Modeling And Application Of Task-state FMRI Oriented Deep Learning

With the development of neuroimaging technology,functional magnetic resonance imaging(fMRI)has become one of the hot research topics,in which task-state fMRI can reflect the response of the brain’s activities in corresponding regions of the brain under conditions of stimulation.Using data modeling method to analyze fMRI data can help us better understand how the brain works.In recent years,deep learning has become one of the commonly used methods because of its outstanding performance in modeling multidimensional data.However,the current deep modeling methods have poor performance because they cannot make full use of the temporal and spatial features of fMRI data.In order to extract the spatial and time series features of fMRI data at the same time,a convolutional recurrent network model(CRNN)is designed in this thesis.In the model,the convolution network module reduces the dimension of data and extracts the spatial feature information,while the recurrent network module extracts the temporal information of data and adds attention mechanism to make the model pay more attention to the information of brain activation state.The model can make full use of the information of data and carry out experiments on public data sets to achieve94.31% classification result,which is better than other comparative models.In order to better analyze the responses of brain functional areas under different task states,based on the model proposed in this thesis and the convolution visualization method,the features learned from the corresponding data are extracted from the model and mapped to the corresponding brain functional areas based on the method of gradient weighted class activation mapping for visualization.Finally,transfer the trained model parameters to the new model for initialization,and carry out classification experiments on small sample data sets.Compared with the normal training method,it not only slows down the over-fitting phenomenon,but also improves the classification performance by 8.34%.