying Different Levels of Isokinetic Exercise Efforts

CHEN Shao-wen1, CUI Dan-ni2, XIA Qing3, XIA Wen-tao3, JIANG Jie-qing1, SHEN Yi-wen1   

  1. 1. Department of Forensic Medicine, School of Basic Medical Sciences, Fudan University, Shanghai 200032, China; 2. Institutes of Biomedical Sciences, Fudan University, Shanghai 200032, China; 3. Shanghai Key Laboratory of Forensic Medicine, Key Laboratory of Forensic Science, Ministry of Justice, Shanghai Forensic Service Platform, Academy of Forensic Science, Shanghai 200063, China
  • Online:2020-04-25 Published:2020-04-28

PDF (PC)

567

摘要： 目的 开发识别不同配合程度下等速膝关节运动力矩-时间图的卷积神经网络（convolutional neural network，CNN）模型。 方法 200名健康青年志愿者分别在30°/s和60°/s角速度下各进行两次、间隔45 min的等速向心右侧膝关节全力和半力屈伸往复运动，收集力矩-时间图。200名受试者随机分为训练集（140名）与测试集（60名），用训练集受试者的力矩-时间图训练CNN模型，再用训练好的模型预测测试集内图形的类别。共进行3次随机取样与模型开发。 结果 在等速膝关节全力及半力运动条件下各收集2 400张力矩-时间图。3次训练的CNN模型分类准确率分别为91.11%、90.49%和92.08%，平均准确率为91.23%。 结论 本研究开发的CNN模型对全力及半力等速力矩-时间图具有较好的区分效果，有助于识别受试者在等速膝关节运动过程中的配合程度。

关键词: 法医学, 等速运动, 膝关节, 伪装, 卷积神经网络

Abstract: Objective To develop a convolutional neural network （CNN） that can identify isokinetic knee exercises moment of force-time diagrams under different levels of efforts. Methods The 200 healthy young volunteers performed concentric isokinetic right knee flexion-extension reciprocating exercises with maximal effort as well as half the effort at angular velocities of 30°/s and 60°/s twice, respectively, with an interval of 45 min. The moment of force-time diagrams were collected. The 200 subjects were randomly divided into the training set （140 subjects） and the testing set （60 subjects）. The moment of force-time diagrams from subjects in the training set were used to train CNN model, and then the fully trained model was used to predict types of curves from the testing set. Random sampling of subjects along with subsequent development of models were performed 3 times. Results Under the circumstances of isokinetic knee exercises with maximal effort and half the effort, 2 400 moment of force-time diagrams were produced, respectively. Classification accuracy rates of the CNN models after the 3 trainings were 91.11%, 90.49% and 92.08%, respectively, and the average accuracy rate was 91.23%. Conclusion The CNN models developed in this study have a good effect on differentiating isokinetic moment of force-time diagrams of maximal effort exercises from those made with half the effort, which can contribute to identifying levels of efforts exerted by subjects during isokinetic knee exercises.

Key words: forensic medicine, isokinetic exercise, knee joint, camouflage, convolutional neural network