ICA Of Complex FMRI Signal Based On Kurtosis Maximization Algorithm

Functional magnetic resonance imaging (fMRI) data are brain signals obtained by the MRI scanner when a specific task is performed by a subject. Because of its advantages of higher resolution on time and space, being able to be tested repeatedly in vivo and having no radiation damage, fMRI has been widely used in brain research and clinical diagnosis. The original fMRI data are acquired as complex image pairs including magnitude and phase information. For simplicity, most current researches only utilize the amplitude information, but ignoring the phase information which is different from the amplitude, resulting in unnecessary loss of performance. With the development of signal processing technology, especially that of ICA (independent component analysis) in the separation of complex signals, analysis of complex fMRI is attracting more and more attention. In this thesis, we carried out researches aiming to solve current shortcomings in ICA for complex fMRI data.In this thesis,our work is as follows:(1) We described the temporal and spatial ICA estimation models, studied two semi-blind methods'principle based on kurtosis maximization algorithm, namely fixed-point algorithm KM-F-R and the gradient algorithm KM-G-R, we described the complex fMRI data acquisition process and analyzed current research topics and problems.(2) The non-circularity of complex fMRI data was studied, we provided a quantitative measurement index for non-circularity degree for complex-valued signals named DOI, calculated the DOIs of simulated and real fMRI signals. Experimental results showed that the estimates of complex fMRI signal have a non-circular nature, which can provide guidance for development of new algorithms.(3) We calculated the quantitative index ISI for the separation signals by blind and semi-blind ICA algorithms. The number of voxels in the brain activation map was significantly increased for the semi-blind algorithm based on kurtosis maximization, but phase information introduced noises into the time courses estimation, which need preprocessing for the phase information.(4) We analyzed the impact of the dimension on ICA performance for decomposing fMRI data. ICA separation experiments were performed after different dimension compression under the same conditions. The separation results showed that when the number of dimensions is too small, the performance decreased due to information loss, and when the number of dimensions is too large, the performance also decreased due to information redundancy.