Analysis of brain fMRI time-series using HRF knowledge-based correlation classifier on unsupervised self-organizing neural network map

Abstract

Brain imaging and particular functional MRI (fMRI), which acquires brain volumes in time, reveals new understanding of the functional/structural relation in neuroscience. During fMRI imaging physiological state changes occur in the brain regions activated from the task paradigm which the subject performs in the scanner. These state changes can be depicted in the small veins of the activated region due to the blood oxygen level dependent (BOLD) effect. For each brain voxel in the fMRI experiment one accumulates a time series vector which has to be analyzed for similarity to the original task paradigm vector and its characteristic hemodynamic response function (HRF). Various analysis methods have been discussed for fMRI analysis, model-based statistical or unsupervised data-driven techniques. The purpose of this paper is to introduce a new method which combines two different approaches. We use an unsupervised self-organizing map (SOM) neural network to reduce the time series vector space by non-linear pattern recognition into a 2D table of representative time series wave-forms. Using a-priori knowledge of the HRF, either derived from a theoretical wave-form model or estimated from a brain region of interest (ROI), one can use correlation analysis between the time series patterns of the SOM table and the HRF to depict regions of activation specific to the HRF. An optional second SOM training with a reduce number of neurons of the best-matching time series to the HRF classification refines the second neural network pattern table. The learned time series pattern of each neuron and the corresponding brain voxels are superimposed onto the subject's brain image for visual investigation.