多變量模式分析已被廣泛的用於展示在皮質活動空間模式中的相關狀態資訊。和單變量的方法相比,多變量分析方式結合了所有感興趣區域內所有體素的訊號。在這篇論文中,我們提出了一個新穎的基於深度神經網路的多變量模式分析方法。這個方法使用了卷積神經網路,而其是一個被證明在圖像分類領域非常有效的一種類神經網路,此方法不僅可以得到比支持向量機更好的分類準確度,而且也可以利用非線性建模方法識別一組具有判別能力的體素,該方法可以獲得表示整個大腦空間所攜帶活動訊息的全局功能圖。我們應用這個方法來分析功能性磁振造影的數據,結果證明了我們所提出的方法之可行性。
Multivariate pattern analysis (MVPA) has been widely used to reveal task-related information embedded in the spatial patterns of cortical activity. Compared to univariate approaches, multivariate methods combine signals from all voxels within regions of interest. In this thesis, we proposed a novel MVPA method based on deep neural networks. This method not only can obtain the higher classification accuracy than Support Vector Machine (SVM), but also can identify a group of voxels with discriminative capability. That is proposed method can obtain a global functional map which represents the information carried by the spatial pattern of activity within the whole brain. We applied proposed method to analyze a set of functional magnetic resonance imaging (fMRI) data and demonstrated the feasibility of the proposed method.