CNN應用於禪定與放鬆休息腦電波之頻率空間特性分類

Classification of Zen-meditation EEG and Resting EEG Spatial-spectral properties by Convolutional Neural Network

王少宏(Shao-Hong Wang)

指導教授 ： 羅佩禎

國立交通大學/電機學院/電機與控制工程系所/碩士(2019年)

摘要

本論文中，我們利用機器學習，自組織映射網路(SOM)去對禪定及放鬆休息狀態下的腦電波加以分群，並試圖探索禪定與休息狀態下腦電波(30通道)質心頻率的空間特性，且利用SOM分群的結果對腦電波進行標記，再利用深度學習，卷積神經網路(CNN)對其產生的腦電圖(brain mapping of 30 centroid frequencies, BMFc)加以分類。為了最佳化SOM分群的結果，本論文比較不同參數設定下的分群結果，並根據分群的結果，可以確定與比較禪定與休息腦電波質心頻率空間屬性的特徵。卷積神經網路(CNN)是一種前饋、誤差反向傳播之深度學習模型，被廣泛應用於圖像的分類，針對不同的分類目的及資料集複雜度，設計者可設計不同之CNN架構。此篇論文比較了不同CNN模型在分類BMFc上之表現。禪定下腦電圖之最高分類率為94.61% (Model-4)，放鬆休息之腦電圖最高分類率為95.88% (Model-2)，禪定下腦電圖之平均分類率為88.87%，放鬆休息之腦電圖平均分類率為88.73%。
關鍵字:腦電波，連續小波轉換，自組織映射網路，卷積神經網路，禪定

關鍵字

腦電波 ； 連續小波轉換 ； 自組織映射網路 ； 卷積神經網路 ； 禪定

並列摘要

This thesis is aimed to investigate the spatial-spectral properties of 30-channel Zen-meditation and resting EEG (electroencephalograph) based on classification of brain mapping of centroid frequency by self-organizing map (SOM) and convolutional neural network (CNN) models. Input data entry is the brain mapping of 30 centroid frequencies (abbreviated as BMFc) extracted from CWT (continuous wavelet transform) coefficients of each channel. Based on the unsupervised learning scheme, SOM mainly performs the matching of the input feature vector Fc and the cluster center representing the quantitative features of the output cluster. In addition, to optimize the clustering results, it is necessary to carefully select the implementation parameters of SOM. From the clustering results, the major spatial-spectral features of Zen-meditation and resting EEG may be determined and compared. With the clustering results by SOM, we are able to label each BMFc and train the CNN models to classify the dataset of BMFc’s. CNN is a hybrid feedforward, error backpropagation model in deep learning that is mainly applied in image classification. The structure of CNN can be designed to match different dataset. This study compares the performance of different CNN models on BMFc classification. The best classification accuracy achieved for classifying Zen-meditation BMFc is 94.61% (Model -4) and for classifying resting BMFc is 95.88% (Model-2). Average classification accuracy is 88.87% for Zen-meditation BMFc and 88.73% for resting EEG.

Keyword: Electroencephalograph (EEG), continuous wavelet transform (CWT), Convolutional neural network (CNN), Self-organizing map (SOM), centroid frequency, Pearson correlation coefficient (PCC), Zen meditation.

並列關鍵字

Electroencephalograph ； continuous wavelet transform ； Convolutional neural network ； Self-organizing map ； centroid frequency ； Pearson correlation coefficient ； Zen meditation

參考文獻

[1] Huang, H-Y., and P-C. Lo. "EEG dynamics of experienced Zen meditation practitioners probed by complexity index and spectral measure." Journal of medical engineering & technology vol.33, no.4 pp. 314-321, 2009.

[2] Lehmann, Dietrich, et al. "Reduced functional connectivity between cortical sources in five meditation traditions detected with lagged coherence using EEG tomography." Neuroimage vol.60, no.2 pp.1574-1586, 2012.

[3] Forman, Steven D., et al. "Improved assessment of significant activation in functional magnetic resonance imaging (fMRI): use of a cluster‐size threshold." Magnetic Resonance in medicine vol.33, no.5 pp.636-647, 1995.

[4] M. Le Van Quyen, C. Adam, M. Baulac, J. Martinerie and F. Varela, “Nonlinear interdependencies of EEG signals in human intracranially recorded temporal lobe seizures,” Brain Research, vol. 792, no. 1, pp. 24-40, 1998.

[5] M.Hämäläinen, R. Hari, R. Ilmoniemi, J. Knuutila and O. Lounasmaa, “Magnetoencephalography—theory, instrumentation, and applications to noninvasive studies of the working human brain,” Reviews of Modern Physics, vol. 65, no. 2, pp. 413-497, 1993.