Deep Learning Based Feature Extraction Methods For FMRI Data And Brain Functional Connectivity Network

Extracting effective features from functional magnetic resonance imaging(f MRI)data and the brain functional connectivity network can provide a new perspective for in-depth understanding of human brain operation mechanism and pathological mechanism of brain diseases,which is an important topic in the field of brain science.The deep learning has become a hot topic in the research of feature extraction from f MRI data and brain functional connectivity network due to its powerful ability of deep network learning.However,the existing methods ignore the high-level temporal characteristics of f MRI data and the sparsity,modularity and the small characteristics path length of the brain network,which makes it hard to obtain the discriminative and effective features.In view of the above shortcomings,this dissertation makes an exploration and research,of which the main contributions and novelties are as follows:1.In order to take the high-level temporal characteristic in the estimation of hemodynamic states,this dissertation presents an approach to extract the f MRI temporal features based on a stacked recurrent neural network model.Firstly,the inversion process has been constructed by the inverse of nonlinear functions in the hemodynamic model,which mapping the f MRI signal to hemodynamic states.Then,three recurrent neural network(RNN)modules are stacked to form a stacked recurrent neural network,each of RNN structure can extract the time series characteristics of the corresponding hemodynamic state.Finally,the f MRI temporal features are used to estimate hemodynamic states.The experimental results on a simulation dataset and a real-world dataset show that this method can extract the temporal characteristics of f MRI data more reasonably,then the constructed functional connectivity based on the estimated hemodynamic states leads to a better performance on the classification of brain functional connectivity.2.To take the sparse connectivity patterns(SCPs)of the human brain into consideration,this dissertation proposes a novel CNN(Convolutional neural network)based model with graphical Lasso(CNNGLasso)to extract sparse topological features for brain disease classification.Firstly,we develop a novel graphical Lasso model for revealing the SCPs at group-level.Then,the SCPs are used to guide the topological feature extraction.Finally,the obtained sparse topological features are used to classify the patients from normal controls.The experiment results on the ABIDE dataset demonstrate that the CNNGLasso outperforms the others on various performances.Besides,the abnormal brain regions derived from the trained model are consistent with the previous investigations,which further proves the application prospect of the CNNGLasso.3.To make full use of the associations between the abnormal connectivity patterns within modules in a brain network and brain diseases,this dissertation proposes a CNN model to extract the multi-level modular features(CNN-MM).More specifically,we first develop a novel algorithm to obtain the modular structure of each node,which are then fed into a CNN model to extract the node-level modular features.Secondly,we minimize the harmonic modularity of the extracted node-level features to reveal the modular structure at whole brain network level.Finally,we employ a deep neural network to further extract high-level features for the classification of the brain disease.The experiment results on a real-world autism spectrum disorder dataset show that our proposed method outperforms other prevalent methods.In addition,the feature analysis based on the trained framework reveals the associations between the modular structures and the brain disease,which provides new insights into the pathological mechanism from the perspective of modular structure.4.In view of the existing deep learning-based brain functional connectivity classification methods ignored the property of small characteristic path length in the brain network,this dissertation proposes a deep neural network model based on random walk and RNN model.Firstly,the random walk algorithm is used to sample on the brain network to obtain node sequences and the corresponding features.Since the sampling process of traditional random walk algorithm is not derivable,we introduce Gumbel softmax to construct a learnable random walk process.In addition,the length loss term is introduced into the optimization target to constrain the length of these random paths,so that it can meet the short path characteristics in the brain network.Then,the RNN model is used to extract the sequence characteristics of these features.Finally,the deep neural network structure is used to extract the high-level features of the neural network for the diagnosis of brain diseases.The experimental results on the real f MRI data show that the method can extract the spatial sequence features of brain network effectively to improve the performance of classification of brain function connectivity.In summary,the deep learning model is used in this paper to extract the discriminant features of f MRI data and the brain functional connectivity.The experiment results show that these methods significantly improve their performance in the estimation of hemodynamic states and the diagnosis of brain disease,which demonstrates the effectiveness of the extracted features based on the proposed methods.