Spatiotemporal Attention Neural Architecture Search For Brain Functional Network Research

In the range of human cognition,as the activity state and function mechanisms of the brain have always been mysterious and difficult to explain,the research of brain science faces great challenges.Task-evoked functional magnetic resonance image(tf MRI)as one kind of functional magnetic resonance image(f MRI)can reflect the scope and mechanism of brain activity truly and reliably for the fact that it collects brain data of different functions such as emotion,memory,language and thinking based on task stimulus.By modeling the brain functional network from tf MRI,one can accurately locate the lesion area of patients and track the changes of brain activity of patients with psychosis.Therefore,it is of important practical significance to research the brain functional network generated by tf MRI data.The traditional methods were initially used in the existing research results of functional brain network.However,traditional methods have problems such as large computation and slow speed.In recent years,deep learning methods were gradually applied to the brain functional network research,and had made some progress.But as the task stimulus are dynamical and unconstraint,tf MRI data have very strong dynamics and variability,which brings great challenge in the understanding of human brain activity state by tf MRI.Therefore,this paper proposes an algorithm for searching brain functional network expression based on spatio-temporal attention neural architecture to realize the representation and analysis of brain functional network under task stimulus.The main work is as follows:(1)Due to the dynamic and unconstrained nature of task stimulus,tf MRI data at different time are dynamic and variable.In this paper,attention mechanism and Gated Recurrent Unit(GRU)are used to describe the correlation between brain network and voxel at different time for overcoming the long-distance dependence of temporal-spatial brain network expression in tf MRI data.(2)The design of common deep learning neural network brain network expression model relies heavily on expert experience,and the diversity of data sets makes the model lack generalization.In this paper,neural architecture is used to automatically search the circular structure GRU and obtain the efficient expression model of tf MRI brain functional network.(3)Experiments were carried out on seven tf MRI data from the open source data set of the Human Connected Genome Project(HCP Dataset)for brain function research.From the generated time series,brain spatial network,functional connectome and interactive connection map of brain regions,the functional brain network characteristics of different stimulus tasks and the coordination mechanism of brain regions were analyzed.Experiments show that this method can validity enhance the fitting degree of time series and task series,generate task-related brain functional network,and discover the coordination mechanism between different brain regions under different stimulus tasks.