Research On Feature Mining And Connection Patterns Of Brain Networks Based On Magnetic Resonance Imaging

The brain is the most advanced and functional organ of human beings.It is a highly interactive and complex information processing system that can produce thinking and control consciousness.Using complex network theory to explore the topological structure characteristics and cooperative work mechanism of brain tissue is one of the current hot spots in brain science research.The popularization of magnetic resonance imaging(MRI)technology has promoted the study of brain networks at the macro-scale,and has provided a great help for capturing and understanding abnormal pathological patterns of brain diseases as well as conducting clinical auxiliary diagnosis.The structural MRI has the advantages of low acquisition cost and good retest performance.However,due to the lack of dynamic characteristics in the time dimension,it is difficult to better quantify the brain structure network at the individual level,which limits the clinical expansion of related research results.Therefore,how to use the progress information reasonably and effectively to construct the brain structure network at the individual level has become an urgent problem to be solved.The effective connectivity(EC)measures the effect of one brain region on another in a particular direction,and has gradually been widely used in the study of pathological patterns of brain diseases.However,the existing EC calculation methods are mostly suitable for the estimation of limited brain areas,rarely focusing on the whole-brain,and only explore the low-level causal interaction effects,with limited information.Therefore,capturing the causal action patterns of brain regions at a higher level from the perspective of the whole-brain is of great significance for further research on the pathogenesis and computer-aided diagnosis of brain diseases.The research work of this study mainly consists of the following two parts:Firstly,we proposed new dynamic morphological features from the perspective of longitudinal research,and constructed the structural similarity network of individual brain combined with a variety of morphological indicators.Machine learning was used to predict the conversion of mild cognitive impairment(MCI)to Alzheimer’s disease(AD).Using leave-one out cross validation and support vector machine(SVM)to train and evaluate the classifier,this method realized the 3-year conversion prediction from MCI to AD,and the accuracy was 92.31%(sensitivity was 100%,and specificity was82.86%).The results showed that dynamic morphological features are more sensitive biomarkers for predicting the conversion from MCI to AD,and the brain similarity network constructed based on them can capture more comprehensive information of abnormal cortical progression patterns.This framework will provide as much help as possible for clinical conversion prediction of MCI and also provide potential model reference for the prediction of other progressive diseases.Meanwhile,the first year is the best time node for MCI conversion prediction,which should be paid more attention in clinical practice.Secondly,we improved a large-scale brain effective connectivity calculation method,and proposed a new high-order effective network based on it.Through topological analysis to reveal the abnormal brain mechanism of major depressive disorder(MDD).The main findings of this work are as follows: MDD patients in highorder effective networks showed a general decline in connectivity values,in which the high-order effective networks were mostly long-term connections across brain regions and correlated with clinical depression scales.The effective connectivity networks measured by high-order method may be more sensitive biomarkers for MDD in clinic.Compared with normal subjects,the abnormal connections of high-order and low-order effective networks of MDD patients involved widely reported default mode network,bilateral limbic network and was also closely related to visual network and cerebellar network.Patients with MDD showed obvious effective network damage,including decreased network efficiency and clustering coefficient,the increased shortest path length,changed small world property and weakened robustness.The indirect relationship represented by the high-order effective network is more stable than the direct relationship implied by the low-order effective network.The high-order effective network will provide a potential powerful tool for the exploration of other brain diseases.