文章摘要

文章信息/Info

Title:: Medical imaging; Image classification; Pneumonia diagnosis; Pathogen classification; Deep learning; Convolutional neural network

作者:: 赵爽1; 魏国辉2; 赵文华2△; 马志庆2; 1.山东中医药大学实验室管理处,济南 250355;2.山东中医药大学智能与信息工程学院,济南 250355

Author(s):: ZHAO Shuang1; WEI Guohui2; ZHAO Wenhua2; MA Zhiqing2; 1. Laboratory Management Office, Shandong University of Traditional Chinese Medicine, Jinan 250355, China；2. College of Intelligence and Information Engineering, Shandong University of Traditional Chinese Medicine, Jinan 250355

Keywords:: Medical imaging; Image classification; Pneumonia diagnosis; Pathogen classification; Deep learning; Convolutional neural network

摘要:: 为提高儿童肺炎的临床诊断准确率，进一步为肺炎治疗精准用药提供依据，本研究提出了一个级联的特征正交融合网络模型对小儿胸片图像进行自动分类，将预处理的图像输入到网络A用来诊断是否患肺炎，然后将网络A的输出作为网络B的输入，判断肺炎的病原体类型。网络A和网络B均以深度残差网络(ResNeXt-50)为基础网络, 首先将压缩-激励模块(squeeze-and-excitation networks, SE-Net)融合到ResNeXt-50中，然后利用空洞卷积获取多尺度特征，并通过自注意力机制获得网络中具有代表性的局部特征，从局部特征中提取与全局特征正交的分量。最后将正交分量与全局特征进行融合，形成最终的特征表征并完成分类。实验结果表明，该模型在二分类模型的分类准确率达到97.78%，在三分类的准确率达到85.13%。该模型具有良好的分类效果，可帮助医生实现对儿童肺炎快速有效的临床诊断。

Abstract:: To improve the accuracy of clinical diagnosis of pneumonia in children and provide a basis for the precise medication of pneumonia treatment, a cascaded convolutional neural network model was proposed to automatically classify children′s chest radiograph images. The preprocessed image was input to network A for diagnosis of pneumonia, and then the network A′s output was used as the input of network B to determine the type of pathogen in the case of pneumonia. Among them, network A and network B were based on the deep residual network (ResNeXt-50). The squeeze-and-excitation networks (SE-Net) was integrated into ResNeXt-50. The representative local information was extracted from the network using multi-scale convolution and self-attention methods to concentrate, the components orthogonal to the global information were extracted from the local information. Finally, the orthogonal component and the global information were complementarily connected to form the final feature representation. The experimental results showed that the model had an accuracy of 97.78% in the binary classification and 85.13% in the triple classification. It indicates that this model has good classification results，can help doctors quickly and effectively realize the clinical diagnosis of childhood pneumonia.