| dc.description.abstract | The human brain is the most complex organ containing billions of neurons that communicate with each other to share information systematically. Learning its behavior through functional magnetic resonance imaging (fMRI) led us to investigate the information flow through the blood oxygen level. This work aims to introduce a method to process the fMRI in a temporal sliding-window configuration, preserving the content and connectivity from one window to another and analyzing the signals' behavior in different visualizations. First, we introduced the modified Principal Component Analysis for sliding-window (mPCASW) to address the traditional Principal Component Analysis (PCA) limitation that cannot be applied directly to sliding-window data while still adapting the PCA characteristics of uncorrelatedness. This method incorporated the sliding-window configuration into the algorithm while trying to achieve the orthogonality in its basis vectors. The advantages of this method are that it is robust to any size of the window and the number of overlapping timepoints, N_hop, and its' basis vectors tend to maintain close to orthogonality even though the basis vectors are divided into different time windows. Our original mPCASW method restricted the number of basis vectors to be the N_hop+1, much less than the traditional PCA or other decomposition methods; therefore, we expanded it to allow more components to be extracted. Using the resulting mPCASW components' basis vectors, we introduced several ways to visualize the decomposed fMRI signal probabilistically, especially during the brain avalanches when most of the brain voxels have high positive activations. The 'standard deviation volume' (SDV) computed from the transform coefficients - the projection of the mPCASW basis vectors into the dataset showed a synchronization pattern across the whole-brain voxels around brain avalanches. Clustering the components' coefficients also highlights the components that contain significant changes. These behaviors appear in large and tiny brain avalanches, indicating that small activations should not be ignored as they could provide information on the fMRI signal content flow. Characterizing deeper into the transform coefficients of each component around large avalanches, we found five major distinct patterns that can be used to classify the characteristic of a specific mPCASW component. Sequential patterns across components' extremal curves of the functional network (FCN) show that several brain network regions have clear time delays as avalanches reach the maximum point. The other components were found to diverge, converge, or both across the regions around the avalanches, reflecting possible synchronization of information happening during the avalanches. Several network regions, mainly the somatomotor (SMN) and dorsal attention (DAN) networks, exhibit the most evident pattern changes during signal transitions. These crucial findings using our mPCASW and visualization methods show that BOLD signals could reveal more interpretations of the dynamics in the human brain and the possible synchronization and desynchronization happening during fMRI avalanches. | |
