北京大学学报自然科学版 ›› 2019, Vol. 55 ›› Issue (5): 859-864.DOI: 10.13209/j.0479-8023.2019.039

上一篇    下一篇

基于步行时足底压力信息的前交叉韧带断裂辅助诊断方法

黄红拾1, 王政飞2, 许国雄2, 李文新2, 张思1, 张东霞1, 敖英芳1,†   

  1. 1. 北京大学第三医院运动医学研究所, 北京 100191 2. 北京大学信息科学技术学院, 北京 100871
  • 收稿日期:2018-06-14 修回日期:2018-11-01 出版日期:2019-09-20 发布日期:2019-09-20
  • 通讯作者: 敖英芳, E-mail: yingfang.ao(at)vip.sina.com
  • 基金资助:
    国家自然科学基金(91646202)、北京大学医学-信息科学联合研究种子基金(BMU20160590)、北京大学第三医院院临床重点项目(BYSY2017012)和广州市产学研协同创新重大专项(201604020095)资助

Anterior Cruciate Ligament Deficiency Auxiliary Diagnosis Based on Plantar Pressure Information during Walking

HUANG Hongshi1, WANG Zhengfei2, XU Guoxiong2, LI Wenxin2, ZHANG Si1, ZHANG Dongxia1, AO Yingfang1,†   

  1. 1. Institute of Sports Medicine, Peking University Third Hospital, Beijing 100191 2. School of Electronic Engineering and Computer Science, Peking University, Beijing 100871
  • Received:2018-06-14 Revised:2018-11-01 Online:2019-09-20 Published:2019-09-20
  • Contact: AO Yingfang, E-mail: yingfang.ao(at)vip.sina.com

PDF

摘要：

为了辨识动态足底压力信息与前交叉韧带断裂的关系, 将步行时的足底压力数据转换成图像, 采用深度学习中的卷积神经网络模型, 在给定足量输入图像与分类结果的情况下, 不断更新神经网络的参数, 建立图像与前交叉韧带断裂的关系。将足底压力测试系统(FootScan®)采集的数据分为训练集和测试集两个部分。训练集用于调整模型的参数, 帮助模型更好地分析并找到足底压力信息与前交叉韧带断裂的关系; 测试集用于模拟诊断, 对比真实情况, 评估准确性, 并评估其作为临床辅助诊断方法的性能。结果表明, 提出的投票法模型的诊断正确率超过90%, 并且从得到足底压力数据到产生诊断结果, 总耗时仅3秒左右。由此得出, 所提出的基于步行时足底压力信息的深度学习模型, 可以在很短时间内辅助诊断前交叉韧带断裂, 为临床辅助诊断及康复提供参考。

关键词: 前交叉韧带断裂, 足底压力, 辅助诊断, 深度学习

Abstract:

To study the identification of dynamic anterior cruciate ligament deficiency based on plantar pressure information, using convolutional neural network, raw plantar pressure data during walking were converted into images to establish the connection between plantar pressure and anterior cruciate ligament deficiency. Given plenty of input images and classification results, convolutional neural network could update its parameters for iterations to fit the connection. Plantar pressure data collected by acquisition system (FootScan®) were divided into two parts, training set and test set. The training set was used for training the deep learning model tune the parameters, which helped the model analyze the data better, while the test set was used to generate diagnosis, compare the results to the ground-truth to evaluate the model’s accuracy, and judge its performance as an auxiliary tool for clinical diagnosis. The results show that trained deep learning model can correctly diagnose over 90% cases in the test set, and only takes about 3 seconds to make a diagnosis. The proposed dynamic plantar pressure information based deep learning model can provide auxiliary diagnosis in very short time, which provides references for the auxiliary diagnosis and rehabilitation in clinical medicine.

Key words: anterior cruciate ligament deficiency, plantar pressure, auxiliary diagnosis, deep learning

版权所有 © 《北京大学学报（自然科学版）》编辑部
技术支持：北京玛格泰克科技发展有限公司