基于AlexNet模型的AD分类  

Efficient Alzheimer’s Disease Classification Based on AlexNet Model

作　　者：张柏雯 林岚[1] 吴水才[1] ZHANG Baiwen;LIN Lan;WU Shuicai(College of Life Science and Bioengineering,Beijing University of Technology,Beijing 100124,China)

机构地区：[1]北京工业大学生命科学与生物工程学院

出　　处：《北京工业大学学报》2020年第1期68-74,共7页Journal of Beijing University of Technology

基　　金：北京市自然科学基金-海淀原始创新联合基金资助项目(L182010);北京市教委计划一般项目(KM201810005033);国家科技支撑计划课题(2015BAI02B03)

摘　　要：阿尔兹海默症(Alzheimer’s disease,AD)一经发现难以逆转,早期诊断对延缓AD的病程发展非常重要.虽然深度卷积网络近年来在图像识别领域有着大量突出的表现,但将从自然图像中训练得到的二维经典的深度网络直接运用到三维的结构磁共振影像(structural magnetic resonance imaging,s MRI)上进行AD疾病状态的分类还存在一些问题.基于194例AD、123例晚期轻度认知障碍(late mild cognitive impairment,LMCI)与105例正常老化(normal control,NC)的s MRI,运用特征迁移学习的方法,从经典的深度卷积模型--Alex Net中提取各阶段受试者的图像特征,并对所提特征进行三维重组,再运用最大池化、主成分分析等方式降维,并运用向前序列选择方法对各分类组进行特征选择,最后运用支持向量机建立分类模型,实现AD、LMCI与NC的分类.在Alex Net的三、四、五层卷积层,AD与NC的分类准确率分别为89. 93%、91. 28%、87. 25%,AD与LMCI的分类结果分别为80. 77%、76. 92%、78. 21%,NC与LMCI的分类结果分别为72. 46%、75. 45%、73. 65%.结果证明,通过经典卷积网络获得的特征,经过三维重组,能够较好地对AD实现分类.Alzheimer’s disease(AD)generally results in irreversible brain damages.Early diagnosis of disease plays an important role in preventing the progression of AD.Deep convolutional neural networks(CNN)have achieved prominent performance in the field of natural image recognition,while some problems exist in applying a classic CNN model on 3D MRI for AD classification.To address these issues,with 194 AD subjects,late mild cognitive impairment(LMCI)123 subjects and 105 normal control(NC)subjects,a hybrid computational strategy was proposed based on a pre-trained AlexNet CNN model and s MRI for AD classification.The feature presentation of pre-trained network was efficiently transferred to AD classification task by using transfer learning,3D features reconstruction,feature reduction using Max pooling and principal component analysis(PCA),and selection feature using sequential forward search and(SFS)method.Then,support vector machines(SVM)was applied to classification.The classification accuracy values on conv3,conv4,conv5 layers of AlexNet were89.93%,91.28%,and 87.25%of AD/NC,respectively,80.77%,76.92%,and 78.21%of AD/LMCI,respectively,72.46%,75.45%,and 73.65%of NC/LMCI,respectively.Results show that features extracted from a classic CNN model and their 3D reconstruction can achieve good performance on AD classification.

关 键 词：Alex Net 三维特征重组 阿尔兹海默症(AD) 结构磁共振图像 迁移学习 特征提取 

分 类 号：R445.2[医药卫生—影像医学与核医学]