Research On Data-driven Brain Effective Connectivity Learning Methods

The study of human connectome not only provides a new perspective for the understanding of the pathological mechanisms of neuropsychiatric disorders,but also provides new brain-network imaging markers for the early diagnosis and treatment of diseases.Previous studies showed that many neuropsychiatric disorders are closely related to abnormal topological changes in brain structural networks and brain functional networks.The brain functional networks mainly include the brain functional connectivity network and the brain effective connectivity network.Among them,the brain effective connectivity network is a graph model consisting of nodes and directed edges,where nodes represent brain regions,directed edges denote the neural activity of the causal effect from one brain region to another brain region,and the parameters associated with the edges depict the intensity of the effective connectivity.In recent years,lots of new brain effective connectivity learning methods have been proposed,which can be divided into model-driven approaches and data-driven approaches.The model-driven approaches are validation methods that are not suitable for analyzing the resting-state functional magnetic resonance imaging(f MRI)data or the case lacking prior information about the model,and can only learn small-scale brain effective connectivity networks.The data-driven methods that do not require prior knowledge and assumptions,and can learn the causality from brain regions directly from the data,have become the mainstream approaches in this field.Among them,Bayesian network method(BN)is a data-driven method with good flexibility and applicability,and it gradually becomes a hot topic in data-driven method research,because of its good performance and no requirement of the threshold.However,BN method suffers from poor search capability,noise sensitivity,and low recognition accuracy on non-stationary data and small-sample data.In view of the above shortcomings,this paper has carried out some in-depth research and exploration.The main contributions and novelties of this dissertation are as follows:First,to handle the challenge that the greedy search method for learning brain effective connectivity is prone to local optimal and has poor search ability,this paper presents two kinds of brain effective connectivity learning methods based on swarm intelligence algorithms.First of all,this paper proposes a brain effective connectivity learning method based on immune algorithm.The method uses K2 score to measure the relationships between data and the brain effective connectivity network structure,and employs four immune operators to improve the global and local search capability of the algorithm,and can efficiently and accurately search for the optimal brain effective connectivity network.In order to learn the strength of the effective connectivity at the same time,a brain effective connectivity learning method based on ant colony algorithm is proposed.This method uses the ant colony algorithm to search the brain effective connectivity network with the optimal K2 score,and obtains the connection strength of the brain effective connectivity network by calculating the pheromone concentration left by the ant colony on the optimal solution path.The experimental results verify the validity of the proposed two methods.Second,regarding the challenge of the high noise of f MRI data and the adverse effect on the learning performance of brain effective connectivity on single-modal data information,this paper studies two methods of information fusion.First of all,this paper proposes a new algorithm to learn the brain effective connectivity network structure using ant colony optimization algorithm combining with voxel activation information.By mining the voxel activation information in f MRI data,this method makes full use of the physiological characteristics of f MRI data,and can reduce the influence of data noise on algorithm performance.Secondly,this paper proposes a new method that learning brain effective connectivity networks from multi-modal data using ant colony optimization.This method utilizes the inherent correlation between structural connectivity and functional connectivity.It not only uses the knowledge of structural constraints obtained from diffusion tensor imaging(DTI)data to compress the search space,but also modifies heuristic function to enhance its enlightening ability.The experimental results validate the superiority of the proposed two methods.Third,aiming at the challenge that the existing methods fail to deal with the problem of f MRI data with small samples,this work explores for both the dynamic and static Bayesian network aspects.Firstly,this paper proposes a brain effective connectivity learning method based on non-stationary dynamic Bayesian network,which uses a new scoring function to measure the relationships between the brain effective connectivity network structure and the non-stationary f MRI data at different time intervals,and then searches for the optimal scoring network structure by the Markov chain Monte Carlo method.This method can not only effectively construct a dynamic brain effective connectivity network,but also has certain advantages in recognition performance.Secondly,a brain effective connectivity learning method based on time-series entropy score is proposed,which first uses conditional entropy to score the connection between brain regions in f MRI data,then uses the transfer entropy to capture the temporal information from the data to help the algorithm determine the direction of information transmission between brain regions,and finally uses the punishment term to prevent the network from complicating.At the same time,the paper theoretically proves the important nature of time-series entropy score function,and provides good theoretical support for its usage.The experimental results show that the two new methods have better solution quality than the existing algorithms.Fourth,to address the challenge that existing methods cannot accurately learn the brain effective connectivity on the non-stationary and small-sample f MRI data,this paper proposes a novel brain effective connectivity learning method based on generative adversarial networks.This method makes full use of the advantages of generating small sample learning against the network.In the course of the game between generator and discriminator,the distribution characteristics of the original f MRI data and the model parameters required by the structural equation model are automatically studied,so as to be able to study the brain effective connectivity accurately.In addition,by combining the recurrent neural network with the structural equation model,the method can effectively obtain the non-stationary temporal information from f MRI data,and further improves the learning performance of the algorithm.The experimental results show that compared with other methods,the new method can learn brain effective connectivity more accurately in the situation of the non-stationary and small-sample f MRI data.