基于多模态医学图像的Alzheimer病分类方法  

Alzheimer’s disease classification method based on multi-modal medical images

作　　者：韩坤 潘海为[1] 张伟[2] 边晓菲 陈春伶 何舒宁 HAN Kun;PAN Haiwei;ZHANG Wei;BIAN Xiaofei;CHEN Chunling;HE Shuning(College of Computer Science and Technology,Harbin Engineering University,Harbin 150001,China;Modern Education Technology Center,Heilongjiang University,Harbin 150001,China)

机构地区：[1]哈尔滨工程大学计算机科学与技术学院,哈尔滨150001 [2]黑龙江大学现代教育技术中心,哈尔滨150001

出　　处：《清华大学学报（自然科学版）》2020年第8期664-671,682,共9页Journal of Tsinghua University(Science and Technology)

基　　金：国家自然科学基金资助项目(61672181)。

摘　　要：多模态医学影像信息已经在计算机辅助检测和诊断中被广泛地应用。在对Alzheimer病(Alzheimer’s disease, AD)的分类与诊断中,结合多个模态医学影像的特征信息能够更加准确且全面地对同一AD主题进行分类与诊断。该文提出了一种基于卷积神经网络的模型结构,分别对AD病患的磁共振图像(MRI)和正电子发射型计算机断层显像(PET)图像进行3D卷积操作来提取各自模态的特征信息,并应用模型融合方法对模态特征信息加以融合,从而得到包含更加丰富的多模态特征信息。最后用全连接神经网络将上述提取的多模态特征信息进行分类预测。通过在AD神经影像学倡议(Alzheimer’s disease neuroimaging initiative, ADNI)公开数据集上的实验结果表明:该文所提出的模型在准确率(accuracy, ACC)和曲线下面积(area under the curve, AUC)的性能评价中都取得了更加优越的结果。Multi-modal medical image information has been widely used in computer-aided detection and computer-aided diagnosis in the medical community.Feature information from multi-modal medical images can be used to accurately classify and diagnosis Alzheimer's disease(AD)characteristics.This paper presents a convolutional neural network model for 3D convolution operations on magnetic resonance imaging(MRI)and positron emission computed tomography(PET)images of Alzheimer's subjects to extract the feature information for the various modalities.Then,models are used to fuse these modal information sets into a rich multi-modal feature information dataset.Finally,this dataset is classified and predicted using a fully connected neural network.Tests on the public data set of the AD neuroimaging initiative show that this model more accurately evaluates accuracy(ACC)and area under the curve(AUC)conditions.

关 键 词：Alzheimer病(AD) 磁共振影像(MRI) 正电子发射型计算机断层显像(PET) 图像分类 深度学习 卷积神经网络 

分 类 号：TP391.7[自动化与计算机技术—计算机应用技术]