類神經網路應用於腦病變源定位分析

Abstract

本論文旨在嘗試利用類神經網路來探討腦病變源(dipole)的定位問 題，主要目的在 估測出引發癲癇症(epilepsy)的單一腦病變源位置及強 度向量。由腦電位分佈圖 (electroencephalogram, EEG)反推病變源發生 位置、強度向量為一高度非線性的對應問題 ，藉由模擬產生病變源與對 應之腦電波值訓練樣本，以訓練類神經網路學習該對應關係， 做為臨床 定位腦病變源資訊的參考。 本文中探討類神經網路對於不同分佈區 域、特性之病變源的定位效果；實驗臨床錄製 EEG訊號的腦病變源定位分 析；並模擬臨床錄製EEG訊號時 無法避免的雜訊干擾，探討不 同訊號雜 訊比(Signal to Noise Ratio)對類神經網路 定位效能衰減的影響。就類 神經網 路模型而言，提出混合式(Hybrid)網路模型，經由實驗證明，其 與常用的多層認知(Multi- Layer Perceptrons,MLPs)網路比較，的確具 有較快訓練時間、較佳定位能力的優點，極適 合做為定位分析的類神經 網路架構。 In this thesis, we use the neural networks to deal with the problem of dipole localization. We aim at estimating location, orientation, and moment strength of a single dipole which induces epilepsy in human brain. The inverseproblem is a highly nonlinear approximation process. We applied current dipolemodel to generate the brain electrical potential distribution on the scalp. The dipole and its corresponding brain potentials were used as training patternsfor the neural networks. A neural network trained to learn correspondence of dipoles to brain potential distribution can be used to estimate the dipole'slocation, orientation, and moment strength. It will be useful for clinicalapplications. In this reasearch, we investigatedthe cppability of neural network in dipole localization. The performance of the neural network depends on region of dipole localization andorientation of dipole moment. we also studied the effect of noise interferencefor the performance of neural network. We found that the accuracy of dipole localization decreased as the signal-to-noise ratio was poor. In addition, we proposed a model of hybrid network . Compared with the conventional multi-layer perceptrons network, the hybri d network indeed requires less training time and achieves better localization results. It might be a feasible neural network model for dipole localization.