Research On FMRI-based Multivariate Pattern Analysis And Applications

The human brain can be regarded as a complex system in which billions of neurons interact with each other dynamically.The development of brain recording technologies provides powerful tools for the exploration of brain.As a noninvasive brain imaging tool,functional magnetic resonance imaging(fMRI)is widely used in brain science research because of its relative high spatial and temporal resolution.The signal-to-noise ratio of the high dimension fMRI data is relatively low,which makes it important to choose proper fMRI analysis methods.Traditional univariate analysis methods mostly construct models on each single voxel,which overlooked the interrelationship of spatially distributed voxels.Therefore,we can hardly detect subtle differences of spatial patterns across cognitive states.Recently,multivariate pattern analysis(MVPA)derived from machine learning fields has attracted much attention in fMRI data analysis.MVPA allows combining brain activities of many voxels to discriminate different stimuli or cognitive states,which is often more sensitive than traditional univariate methods.In this dissertation,we focused on using and improving MVPA methods to explore neural representation of external stimulus or cognitive states.The main contents and contributions of the dissertation are as follows:1.Considering the structural information of fMRI data,we proposed a spatialconstrained multi-target regression model for voxel activity prediction.Most of the traditional methods constructed models for each voxel,overlooked the correlation between voxels.We utilized the spatial smoothness property of fMRI data,proposed a method to predict voxel activities of several voxels simultaneously.We used a regularization term to constrain that model weights of adjacent voxels should be similar.One the other hand,we further obtained spatial smooth activity predictions by combining the searchlight strategy.We compared the proposed method with several state-of-the-art methods on publicly available fMRI dataset,results suggest that the proposed method outperforms compared methods with regarding to encoding accuracy and decoding accuracy.2.We proposed a voxel selection method based on sparse Bayesian learning for neural decoding.Considering the correlation between voxels,we first divided all voxels into several groups according to prior structural information.And then we select discriminative voxel groups according to a group sparsity method for cognitive states decoding.Under the Bayesian framework,all parameters of the proposed model can be tuned automatically,avoiding the selection of hyper-parameters.Experimental results based on two publicly available fMRI datasets and simulated datasets demonstrate that the proposed method achieved better classification accuracies and yielded more stable solutions than several state-of-the-art methods.And the selected voxels are clustered in a few brain regions,which benefits the interpretation of results.3.We investigated cross-modal representation of individual familiarity by using cross-modal decoding methods.In the fMRI experiment,we presented familiar or unfamiliar facial images and spoken names to subjects,to evoke representation of individual familiarity in different stimulus modalities.We used region-of-interest-based and searchlight-based cross-modal MVPA methods to decode personal familiarity.We used data from one modality to training the classifier,and used data from the other modality to evaluation the performance of the classifier,so that we can identify brain regions that have similar spatial patterns across modalities.We identified several regions that support cross-modal classification of personal familiarity,including the precuneus,posterior superior middle temporal gyrus and medial prefrontal cortex,suggests the modality invariant representation of personal familiarity in these regions.Our results provide insights in the neural mechanisms of personal familiarity perception.4.We investigated how top-down processing modulates whole-brain functional connectivity(FC)by using brain network-based MVPA methods.Task-evoked FC is often overlapped with resting-state FC,which hinders the detection of subtle difference of brain network across cognitive states.We used a novel inter-subject functional correlation(ISFC)method to extract task-evoked FC when subjects are processing naturalistic stimuli according to their internal goals.Using MVPA methods,we successfully predicted the task goals of the participants with ISFC patterns in a high accuracy(90%),suggests that the task states can be represented by brain network.We further found that many ISFC measures supported task goals classification.Our findings suggest that goal-directed processing of naturalistic stimuli systematically modulates large-scale brain networks but is not limited to the local neural activity or connectivity of specific regions.