發展類神經網路用於禪定組與控制組的腦電波源定位

Development of Artificial Neural Network for Source Localization of Zen-meditation and Resting EEG

葉家祺(Chia-Chi Yeh)

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(2018年)

摘要

腦功能的源定位是大腦研究領域的重要議題,建立大腦的電位模型是解決此問題的其一方法。一般而言,我們透過不斷嘗試偶極子的參數,直到找到偶極子產生的電位和真實量測的電位有最佳的對應。假設頭為一有多層殼並帶有不同導電度的半球模型,將小範圍內的神經元視為一產生電位的偶極子,以此建立電位模型模擬真實的腦電位訊號。在採用一近似的模型來建立多層半球殼的腦電位模型,並開發一套尋找最適偶極子位置與強度的演算法。然而,上述的方法可能非常花時間並且難以收斂。
此篇論文採用類神經網路實現在腦模型中的腦功能的源定位,並分析偶極子的特徵,在類神經網路部分,我們採用倒傳遞前饋網路以達到減少運算時間的目的,而類神經網路的分析結果也將和上述提及的腦電位模型為基礎發展出的演算法做比較。在這裡,我們探討在類神經網路架構以及訓練資料數量的選擇,對於結果和計算效率的影響中做權衡。最後,我們利用訓練好的網路分析 15 分鐘實驗組(禪定)與控制組(閉眼休息)的訊號,並比較實驗組與控制組腦功能的空間特性差異。
結果顯示類神經網路在腦功能的源定位分析中,提供了較有效率的方法,此外,對於實驗結果類神經網路也比腦電位模型為基礎發展出的演算法有較好的表現。在腦功能的空間特性差異,實驗組的第一第二偶極子皆集中在腦深部,左顳葉、左額葉區域,特別是在深部的額葉-中樞區域,此為禪定者將意識集中在禪心脈輪的結果。控制組的第一偶極子較易出現在左頂葉及左右枕葉,這些區域負責處理視覺處理以及空間感官認知處理。第二偶極子則多集中在右顳葉、右額葉及左枕葉。

關鍵字

腦電波源定位 ； 類神經網路

並列摘要

Source localization of specific brain functioning is an important task in the field of brain research. The general way is to iteratively determine the location and orientation of the dipole source until optimal matching is reached between the dipole-generated brain potentials and the measured potentials on the head. The idea is based on constructing a dipole model to simulate real EEG. The dipole model considers the brain as a volume conductor with multiple concentric shells and assumes
active neurons within a small region of the brain forming a current dipole. However, the scheme mentioned above is remarkably time-consuming and, worst of all, the iteration process may not
converge.
This thesis is aimed to develop the artificial neural network (ANN) to realize the brain focal activities modeled by the dipoles embedded in a four concentric-shell spherical model. Particularly feed forward back-propagation neural networks are studied in order to reduce the calculation time. The performance of ANN scheme is compared with which of the dipolar-model based numerical method. In both one and two-dipole localization, we investigate the effect of ANN structure and number of training data on the ANN performance. These factors are used to manage the trade-off between source-localization performance and computational efficiency. Finally, to compare the spatial characteristics of focalized sources (dipoles) between Zen-meditation EEG and resting EEG, we apply the trained network to determine the focalized sources of 15-minute Zen-meditation and
resting EEGs and to analyze spatial statistics of dipole features.
Our results show that the ANN provides an efficient and effective scheme for dipole source
localization. Besides, the performance of neural network is better than the performance of numerical
method based on the simulated brain potentials. In dipole features analysis, both primary and
secondary dipoles of Zen-meditation EEG concentrate in deeper brain, with higher probability of
locating in left temporal, left frontal regions and fronto-central regions. It reflects the effects of Zen-
meditation practice at ChanXin-FaYan mailuns. With respect to the resting EEG, primary dipoles tend to locate in left and right parietal, which are responsible for reasoning as well as integration of sensory information among various modalities and perception and cognition processing of spatial information. In resting EEG, secondary dipoles relatively distribute in left parietal, right temporal and frontal regions evenly.

並列關鍵字

EEG ； Source localization ； Artificial Neural Network

參考文獻

doi: 10.3389/fnhum.2012.00133. PMID: 22615691, 2012.

Medicine, 18(6), pp. 589-596, doi:10.1089/acm.2011.0038. PMID: 22784346, 2012.

doi:10.1017/S0033291709991747. PMID:19941676, 2010.

cognition, 19(4), pp. 1110-1118, doi: 10.1016/j.concog.2010.01.007. PMID: 20167507, 2010.

medicine, 31(03), pp. 499-507, doi: 10.1142/S0192415X03001132. PMID: 12943181, 2003.