基于深度学习的髋关节应力分布算法研究

Research on hip joint stress distribution algorithms based on deep learning

针对髋关节软骨的应力分布算法研究问题,设计了一个基于深度学习模型来代替有限元分析。该深度学习模型分为无监督学习模块和有监督学习模块,首先使用无监督学习模块对髋关节的软骨和股骨进行形状编码;之后实现对应力分布数据的编码与解码,使得应力数据能够与神经网络相结合;然后通过监督学习,利用编码好的应力数据进行监督,使神经网络学习得到一个从髋关节软骨和股骨的形状码到应力分布的应力码的映射关系;最终得到一个拟合的深度学习模型。此模型能够在一定程度上模拟有限元分析方法,但是由于其平均绝对误差和归一化平均绝对误差比较大,所以还不能完全替代有限元分析方法。在此基础上,进一步探索了新模型在特征利用上的局限,并提出了改进的方向。

Aiming at the problem of the stress distribution algorithm of hip cartilage,a deep learning model to replace the finite element analysis(FEA)was proposed.This deep learning model was divided into unsupervised learning module and supervised learning module.Firstly,an unsupervised learning module was adopted to encode the shape of hip cartilage and femur.Then the coding and decoding of stress distribution implement was implemented so that stress data can be combined with the neural network.Next a supervised learning module supervised by the stress data was used,and the model uses neural networks to learn a mapping relationship from the shape code of the hip cartilage and femur to the stress code of the stress distribution.Finally,a fitted deep learning model was obtained.This deep learning model can simulate the FEA method to a certain extent.But the mean absolute error and the normalized mean absolute error are still larger than that of the FEA method,so the FEA method cannot be completely replaced by our deep learning model.Meanwhile,the limitations of the deep learning model in the use of input features were studied,and a direction to improve the performance of the model was proposed.