Research On Classification Method Of Brain Functional Connectivity Data Based On Neural Networks

Resting-state functional magnetic resonance imaging(rs-fMRI)is a type of functional magnetic resonance imaging(fMRI),which is mainly used to assess the interaction of brain functional areas when subject is in resting state or not performing a clear task.Analysis of rsfMRI data can help to explore the functional organization of the brain and study its changes in neurological or psychiatric diseases.Existing studies have found that changes of brain functional connectivity patterns are closely related to the clinical manifestations of some mental diseases,such as schizophrenia,Alzheimer's disease,autism and so on.It's helpful for the understanding of brain diseases and the discovery of potential biomarkers that conducting research on brain functional connectivity data based on machine learning method and analyzing the correlation between the functional connectivity of brain regions and brain diseases.In this paper,we study the classification of brain functional connectivity data based on neural networks.The deep neural networks have achieved excellent results in classification tasks and have been applied to the analysis of functional connectivity.However,it's difficult to explain the internal logic of neural networks when using them to make predictions.Regarding the issue above,taking advantage of the interpretability of piecewise linear neural network(PLNN),we propose a fully connected neural network(FCNN)based on PLNN,which is used to classify the brain functional connectivity data.The functional connectivity vector is input into the FCNN model for training to realize the classification of rs-fMRI data.Then,based on the interpretability of the PLNN,an interpretation method for the FCNN model is proposed to explain the feature contribution of the classification judgment of the model on a single sample;and statistically explain the feature contribution of the sample group in the model classification judgment.We conduct classification experiments on autistic patients and healthy controls on the ABIDE I dataset to verify the FCNN model.The experimental results show that the FCNN model has good classification performance.We also analyze the contribution of functional connectivity features in the classification model.Aiming at the characteristics of connection and cooperation between the brain regions of the brain,this paper proposes a graph neural network model(GAT-NR)based on graph attention network(GAT)to classify brain functional connectivity data.First,the brain regions are used as the nodes of the graph,and the value of functional connectivity between the brain regions are used as the weights of the edges between the nodes,so that the rs-fMRI data is constructed as graph data and input into the graph neural network model for training and classification.For the GAT-NR,the graph attention layer is used to output the feature representation of the node,and then the node representation is mapped by function to obtain the node information and the information of each node is weighted,and finally the prediction result is output.Classification experiments on autistic patients and healthy controls were conducted on the ABIDE I dataset,and the results show that the model has advantages in classification effects.In addition,the saliency map is used to calculate the gradient to analyze the feature contribution in the model classification judgment.