- 學位論文
獨立成份方法分析磁振造影資料 :個別到群組的數據分析
Data analysis on MRI data using ICA :from single to group dataset
摘要
獨立成份分析法（Independent Component Analysis, ICA）把混合的訊號分解為各自獨立的訊號源，常被用在MRI，處理訊號分離或功能性定位的研究。在目前MRI 的資料處理中，感興趣的訊號皆可視為互為獨立性，例如，在血流灌注的研究(perfusion MRI)，不同的組織所發生之信號各有其特性，且在空間上的貢獻各自獨立，ICA利用這個特性來做為組織分割的依據，將受局部磁場不均影響的灰質、白質以及藉於微血管與血管間的間質組織的區域分割出來；而對於功能性的MRI資料，各種刺激所引起的 Blood-Oxygen-Level-Dependence (BOLD)訊號源在空間的貢獻亦為獨立，因此可將受刺激影響的區域分割出來，做為功能上位置的確認。 決定訊號源的個數在資料處理中，可以有效地把需處理的資料量減縮，達到節省電腦運算容量及加快運算速度，然而，在傳統的評估方法(例如, AIC、BIC、MDL)，群組間與群組內的變異性往往會造成高估的現象，而AR(1) 曲線擬合(curve fitting)的方式能較保守地估計訊號源的個數，較不受傳統方法對雜訊假設為常態分佈的限制。在本論文中，用ICA來分析一系列的MRI資料(從單一資料處理，單一群組到多群組資料)，為因應資料型態的不同，ICA技術的發展也隨之改進：第一部份為將ICA應用在血流灌注的研究，找出受顯影劑影響的血管周圍組織區域；第二部份為將群組 ICA應用到單一群組的藥物濫用研究，增加對藥物腦功能的了解並跟傳統血流動力模組的結果做比較；第三部份是為因應多群組資料的處理，提出以AR(1)模組為概念的群組式訊號源估計法來進行訊號源個數的評估與探討。第四部份是把多群組ICA結合群組式訊號源估計法的技術，應用到腦功能預設系統(Brain Default System)的研究，探討人類在休息狀態(Resting State)時，腦功能的進行在時間軸上的一致性。 初步結果顯示，利用ICA來分割血管周圍的組織訊號，以此訊號來決定的動脈輸入函數(Arterial Input Function, AIF)可以幫助量化血流灌注的生理參數；再則，藉由ICA非模組式的運算方式，運用到先驗知識不足的藥物濫用腦功能上，可以幫助傳統建模的困難以及提供額外的資訊；同樣地，對於龐大而且沒有剌激輔助實驗的腦預設系統，基本腦功能的進行縱使在休息的狀態也穩定地存在著，ICA搭配資料量縮減的方式能夠避免群組間及群組內的變異所造成的誤差，達到客觀分析及立竿見影的優點。 經過這一系列有系統的研究與探討，對於ICA的應用廣度從單一個體資料處理到單(多)群組的資料處理可以落實，研究的結果也從過去的研究中得到驗證。未來的工作方向可以著重在將這方向的研究推廣到腦功能連結圖譜(Brain Connectivity)，以及將影像空間的訊息加入ICA中，提昇ICA在訊號分解上的效率。
並列摘要
Independent component analysis (ICA) decomposes mixing signal into their constituent components and is commonly used in the research of signal segmentation and functional localization in magnetic resonance imaging (MRI) field. Recently, analysis on MRI data show that the interesting source signals attributing to brain activity can be regarded as independent. In perfusion study, ICA utilizes the temporal-spatial independence to segment the tissue into gray, white matter and the tissue surrounding vessel which is affected by the local field inhomogeneity during the contrast agent passage. Besides the perfusion study, ICA is also applied to locate the brain activity region by the level of blood-oxygen dependence under task delivery in function study. Determination of the data dimension in data analysis could reduce the computer loading and increase the computation speed. However, the traditional estimation methods (i.e. Akaike information criterion (AIC), Bayesian information criterion (BIC) and minimum description length (MDL)) over-estimate the dimension number due to the variation of between- and within-subject. This over-estimated situation can be decreased by a conservative method: the fitting of auto-regression model with first order, acronymic in autoregressive model of order one (AR (1)). It estimates the dimension of data by fitting the noise part of data because it assumes that the noise contribution to data is colored. In this work, ICA combined and the extended AR(1) method is applied to a series of MRI data from single dataset, single-group dataset to multi-group dataset. In the first part, ICA is used to segment the tissue around vessel in perfusion MRI study. In the second part, extended ICA is applied to single group dataset for drug abuse investigation. The resultant performance is compared with the traditional model based method. In the third part, an extended AR(1) method is used to assess the data dimension in order to handle the analysis of a giant dataset with multi-groups. In the final part, an application to the brain default system is involved to study the brain function consistency across time series. Preliminary result showed that the ICA performed excellent signal decomposition on the data. The partial volume problem for the region around the vessel is alleviated by ICA and a better arterial input function (AIF) for quantifying physiological parameters can be achieved. Subsequently, ICA utilize a non-parametric model of drug effect as compared with the traditional parametric model and it also provides extra information for the drug study such as behavior task, physiology test during scan delivery. With the same concept, ICA also helps the analysis on the brain function under resting state. The results showed that the brain default function is consistent existing across time frame and it is also exposed that ICA combined with our home-made dimension estimation method could alleviate the overestimation of dimension caused by the variation of within- and between-subject. In conclusion, ICA is a powerful tool in analyzing data. The relative research such as brain connectivity under brain resting state and the extra information involved to ICA is worth investigation and development.