基于卷积神经网络的儿童病毒性脑炎磁共振影像分类与早期诊断研究  

Classification and early diagnosis of children viral encephalitis on MRI images based on convolutional neural network

作　　者：黄坚 余卓 徐璐[4] 周海春[5] 俞刚[1,2] HUANG Jian;YU Zhuo;XU Lu;ZHOU Haichun;YU Gang(Department of Data and Information,the Children's Hospital Zhejiang University School of Medicine,Hangzhou 310052,China;Sino-Finland Joint AI Laboratory for Child Health of Zhejiang Province,Hangzhou 310052,China;Department of Scientific Research,Huiying Medical Technology(Beijing)Co.,Ltd.,Beijing 100192,China;Department of Neurology,the Children's Hospital of Zhejiang University School of Medicine,Hangzhou 310052,China;Department of Radiology,the Children's Hospital Zhejiang University School of Medicine,Hangzhou 310052,China)

机构地区：[1]浙江大学医学院附属儿童医院数据信息部,杭州310052 [2]浙江-芬兰儿童健康人工智能联合实验室,杭州310052 [3]慧影医疗科技(北京)股份有限公司科研部,北京100192 [4]浙江大学医学院附属儿童医院神经内科,杭州310052 [5]浙江大学医学院附属儿童医院放射科,杭州310052

出　　处：《磁共振成像》2023年第1期54-60,共7页Chinese Journal of Magnetic Resonance Imaging

基　　金：国家重点研发计划(编号:2019YFE0126200);国家自然科学基金面上项目(编号:62076218)。

摘　　要：目的构建基于卷积神经网络的儿童病毒性脑炎MRI分类与早期诊断模型,探讨其对儿童病毒性脑炎早期诊断、精准治疗和改善患儿预后的价值。材料与方法收集浙江大学医学院附属儿童医院2020至2022年期间颅脑MRI影像数据1077例,其中病毒性脑炎患儿577例,非病毒性脑炎儿童500例。运用卷积神经网络中的Squeeze-and-Excitation Residual Networks(SE-ResNet)模型构建儿童病毒性脑炎MRI分类与早期诊断模型并与Convolutional Block Attention Module Residual Networks(CBAM-ResNet)、Mobile Networks(MobileNet)、Residual Networks(ResNet)、Shuffle Networks(ShuffleNet)模型进行了对比。结果所有模型在训练集上都达到了收敛。SE-ResNet、CBAM-ResNet、MobileNet和ShuffleNet模型在训练集训练100轮后准确率都达到90%以上,而只有CBAM-ResNet模型和本研究选用的SE-ResNet模型在验证集上同样取得了90%以上的准确率。在测试集上,CBAM-ResNet具有最高的准确率73.91%,ResNet具有最高的召回率75.45%,但只有本文所用SE-ResNet模型在准确率和召回率都达到较高水平,并且取得最好的F1得分和曲线下面积(area under the curve,AUC)值:准确率为70.83%,召回率为72.73%,AUC为0.77,F1得分为0.7183。结论运用人工智能技术结合MRI实现儿童病毒性脑炎早期诊断是可行的,本研究为进一步实现全面的儿童脑炎早期诊断、精准治疗和改善脑炎患儿预后提供了理论和应用基础。Objective:To establish a magnetic resonance imaging(MRI)classification and early diagnosis model of children viral encephalitis based on convolutional neural network(CNN),and to explore its value in early diagnosis,precise treatment and improvement of prognosis of children viral encephalitis.Materials and Methods:A total of 1077 cases of brain MRI data were collected from the Children’s Hospital of Zhejiang University School of Medicine from 2020 to 2022,including 577 cases with viral encephalitis(VE)and 500 cases without VE.The Squeeze-and-Excitation Residual Networks(SE-ResNet)model in CNN was used to construct the MRI classification and early diagnosis model of children viral encephalitis,and was compared with Convolutional Block Attention Module Residual Networks(CBAM-ResNet),Mobile Networks(Mobile Net),Residual Networks(Res Net),and Shuffle Networks(Shuffle Net)models.Results:All models converged on the training set.The accuracy of SE-ResNet,CBAM-ResNet,MobileNet and ShuffleNet models all reached more than 90%after 100 rounds of training in the training set,while only CBAM-ResNet model and SE-ResNet model selected in this study also achieved more than 90%accuracy in the validation set.In the test set,CBAM-Res Net had the highest accuracy rate of 73.91%,Res Net had the highest recall rate of 75.45%,yet only SE-Res Net model used in this work had a high level in both accuracy and recall,and achieved the best F1 and area under the curve(AUC):the accuracy rate was 70.83%,the recall rate was 72.73%,the AUC was 0.77,and the F1 score was 0.7183.Conclusions:The results in this work showed that it is feasible to realize the early diagnosis of viral encephalitis in children by using artificial intelligence technology combined with MR images,and provided theoretical and practical foundation for further achieving the early diagnosis and precise treatment of children viral encephalitis and improving the prognosis of children with encephalitis in a comprehensive way.

关 键 词：儿童疾病 病毒性脑炎 磁共振成像 SE-ResNet 深度学习 分类模型 早期诊断 

分 类 号：R445.2[医药卫生—影像医学与核医学] R725.1[医药卫生—诊断学]