Analysis Of Spatial Source Phase From Complex-Valued FMRI Data And Classification Of ICA-CNN

Spatial source phase is the phase information of spatial maps(SM)extracted by datadriven blind source separation algorithms from complex-valued functional magnetic resonance imaging(fMRI)data,which has superior denoising performance.However,the uninque information contained in spatial source phase and its ability on identification of group difference and classification of patients and healthy controls have not been studied.As such,this thesis carries out the following novel studies based on resting-state fMRI(rs-fMRI)from 82 subjects including schizophrenia(SZ)and healthy controls(HC):(1)The thesis proposes a group difference identification method based on variance analysis and a method for verifying differences based on resamplings.First,construct voxellevel spatial source phase vectors for SZ and HC,respectively,and obtain SZ-HC variance difference maps based on homogeneity of variance test(F-test).Then define variance difference index to measure and highlight the spatial difference in region of interests.Robustness of the difference is then verified by two resampling methods.Results of default mode network(DMN)and auditory cortex(AUD)show that the spatial source phase has higher sensitivity and robustness in identifying spatial differences compared with the widely used spatial source magnitude.(2)The thesis proposes a single-component ICA-CNN classification framework and data augmentation based on mulitiple kinds of model orders of ICA.The SM estimates of the components of interest were obtained by ICA,and then converted into 2D slices to construct samples for CNN.ICA extracts single-component primary features from fMRI data,and CNN models extract deep features from ICA components for classification.Results of DMN,left frontoparietal network(FPNL)and right frontoparietal network(FPNR)indicate that the proposed ICA-CNN has higher classification performance compared with single-ICA framework without CNN,the accuracy,sensitivity and specifity for classification based on spatial source phase is better than spatial source magnitude and magnitude data,and data augmentation based on multiple model orders of ICA solves the shortage of fMRI data to some extent.(3)The thesis proposes a multiple-components ICA-CNN classification framework based on post-ICA fusion and post-CNN fusion.Post-ICA fusion combines different resting-state networks(RSNs)from the data level,and post-CNN fusions connect different RSNs at the feature level and single-model decision-making results,repectively.Then,comprehensive prediction is carried out.Results of different numbers of RSNs indicate that the fusion of more RSNs is beneficial to improve the classification performance.Moreover,post-CNN feature fusion has the strongest feature learning ability,post-ICA post fusion and post-CNN feature fusion are better RSN fusion strategies compared with post-CNN voting decision.