RNN-based longitudinal analysis for diagnosis of Alzheimer's disease
RNN-based longitudinal analysis for diagnosis of Alzheimer's disease
复制标题
DOI：
10.1016/j.compmedimag.2019.01.005
复制
发表时间：
2019-04-01
影响因子：
5.7
通讯作者：
Liu, Manhua
中科院分区：
文献类型：
--
作者：
Cui, Ruoxuan;Liu, Manhua
阿尔茨海默病(AD)是一种不可逆转的神经退行性疾病，伴有记忆和其他精神功能的进行性损害。磁共振成像(MRI)作为一种重要的脑成像手段被广泛应用于AD的诊断和病情监测。序列磁共振成像的纵向分析对于建模和测量疾病沿时间轴的进展以获得更准确的诊断是重要的。现有的方法大多是利用核磁共振成像技术提取捕捉大脑形态异常及其纵向变化的特征，然后设计一个分类器来区分不同的群体。然而，这些方法有几个局限性。首先，由于特征提取和分类器模型是独立的，提取的特征可能不能完全捕捉与AD相关的脑异常的特征。其次，对于一些受试者来说，纵向磁共振图像可能在某些时间点丢失，这导致难以提取用于纵向分析的一致特征。本文提出了一种基于卷积神经网络和递归神经网络相结合的分类框架，用于AD诊断中结构磁共振图像的纵向分析。首先，构造卷积神经网络(CNN)来学习磁共振图像的空间特征，用于分类任务。然后，在多个时间点的CNN输出上构建具有级联三个双向门控递归单元(BGRU)层的递归神经网络(RNN)，用于提取AD分类的纵向特征。该方法不需要单独进行特征提取和分类器训练，而是联合学习空间特征和纵向特征以及疾病分类器，从而达到最优性能。此外，该方法还可以利用RNN对不同时间点的成像数据进行纵向分析建模。我们的方法被用来自阿尔茨海默病神经成像计划(ADNI)数据库的830名参与者的纵向T1加权MR图像进行评估，其中包括198名AD、403名轻度认知损害(MCI)和229名正常对照(NC)。实验结果表明，该方法对AD和NC的分类正确率分别为91.33%和71.71%，对纵向MR图像的分类正确率为71.71%。(C)2019爱思唯尔有限公司。保留所有权利。
Alzheimer's disease (AD) is an irreversible neurodegenerative disorder with progressive impairment of memory and other mental functions. Magnetic resonance images (MRI) have been widely used as an important imaging modality of brain for AD diagnosis and monitoring the disease progression. The longitudinal analysis of sequential MRIs is important to model and measure the progression of the disease along the time axis for more accurate diagnosis. Most existing methods extracted the features capturing the morphological abnormalities of brain and their longitudinal changes using MRIs and then designed a classifier to discriminate different groups. However, these methods have several limitations. First, since the feature extraction and classifier model are independent, the extracted features may not capture the full characteristics of brain abnormalities related to AD. Second, longitudinal MR images may be missing at some time points for some subjects, which results in difficulties for extraction of consistent features for longitudinal analysis. In this paper, we present a classification framework based on combination of convolutional and recurrent neural networks for longitudinal analysis of structural MR images in AD diagnosis. First, Convolutional Neural Networks (CNN) is constructed to learn the spatial features of MR images for the classification task. After that, recurrent Neural Networks (RNN) with cascaded three bidirectional gated recurrent units (BGRU) layers is constructed on the outputs of CNN at multiple time points for extracting the longitudinal features for AD classification. Instead of independently performing feature extraction and classifier training, the proposed method jointly learns the spatial and longitudinal features and disease classifier, which can achieve optimal performance. In addition, the proposed method can model the longitudinal analysis using RNN from the imaging data at various time points. Our method is evaluated with the longitudinal TI-weighted MR images of 830 participants including 198 AD, 403 mild cognitive impairment (MCI), and 229 normal controls (NC) subjects from Alzheimer's Disease Neuroimaging Initiative (ADNI) database. Experimental results show that the proposed method achieves classification accuracy of 91.33% for AD vs. NC and 71.71% for pMCI vs. sMCI, demonstrating the promising performance for longitudinal MR image analysis. (C) 2019 Elsevier Ltd. All rights reserved.