Shi Jiahao, Liu Baolin
College of Intelligence and Computing, Tianjin Key Laboratory of Cognitive Computing and Application, Tianjin University, Tianjin, 300350, P. R. China.
School of Computer and Communication Engineering, University of Science and Technology Beijing, Beijing, 100083, P. R. China.
Med Phys. 2020 Jul;47(7):2902-2915. doi: 10.1002/mp.14183. Epub 2020 May 13.
This work aims to establish a classification framework for the diagnosis of mild cognitive impairment (MCI) at different stages (early MCI and late MCI) through direct analysis of resting-state functional magnetic resonance imaging (rs-fMRI) signals and using the accuracy (total correct rate), specificity (correct rate of late MCI) and sensitivity (correct rate of early MCI) to validate its classification performance.
All fMR images of subjects were parcellated into 116 regions of interest (ROIs) by applying the Anatomical Automatic Labeling (AAL) template, and the average rs-fMRI signals of each ROI were extracted. The Hilbert-Huang transform (HHT) was introduced into the framework to decompose each rs-fMRI signal into a series of intrinsic mode functions (IMFs) and to analyze these nonstationary and nonlinear time-series from the perspective of multiresolution. After obtaining the instantaneous frequencies and amplitudes of all IMFs of a signal, the Hilbert weighted frequencies (HWFs) were calculated and combined into a vector as the feature of the corresponding ROI. Support Vector Machine (SVM) was implemented to classify MCI at different stages. We used the independent two-sample t-test as the feature selection method and measured the classification performance through the leave-one-out cross-validation (LOOCV) method.
Results on 77 early MCI (eMCI) and 64 late MCI (lMCI) with baseline rs-fMRI data from the Alzheimer's Disease Neuroimaging Initiative (ADNI) yielded 87.94% classification accuracy. Some of the brain regions with significant differences found by previous studies have been confirmed in this work. We found that HWF characteristics exhibited a significant downward trend in all cerebellar regions. The rs-fMRI signals in differential brain regions have not changed completely, but only altered in some narrow frequency bands. The analysis results showed that during the progress of MCI, the main changes of rs-fMRI were concentrated in IMF3, while IMFs with other indexes also contained HWF features with high SVM weights, such as Orbitofrontal superior frontal gyrus in IMF2, Insula in IMF4, and Lobule Ⅲ of vermis in IMF5, indicating that other IMFs provide important information for the diagnosis of MCI as well.
This work confirmed the classification ability of HHT-based classification framework in classification of at different stages of MCI. Through the analysis, we found that during the progress of MCI the main changes of rs-fMRI were concentrated in IMF3, and HWF characteristics showed a significant downward trend in all cerebellar regions.
本研究旨在通过直接分析静息态功能磁共振成像（rs-fMRI）信号，建立一个针对不同阶段（轻度认知障碍早期和晚期）轻度认知障碍（MCI）诊断的分类框架，并使用准确率（总正确率）、特异性（晚期MCI正确率）和敏感性（早期MCI正确率）来验证其分类性能。
通过应用解剖自动标记（AAL）模板，将受试者的所有功能磁共振图像分割为116个感兴趣区域（ROI），并提取每个ROI的平均rs-fMRI信号。将希尔伯特-黄变换（HHT）引入该框架，将每个rs-fMRI信号分解为一系列固有模态函数（IMF），并从多分辨率角度分析这些非平稳和非线性时间序列。在获得信号所有IMF的瞬时频率和振幅后，计算希尔伯特加权频率（HWF）并组合成一个向量，作为相应ROI的特征。采用支持向量机（SVM）对不同阶段的MCI进行分类。我们使用独立两样本t检验作为特征选择方法，并通过留一法交叉验证（LOOCV）方法测量分类性能。
对来自阿尔茨海默病神经影像倡议（ADNI）的77例轻度认知障碍早期（eMCI）和64例轻度认知障碍晚期（lMCI）的基线rs-fMRI数据进行分析，分类准确率为87.94%。先前研究发现的一些存在显著差异的脑区在本研究中得到了证实。我们发现HWF特征在所有小脑区域均呈现显著下降趋势。不同脑区的rs-fMRI信号并非完全改变，而是仅在某些狭窄频段发生改变。分析结果表明，在MCI进展过程中，rs-fMRI的主要变化集中在IMF3，而其他指标的IMF也包含具有高SVM权重的HWF特征，如IMF2中的眶额上额回、IMF4中的脑岛以及IMF5中的蚓部Ⅲ小叶，这表明其他IMF也为MCI的诊断提供了重要信息。
本研究证实了基于HHT的分类框架在MCI不同阶段分类中的能力。通过分析，我们发现在MCI进展过程中，rs-fMRI的主要变化集中在IMF3，且HWF特征在所有小脑区域呈现显著下降趋势。
