基于DenseNet-SRC的多组慢性鼻窦炎计算机辅助诊断模型研究

Computer aided diagnosis model of multi-group chronic rhinosinusitis based on DenseNet-SRC

目的:基于稀疏表示分类的密集神经网络(DenseNet-SRC),搭建多组慢性鼻窦炎计算机辅助诊断模型,实现多组慢性鼻窦炎自动高效分类识别,以辅助医生诊断,提高诊断精确度。方法:首先,使用迁移学习预训练密集神经网络模型初始化参数;其次,采用蝶窦、额窦、筛窦、上颌窦4组慢性鼻窦炎数据训练密集神经网络,提取全连接层的特征向量;然后,采用稀疏表示分类器(SRC)对特征向量进行表示,求解系数矩阵,利用残差相似度进行分类;最后,将DenseNet-SRC与DenseNet-SVM模型及其他卷积神经网络模型进行对比实验。结果:本研究提出的DenseNet-SRC模型的分类准确度达到99.71%,高于DenseNet-SVM及其他卷积神经网络模型。结论:基于DenseNet-SRC的多组慢性鼻窦炎计算机辅助诊断模型,具有较好的分类识别准确率,能提高临床诊断的精确度,为多组慢性鼻窦炎的诊断提供参考。

Objective To build a computer-aided diagnosis model for multi-group chronic rhinosinusitis based on a dense neural network with sparse representation classification(DenseNet-SRC),and to achieve automatic and efficient classification and recognition of multi-group chronic rhinosinusitis to assist physicians in diagnosis and improve diagnostic accuracy.Methods First,the initialization parameters of the dense neural network model were pretrained using migration learning.Second,four groups of chronic rhinosinusitis data including sphenoid sinus,frontal sinus,ethmoid sinus,and maxillary sinus were used to train the dense neural network,and the feature vectors of the fully connected layer were extracted.Then,the sparse representation classifier(SRC)was used to represent the feature vectors,solve the coefficient matrix and use the residual similarity to classify.Finally,the Dense Net-SRC was compared with the Dense Net-SVM model and other convolutional neural network models for experiments.Results The classification accuracy of the Dense Net-SRC model proposed in this paper was 99.71%,which was higher than that of Dense Net-SVM and other convolutional neural network models.Conclusion The Dense Net-SRC-based computer-aided diagnosis model for multi-group chronic rhinosinusitis has better classification and recognition accuracy,which can improve the accuracy of clinical diagnosis and provide an effective reference for the diagnosis of multi-group chronic rhinosinusitis.