腦電位分佈圖分類以提高類神經網路之定位效能

Abstract

腦病變源之定位問題在EEG(electroencephalograph)研究領域中一向 佔有極重要之地位.過去為了解決此問題,電偶極模型被廣泛地用來描述腦 病變源和頭皮上腦電位之關係,而除了電偶極模型之數學表示式外,類神經 網路(ANN)常被用來解決一高度非線性之定位問題.先前之研究指出使用 ANN作定位之精準度是不足與數值方法作比較.而於本論文中,則預 先對腦 電位分佈圖作分類處理,以期改善ANN之定位效能. 本論文之構思 是依據腦電位之分類群數,建立一套由數個NN所組成之病變源定位系統. 其中,每個ANN則一一針對給定之類群作病變源定位之訓練.因此,對某一腦 電位分佈,預先以K-means分群演算法作判別的動作.繼之,才以判為所屬群 集之相對應子網路進行定位的工作. 除了此一整合性ANN系統外, 由於RBF NN本身對輸入資料具有分類之功能,因此,本文更廣泛地評估RBF NN應用在腦病變源定位之可行性.但,使用RBF NN 之方法卻需要更長的訓 練時間,才能達到整合性ANN系統的效能. Source localization is of great interest in the EEG( electroencephalograph)study. The current dipole model has been widely used to solve the problem. In addition to the mathematical model, artificial neural network(ANN) model has been applied to solve this highly nonlinear problem. Pervious study showed that the accuracy of localization using the ANN is poor compared with the mathathical approach. In this thesis, we presume a pre-classification of the brain potential mapping will improve the performance of ANN. The idea is to contruct an integrated system involving a number of ANN models according to the number of clusters of the brain potential mappings. Each ANN model is intentionally trained to localize the dipole source for a given cluster. Hence, the brain potential mapping is first classified by the K-means clustering algorithm. Then, according to the classified cluster, the corresponding sub-ANN model is selected to perform the task of dipole localization. In addition to the integrated ANN system, we extensively investigate the feasibility of the RBF(Radial Basis Function)ANN in dipole location since the RBF ANN itself possesses the capability of pattern classification. However,the RBF ANN approach requires longer training time to archive the performance the integrated ANN system.