Research On Brain Architecture Via Functional MRI

Modern magnetic resonance technology(MRI)enables looking into the human brain in a non-invasive way.With MRI,not only it's possible to inspect the anatomy and structure of the human brain,but also and more excitingly,it's possible to measure and record the brain functional activity.This shed a light,a routine and a clue to decipher the mystery of brain architechtures of structure and function.With the development of neuroscience,cognitive science and psychology,more complex models are involved into brain function research,among which the most popular one is the network hypothesis.There are two streams in the brain network research.One is the “node-edge” model which is in a large scale,the other one is the voxel level “component analysis” stream.The key problem of “node-edge” model is the definition of nodes and the crucial issure of voxel level “component analysis” model is the definition of relationship among components.In this thesis,we adopt DICCCOL as the nodes of the “node-edge” model,and we use sparsity,which reflect the intrinsic nature of neural activity,as the principle of component analysis.Recently DICCCOL(Dense Individualized and Common Connectivity-based Cortical Landmarks)system has attracted a lot of attention in the neuroimaging field as a system for structural brain network,which also naturally serves as a system of common brain nodes and a platform of statistical analysis.We will discuss the important functional architecture research based on DICCCOL system.DICCCOL system performs well for “node-edge” model,however,DICCCOL itself is a set of sampling points on the brain surface,and it could only support network analysis in a large scale.In order to sufficiently decode the functional architecture of the brain,methodology that could support big data mining is needed for voxel level brain fMRI signal analysis.Recently,sparse coding is quite active in the field of computer vision,but it's rarely applied on medical image analysis,like fMRI.The other part of this thesis will focus on the sparse coding for brain network representation.Briefly,the thesis can be summarized in the following aspects:1.Propose new methods about network activation,from nodes to connections.? Group-wise activation detection based on DICCCOL.For a long time,group-wise activation detection is based on the alignment reached by registration methods,but registration methods cannot deal with the dramatic individual variability.DICCCOL system addressed this challenge from the feature space of fiber connectivity.So that based on the correspondence built up by DICCCOL,we developed a new group-wise network nodes activation detection method.This method significantly improves the sensitivity of activation detection.? Group-wise connection activation detection.Modern neuroscience believes that brain function is not easily realized by regional activation,but is based on complex interaction and cooperation of regions at the network level.Thus,how to define the network level activation is another question to answer.The DICCCOL naturally serves as a node system for our brain network research.Based on DICCCOL,we explore the dynamic interaction among the brain nodes,and the activation detection on the dynamic interaction help us define the connection level activation.2.Propose new methods to model complex brain network architecture.? Multi-stage information system in cognitive process.It's believed that there exists top-down and bottom-up information flow in the brain cognition.Information is transmitted and interacted by different brain regions.However,this has rarely been modeled within the neuroimaging field.Also based on the correspondence realized by DICCCOL system,we proposed a new model to describe the hierarchical information flow of human brain cognition.? Research about network overly based on DICCCOL.Sparsity is the intrinsic nature of neural activity.Diverse sparse combination of neuron activation could realize diverse brain function,i.e.,a certain neuron's interaction with different neurons could conduct different brain function.Thus,whether this sparsity could be captures in the fMRI data is a question to answer.Sparse coding is applied on group-wise DICCCOL fMRI signals,meaningful brain networks with overlay patterns are detected.3.Sparse representation of whole brain networks.? Functional network representation based on sparse coding.Based on sparse coding on DICCCOL fMRI signals,meaningful networks could be detected.So that we apply dictionary learning and sparse coding on whole brain fMRI signals to find networks with higher resolution and complexity.Our results indicate that sparse coding could find meaningful brain networks.? Holistic atlas of functional networks and their interaction(HAFNI).Our research found fMRI signals can be decomposed into hundreds of networks with correspondence among subjects.We summarize from these networks to build a whole brain atlas of functional networks.Overlay patterns widely exist among these networks.4.Method extension and clinical application of sparse representation of brain networks.? Supervised sparse coding method for f MRI analysis.Traditional model driven method usually overview the huge information hidden in the fMRI data,but pure data-driven dictionary learning and sparse coding is kind of arbitrary in brain network analysis.We propose a supervised sparse coding method for fMRI analysis.Temporal and spatial information could be supervised for the learning of certain networks,on the other hand,other networks could be learned in an automatic way.? Group-wise sparse coding on clinical data.Sparse coding could detect brain networks with correspondence among subjects,but how to benefit clinical research is a question for us to answer.We developed a group-wise sparse coding method and applied it on a data set of PAE(Prenatal Alcohol Exposure)subjects.Our results suggest that groupwise sparse coding could find significant difference between normal group and disease group from the perspective of functional network.