基于高光谱图像的中医舌苔和舌质分类研究  

Classification of tongue coating and tongue texture in traditional Chinese medicine based on hyperspectral image

作　　者：蔡轶珩 潘建军 郭雅君 谢锦 任传云[2] 肖永华[2] CAI Yiheng;PAN Jianjun;GUO Yajun;XIE Jin;REN Chuanyun;XIAO Yonghua(School of Information and Communications Engineering,Beijing University of Technplogy,Beijing 100124;Dongzhimen Hospital,Beijing University of Chinese Medicine,Beijing 100700)

机构地区：[1]北京工业大学信息学部,北京100124 [2]北京中医药大学东直门医院,北京100700

出　　处：《北京生物医学工程》2023年第6期559-565,611,共8页Beijing Biomedical Engineering

基　　金：国家重点研发计划(2017YFC1703302)资助。

摘　　要：目的舌苔和舌质分类对于后续的舌象客观化诊断具有重要的作用,高光谱图像包含大量的数据信息,能够有助于分类效果提升。但是高光谱图像信息量巨大,且传统的方法提取特征不够充分,如何有效提取数据信息并促进舌诊客观化仍是个值得深究的问题。因此,本文提出面向高光谱舌图像的深度学习算法,利用深层网络来提取高光谱图像的数据信息,从而提升舌苔和舌质的分类效果。方法使用高光谱相机对图像进行采集,对采集的图像构造谱图进行预处理从而简化输入数据的冗余性;为了提取高光谱舌图像的数据信息,在算法上设计了一种可以获取底层特征的残差网络结构,加入了跳跃连接并在每个卷积层前加入批量归一化(batch normallization,BN)和带参数的ReLU(parametric rectified linear unit,PReLU)激活函数来提前激活网络,因此可以提取深层的光谱空间特征以提升分类精度。结果在高光谱舌图像数据集上的实验表明,本文算法分类精度达到93.9%,优于传统的基于RGB图像分类方法和CNN(convolutional neural network)与VGG(visual geometry group)网络。分类结果图显示,除了舌苔和舌质交界处光谱曲线相差不大的地方会有误分类的现象,分类结果与标签图基本一致。结论该深度学习算法可以较好地完成舌苔和舌质分类任务,为后续舌象特征信息的计算机自动分析提供良好基础。Objective The classification of the tongue coating and tongue texture is important for the subsequent objective diagnosis of the tongue,and hyperspectral images contain a large amount of data information that can help to improve the classification effect.However,the amount of information in hyperspectral images is huge and traditional methods are not sufficient to extract features,so how to effectively extract data information and facilitate the objectification of tongue diagnosis is still a problem worthy of further investigation.Therefore,this paper proposes a deep learning algorithm for hyperspectral tongue images,with a deep network to extract data information from hyperspectral images,so as to improve the classification effect of the tongue coating and tongue texture.Methods The images are acquired by using hyperspectral camera,and the acquired images are preprocessed to construct spectrograms so as to simplify the redundancy of the input data.In order to extract the data information of the hyperspectral tongue images,algorithmically we design a residual network structure that can acquire the underlying features by adding jump connections and adding batch normallization(BN) and parametric rectified linear unit(PReLU) activation functions before each convolution layer to activate the network in advance,so that deep spectral spatial features can be extracted to improve the classification accuracy.Results Experiments on hyperspectral tongue image dataset show that the classification accuracy of our algorithm reaches 93.9%,which is better than the traditional RGB image based classification methods and CNN(convolutional neural network),and VGG(visual geometry group) networks.The classification result map shows that our classification results are basically consistent with the label map,except for the misclassification at the junction of tongue coating and texture where the spectral curves do not differ much.Conclusions The deep learning algorithm can better perform the task of classifying tongue coating and tongue

关 键 词：高光谱 舌苔 舌质 分类 深度学习 

分 类 号：R318􀆰[医药卫生—生物医学工程]