Low-dimensional Representation Of Magnetic Resonance Images Of Human Brain

Functional magnetic resonance imaging(fMRI)can non-invasively explore the functional activity of the human brain and has developed into one of the most important techniques for assessing individual cognition.However,the characteristics of fMRI that less sample-size and high dimensionality always make the studies which related with individual cognition being in trouble.And the reason why the characteristics of fMRI exist is that acquirement of fMRI is so expensive that sample size is quite small(tens to hundreds)and the four-dimensional property of fMRI(the feature size of one sample is tens or even hundreds million).Although the existing dimensionality reduction methods can mitigate curse-of-dimensionality effects to a certain extent,they can’t explore the effective information contained in fMRI in a more accurate or even complete way.Therefore,this study will investigate the methods of low-dimensional representation for fMRI.Considering that many studies related to the function of brain are always based on RS-fMRI,TS-fMRI and NS-fMRI,the study will make a preliminary exploration on the low-dimensional representation of RS-fMRI,TS-fMRI and NS-fMRI,respectively.The main contents of this paper are as follows:(1)Based on the low-dimensional representation of RS-fMRI,individual brain age is assessed.In this work,the traditional unsupervised dimensionality reduction algorithm:multidimensional scaling(MDS)is improved by constructing a new space according to the information contained in labels and the feature space is projected into the new space by linear mapping.The results show that the proposed supervised multidimensional scaling(SMDS)can effectively reduce the dimensionality of the functional connectivity extracted from RS-fMRI,and thus we can effectively assess the brain age.In addition,considering that the brain functional features are projected into low-dimensional space by a linear mapping,the features that play important roles in brain age assessment can be easily inferred based on SMDS.And the work will help researchers understand the aging mechanism of the human brain.(2)Based on the low-dimensional representation of TS-fMRI,visual objects are decoded.In order to accurately capture the features related to visual stimuli while reducing dimensionality of data as much as possible,this study used T-test to capture the voxels sensitive to visual stimuli at the feature extraction stage,and then used shared response model(SRM)to reduce dimensionality of the time series of sensitive voxels.And the low-dimensional features are used to decode the visual objects.The results show that the captured information related to visual stimuli from the time series of sensitive voxels is more accurate by SRM than the information obtained based on the original features and the features reduced by principal component analysis(PCA),locally linear embedding(LLE)and isometric mapping(ISOMAP).That is SRM can decode visual objects more accurately based on the time series of sensitive voxels.(3)Based on the low-dimensional representation of NS-fMRI,individual physiological/cognitive parameters are assessed.This work introduced convolutional self-coding neural network to perform low-dimensional representation of the signal intensity of brain regions or functional connectivity between brain regions extracted from NS-fMRI,and then verified the effect of low-dimensional representation of NS-fMRI with three tasks including individual gender classification,grip strength prediction,and individual identification.The results show that the convolutional self-coding network can effectively extract information related to gender and grip strength.However,some personalized differences are lost when the dimensionality of functional connectivity is reduced.The innovations of this study are as follows:(1)We firstly advanced a novel supervised dimensionality reduction technique for regression purposes: SMDS,and effectively reduce the dimensionality of functional connectivity extracted from RS-fMRI.(2)The low-dimensional representation of TS-fMRI is effectively achieved by using the brain activity response coefficients and shared response model,and which is good for decoding visual objects.(3)The convolutional self-coding neural network is used to reduce the dimensionality of the intensity of signal changes in brain regions and functional connectivity between brain regions extracted from NS-fMRI,and that will help researcher assess individual physiological/cognitive parameters based on fMRI.