Research On Methodologies Of Structural MRI-based Connectivity Extraction And Their Applications

Using complex network analysis to explore the topological characteristics and collaborative working mechanism of the brain is one of hot topics of current brain research.In recent years,magnetic resonance imaging(MRI)technology has greatly promoted the development of brain network research,and has systematically revealed the changes of brain functional and anatomical organization caused by different external stimuli or mental disorders.Structural MRI(sMRI),as a routine project for clinical examination,has the advantages of its low data acquisition cost and large possession amount,which should have large application prospects in clinical research and computer-aided diagnosis.However,due to the lack of time dimension in sMRI,it is difficult to realize connectivity extraction at the individual level,and thus the conclusion of group-level network analysis cannot be extended to individual clinical research.Therefore,developing reasonable and valid method for the construction of brain structure network at the individual level is an important issue that needs to be addressed.In this dissertation,we used machine learning approaches to investigate the brain network construction method based on sMRI.Two methods of quantifying structural connectivity were proposed and examined in the automatic diagnosis and mechanism research of brain disorders.Results indicated that our study provided reliable biomarkers for revealing the pathogenesis of brain disorders and for realizing computer-aided diagnositic purpose for the corresponding diseases.The main contributions and innovations are as follows:Frist,we developed a novel method for extracting structural connectivity of individual brain by using the multi-scale dissimilarity information between parcellations of the brain.This method quantified the structural covariation pattern among the cortical regions by combining the information of both global and local morphological differences.The structural connections achieved accuracy of 89.88% and 85.43%,respectively,in the identification of patients with Alzheimer's disease(AD)and mild cognitive impairment(MCI)from the normal control(NC)subjects.Moreover,it predicted the MCI-to-AD conversion within 3 years with 75.19% accuracy.The inclusion of cortical thickness and genetic information further enhanced the recognition performance.This study confirmed the validity of our proposed network construction method and indicated that abnormal cortical structural connections could serve as efficient biomarker for early diagnosis of AD and MCI.Second,we introduced a method of constructing sparse individual brain network based on multiple structural features.Although the aforementioned method achieved evident classification performance,it omitted the sparse interaction mode among brain regions and the “one-to-multiple” impact pattern that the structural changes of one region may simultaneously influence multiple other regions.More importantly,the feature of cortical morphology may not be comprehensively characterized by using single structural property.In view of this,we extracted structural connectivity based on multiple structural features by using the sparse regression model,and then constructed a brain network —named multi-feature based sparse network(MFN).We found that MFN can effectively characterize abnormalities in network organization caused by AD and MCI,and the accuracy of identifying AD and MCI patients from the NCs reached 96.42% and 96.37%,respectively,which almost reached the current clinical evaluation accuracy.The results were highlighted by a press release of the Society of Biological Psychiatry(SOBP).In addition,we found that MFN can effectively capture the abnormalities of cortical structures in patients with Autism Spectrum Disorder(ASD),and its diagnostic efficiency reached 78.63%.These results confirmed the validity of the proposed structural connectivity and further elucidated that cortical structural connectivity may have important application value in pathological research and computer-aided diagnosis of mental disorders.In the last part,we applied structural network construction method to investigate how painful stimulation influence functional brain network.Currently,the short duration of stimulation in single-trial pain experiments and small trial numbers limited the construction of time-series based network.If we treat the brain activity of each trial as a static picture,it should have similar data form as structural image,and thus structural network analysis could be performed.Results showed that pain stimuli significantly influenced the topological metrics of functional network and evoked an evident reorganization across the whole brain,which may enable the brain to prioritize the processing of noxious painful stimuli.Moreover,using the topological metrics of individual functional networks can effectively predict subjective ratings of pain.These findings revealed how functional network adapt dynamically,on a second-by-second basis,to pain,and may have implications for understanding what creates the conscious experience of pain.The study also laid a theoretical basis for building functional imaging-based pain prediction model.