Li Zheng1, Long Qian1, Dandan Zheng2, and Jiahong Gao3,4
1Department of Biomedical Engineering, College of Engineering, Peking University, Beijing, China, People's Republic of, 2GE Healthcare, MR Research China, Beijing, Beijing, China, People's Republic of, 3Beijing City Key Lab for Medical Physics and Engineering, Institute of Heavy Ion Physics, School of Physics, Peking University, Beijing, China, People's Republic of, 4Center for MRI Research, Academy for Advanced Interdisciplinary Studies, Peking University, Beijing, China, People's Republic of
Synopsis
The early detection of MCI is of paramount importance for possible delay of the transition from MCI to AD. Recently, several resting-state fMRI based neural imaging studies have been applied for MCI diagnosis by the aid of pattern classification recently. In current study, CEEMD-based high-dimensional pattern classification framework was proposed to identify MCI individuals from subjects who experience normal aging with an accuracy of 93.3 percent, compared to conventional method for brain oscillation separation. In addition, the most discriminant regions selected by our method also reflected the association with MCI, to some degree.Introduction and Purpose
Mild cognitive
impairment (MCI), often a prodromal stage of Alzheimer’s disease (AD), is a
good target for early diagnosis and therapeutic interventions of AD. The early
detection of MCI is of paramount importance for possible delay of the
transition from MCI to AD. Nevertheless, MCI is difficult to diagnose due to
its very mild symptoms of cognitive impairment.
Recently, several resting-state function MRI based neural
imaging studies have been applied for MCI diagnosis by the aid of pattern
classification [1, 2, 3]. To improve the accuracy of classification, in current study, a high-dimensional
pattern classification framework was proposed to identify MCI individuals from
subjects who experience normal aging based on multi-spectrum brain networks[3].
Method
Eight-one subjects including 44 patients with MCI and 37 normal controls (NC) were analyzed. The data were acquired on a 3.0 Tesla (T) MRI scanner. An overview of our study method is shown in Figure 1 and described as follows. First, Standard image preprocessing was performed by using both FMRIB Software Library (FSL) and Analysis of Functional NeuroImaging (AFNI). Then, a data-driven method, complementary ensemble empirical mode decomposition (CEEMD)
[4], adaptively decomposes the time course of each voxel into five time series termed intrinsic mode functions (IMFs). Each IMF occupies a unique frequency range: IMF1 0.1-0.25Hz, IMF2 0.04-0.1Hz, IMF3 0.02-0.04Hz, IMF4 0.01-0.02Hz, IMF5 0-0.01Hz. In addition, traditional methods of brain oscillation separation, Fourier Transform and Wavelet Transform were used to decompose signal similarly. After signal decomposing, Pearson’s correlation coefficient was applied to construct a set of frequency-specific inter-regional correlation matrices for each subject (90 x 90, AAL template). In the next step, graph theoretic analysis was used to characterize topological properties for each subject. And t-test method and a nested full leave-one-out cross-validation strategy were used to feature selection. Finally, pattern classification method was performed with graph theoretic analysis and support vector machines. Receiver operating characteristic (ROC) curve was used to analysis the classifying ability of a classifier.
Results
As show in table 1, our method show highest accuracy to identify MCI individuals from normal subjects compared with traditional methods of brain oscillation separation, Fourier Transform and Wavelet Transform. The classification accuracy of CEEMD method increased by more than around 10 % while the AUC value of that increased by more than 0.097, indicating significant improvement in diagnostic power. The ROC curves of three compared conditions are shown in Figure 2.
14 regions most frequently selected during the construction of optimal SVM model were are defined as the most discriminant regions and graphically displayed in Figure 3, involving Heschl gyrus (right and left), Amygdala (right and left), Superior frontal gyrus, medial (right), Hippocampus (right and left), Putamen (left), Paracentral lobule (right), Parahippocampal gyrus (left), Superior frontal gyrus, orbital part (right), Thalamus (left), Precuneus (right), Anterior cingulate gyri (right).
Discussion and
Conclusion
In this study, CEEMD-based high-dimensional pattern classification framework was proposed to accurately identify MCI individuals from from age-matched healthy controls. Compared to conventional method of brain oscillation separation, our meathod demonstrated higher classification performance with an accuracy of 93.3 percent. This result may be associated with intrinsic brain oscillations with distinct frequency bands based on fMRI. In addition, the most discriminant regions such as Hippocampus, which is responsible for memory and learning, and Frontal Gyrus associated with attention function reflects the association with MCI, to some degree. So the proposed method is proved rationality and effectiveness from side-taking. In summary, our method demonstrated for the first time that the CEEMD-based multi-spectrum brain network showed the higher classification accuracy than traditional method of brain oscillation separation. Importantly, our study provides an effective method for brain oscillation separation, by which a more useful multi-spectral classification framework for identification of MCI may be discovered.
Acknowledgements
No acknowledgement found.References
[1] Ni H, Zhou L, Ning X, et al. Exploring multifractal-based features for mild Alzheimer's disease classification [J]. Magnetic resonance in medicine, 2015.
[2] Jie B, Zhang D, Wee C Y, et al. Topological graph kernel on multiple thresholded functional connectivity networks for mild cognitive impairment classification [J]. Human brain mapping, 2014, 35(7): 2876-2897.
[3] Wee C Y, Yap P T, Denny K, et al. Resting-state multi-spectrum functional connectivity networks for identification of MCI patients [J]. PloS one, 2012, 7(5): e37828.
[4] Yeh J R, Shieh J S, Huang N E. Complementary ensemble empirical mode decomposition: A novel noise enhanced data analysis method[J]. Advances in Adaptive Data Analysis, 2010, 2(02): 135-156.