基于迁移学习和深度卷积神经网络的胸腰椎骨折AI分类研究

Research on AI classification of thoracolumbar fractures based on deep convolutional neural network and transfer learning

传统的胸腰椎骨折影像辅助分类方法准确率低、泛化能力差,为此提出一种基于深度卷积神经网络方法辅助诊断的胸腰椎骨折AI分类方法。收集四川省中西医结合医院胸腰椎骨折患者CT影像图片共698张,建立数据集,其中单纯压缩性骨折(A类)279张,爆裂性骨折(B类)295张,正常(C类)124张。对传统卷积神经网络模型ResNet-50进行改进并融入迁移学习,对数据集进行训练,获得胸腰椎骨折AI分类模型。采用混淆矩阵评估预测模型分类性能,模型的训练集和验证集准确率分别为95.75%和96.36%,表明训练得到的智能分类模型具有较好的准确率和泛化能力。本文提出胸腰椎骨折影像辅助分类方法,可以提高人工诊断的效率和准确率。

The traditional thoracolumbar fracture image-assisted classification method has low accuracy and poor generalization ability.Therefore,based on deep convolutional neural network,this paper proposes an AI classification method for thoracolumbar fracture for auxiliary diagnosis.Firstly,a total of 698 CT images of patients with lumbar spine fractures were collected from Sichuan Integrative Medicine Hospital,and a data set was established,including 279 compression fractures(category A),295 burst fractures(category B),and 124 normal(category C).Secondly,the convolutional neural network model ResNet-50 was modified and combined with transfer learning to train,verify and test the data set to obtain the AI classification model of thoracolumbar fracture.Then,the Confusion Matrix is used to evaluate the prediction model.The accuracy of the training set and the validation set of the model is 95.75%and 96.36%,respectively,indicating that the model obtained by training has good accuracy and generalization ability.This paper proposes an image-assisted classification method for thoracolumbar fracture,which can improve the efficiency and accuracy of manual diagnosis.