| 摘要: | 典型的功能性磁振造影(fMRI)分析一般皆奠基於線性假設的血液動力模型非時變性的特性並且專注於群體間差異的結果。近年來靜息態功能性磁振造影這一類新型的分析技術引起了原本較少探討的大腦的動態變化以及個體差異等新興議題。在目前的動態分析方法之中，希爾伯特黃轉換(HHT)建立在非線性以及非穩態架構之下，不僅提供了較佳的時頻分析解析度以及在生醫訊號中所可以代表著生理意義。然而傳統HHT分析方法原為針對一維的時間信號進行分析，鮮少用在四維的功能性磁振造影信號上。因此在論文初步應用HHT動態分析技術功能性磁振造影以探討大腦在運作或靜止時的動態變化。
具體來說，HHT是由經驗模態分解法(EMD)以及希爾伯特時頻分析(HSA)兩部分所構成。我們嘗試利用EMD於提升fMRI信號之敏感度，並以HSA觀察信號的時頻變化。在第一部分中，我們利用整體經驗模態分解法(EEMD)取代傳統的經驗模態分解法以解構原始fMRI信號，因為EEMD對於fMRI生理雜訊的干擾較不敏感。我們利用了四種在fMRI信號上可能發生的非穩態模擬訊號來評估EEMD是否能有效的應用於fMRI。在第二部分中，利用HSA觀察手指運動時以及靜息態磁振造影狀態時fMRI訊號的動態變化，並與wavelet時頻分析進行比較。另外，由於EEMD處理fMRI信號分析上的時間過於冗長，我們利用圖形運算單元(GPU)對於EEMD分析在MATLAB程式平台上進行加速。根據以上總述，在本論文中我們成功將HHT應用在fMRI資料處理，可確實提升fMRI訊號靈敏度，並可解析更細微的大腦動態資訊。
Traditionally, functional Magnetic Resonance Image (fMRI) is based on the observation of hemodynamic response function (HRF), and the fMRI analysis follows the assumptions including the time invariance and group analysis. Recently, the resting-state fMRI technology attracts public attention and addressed new issues like brain dynamics and individual differences, which are rarely explored. Among all dynamic analyses methods, the Hilbert-Huang Transform (HHT), based on non-linearity and non-stationary, not only provides better resolutions in both time and frequency domains, but provides physiological meanings in biomedical signals. However, original HHT was used to decompose one dimensional signal and hardly applied to the 4D fMRI signal. Therefore, in this thesis, we preliminarily applied the HHT to both task-based and resting-state fMRI data to extract the dynamic information in brain circuits.
Specifically, because HHT is composed of empirical mode decomposition (EMD) and Hilbert spectral analysis (HSA), we attempts to apply both EMD and HSA on fMRI dataset for two purposes: contrast enhancement and dynamic analysis, respectively. In the first part, we applied Ensemble Empirical Mode Decomposition (EEMD) to replace EMD because EEMD is relatively insensitive to the intrinsic noise of fMRI signal. We conducted four types of non-stationary simulations of fMRI signal to evaluate the effectiveness of EEMD. In the second part, we performed HSA analyses on both finger-tapping and resting-state fMRI datasets to observe the dynamic process within different brain regions, and compared the results with wavelet analysis. In addition, because of the long processing time of EEMD analysis, we applied the parallel computing of Graphic Processing Unit (GPU) for EEMD acceleration on the MATLAB platform. In summary, we successfully applied HHT on fMRI datasets for enhancing the sensitivity to task-induced activation and for exploring the detailed brain dynamics among brain areas. |
