Blind source separation methods for analysis and fusion of Multimodal brain imaging data
Kyathanahally, Sreenath Pruthviraj
Type of Degreethesis
MetadataShow full item record
With the advent of functional magnetic resonance imaging (fMRI), several studies have emerged to decipher the functioning of the brain. However, three major issues limit the inferences that one could obtain from the brain using fMRI. First, fMRI time series at each voxel in the brain can be modeled as a convolution of the hemodynamic response function (HRF) and latent neuronal signals representing neural activity. Consequently, it is not a direct measure of neural activity. In the first chapter, we performed blind deconvolution of the HRF to recover latent neuronal signals to demonstrate that important resting state networks (RSNs) in the brain such as the default mode network (DMN) have a neural origin. Second, fMRI is a 4-dimensional multivariate signal which is a complex mixture of unknown independent source signals. In both second and third chapters, we used independent component analysis (ICA) based blind source separation techniques for identifying RSNs in humans and dogs, respectively. Third, fMRI has low temporal resolution which limits our ability to make inferences about fast neuronal processes. Therefore, in the fourth and fifth chapters, we performed multimodal electroencephalography (EEG)-fMRI imaging such that the superior temporal resolution of EEG can be fused with the superior spatial resolution of fMRI so as to obtain high spatio-temporal resolution. Specifically, in the third chapter, we used joint ICA, which assumes a single mixing matrix for both modalities, to investigate decision-making in the brain. In the fourth chapter, we used parallel ICA, which assumes different but correlated mixing matrices for the two modalities, to demonstrate that the neural basis of the DMN exists in sub-second neural fluctuations, making RSNs more relevant to the time-scale of neural activity. Importantly, unlike previous studies, our fusion ICA approaches do not downsample or sacrifice the native resolution of either modality.