The Research Of FMRI Data Processing Approach Based On Frequency Analysis And Structural Clustering

With the development of research in diagnosis of mild cognitive impairment,which in-clude depression,Parkinson,Alzheimer’s disease etc.,data mining plays an important role in the brain functional cognitive data analysis,especially for the analysis of multi-attributes and functional cognitive pattern.Two issues are found in the process of fMRI data analysis,e.g.the potential delay in extracting brain activity region and the reliability in extracting brain functional connectivity pattern.First,while extracting the brain activity region for functional cognitive test,if there exists uncertain potential delay between the functional cog-nitive stimulus and hemodynamic response signals which is possible for the multi-subject specificity,the accuracy of extracted brain activity region might be in a low level.As for operability,resting state test becomes a new hotspot in the research of mild cognitive im-pairment diagnosis.It has been found that the functional connectivity pattern is an efficient feature for mild cognitive impairment diagnosis.But for the model of functional connectiv-ity pattern extraction,if there exists significant difference between the assumption of model and the practical data,then the reliability for extracting pattern might be decreased.With a in-depth research of the above issues,a space affine matching approach is pro-posed in this thesis for the potential delay problem.As for the reliability problem in ex-tracting brain functional connectivity pattern and large scale brain functional connectivity networks,a combinational structural clustering approach is proposed.Furthermore,in or-der to ulteriorly develop the reliability for extracted pattern,in the small scale networks temporarily,a novel approach named structural clustering analysis based on k-cliques is proposed in this thesis.The major contributions and innovation points in this thesis are as follows:Firstly,in order to solve the potential delay problem,the space affine matching approach is presented by introducing the time domain and frequency domain features.The time do-main feature is used to discern different stimuli,while the frequency domain feature is used to eliminate the delay.The algorithm is to match fMRI time series by our affine feature,in which a normal vector is estimated using gradient descent to explore the time series match-ing optimally.The experimental results for 24 subjects with receiver operating characteristic curve illustrate that our approach significantly outperforms GLM method while there exists delay between hemodynamic response signal and cognitive stimuli signal.Secondly,in extracting the brain topological networks pattern for the diagnosis of mild cognitive impairment with traditional approaches,it is assumed that brain sub-networks ex-tracted from the brain functional connectivity networks are similar with each others for sam-ple size or the region of interest size.And this might be a potential systemic error for func-tional connectivity pattern analysis.In this work,we use unsupervised approach to extract brain functional connectivity pattern.In traditional unsupervised clustering approaches,the count or proportion of sharing nodes are used as the metric for nodes density in topological networks.The robustness of functional connectivity pattern extracted by these approaches might be in a low level if the size for each sub-network is imbalance.In this case,an algo-rithm named combinational structural clustering is proposed in this thesis with the structural similarity and the loop with 3—nodes features.Thanks to the characteristic of these two features,the proposed algorithm shows robustness for the noised,pattern size imbalanced topological networks.The experimental results show the decrease of false positive outli-er compared with previous approaches while very few false negative outlier is extracted.Furthermore,the performance is also improved.In general,compared to the previous ap-proaches,the accuracy of extracted pattern increases,this approach could also be applied to the large scale networks for brain functional connectivity analysis.Thirdly,our in-depth study shows that loop with 3-nodes could be universally de-scribed as k-clique.An algorithm named structural clustering analysis based on k—cliques is proposed in this thesis.Even though the computational cost is higher than previous ap-proaches,studies in this work show that it is an accurate and robust clustering algorithm.With optimization based on graph theory and computational theory,its’ computational cost is decreased to an acceptable level for the small scale networks and current personal com-puting platform.The simulation topological networks analysis shows that the accuracy and robustness of algorithm,in structural clustering analysis based on k-cliques,are superior to traditional structural clustering approaches for ER random graph with balance pattern size,all kinds of scale free networks.The resting state functional connectivity analysis shows that the results of brain functional connectivity pattern extracted by the approach proposed in this work are verified mutually via the previous works which suggests this algorithm has great significance in medical fields.Above all,structural clustering analysis based on k-clique is an accurate,robust and well signified algorithm in medical research for functional con-nectivity pattern analysis,especially for the functional connectivity networks as complex networks.The main issues proposed in this thesis are to improve the performance of fMRI analysisapproach for the cognitive datasets.In allusion to practical issues concerned in real MCI diagnosis,the algorithm proposed in this paper will increase the accuracy and robustness of diagnosis result which may lead the further improvement of feasibility and creditability in the diagnostic procedure.In general,the approaches proposed in this thesis are with the great meanings for the MCI diagnosis and research.