Research On Identification Of Brain Response Patterns Based On Functional Magnetic Resonance Imaging

Exploring the organizational architecture of human brain has been a tense interest and great challenge in neuroscience community.For decades,task-based functional magnetic resonance imaging(tfMRI)has been a powerful and popular noninvasive neuroimaging technology for the study of brain activities.To model the very informative but complex tfMRI time series data,a variety of analysis approaches have been developed and greatly advanced our understanding of the brain networks and associated hemodynamic patterns.Although prior work has proven great advantages,a possible drawback is that the methods are usually based on prior hypothesized models such as the hemodynamic response function or other mathematical or statistical assumptions.Besides,there is always a large amount of data for brain images,and it relies on methods like region of interest(ROI)or down sampling to reduce the data size.In response to the limitation of existing methods,this thesis introduces the study on natural brain response activities with deep learning models and machine learning methods based on tfMRI data,in which diverse and complex spatial and temporal patterns of brain response activities could be identified.In response to the hypothesize dependency in modeling brain tfMRI data,a Deep Recurrent Neural Network(DRNN)model is proposed.The task stimulus information is sequentially processed through the model,automatically generates the whole brain voxel signals and models the natural brain response activities.The proposed DRNN framework can efficiently model the natural brain response activities without hypothesizes,and not only identify well-shaped functional brain networks,but also more response activity patterns at multiple time scales simultaneously.In response to the limitation of identifying and interpreting complex brain networks,a supervised dictionary learning and sparse coding method is proposed.With the data-driven regressors derived from the DRNN model,the supervised dictionary learning approach fixes the data-driven regressors as predefined dictionary atoms,only optimizes the other portion of dictionary atoms,then brain networks can be obtained from the sparse coefficient matrix.The proposed framework can identify not only individual and well-shaped functional brain networks oriented from task stimulus information,but also latent and concurrent resting state networks.In response to the limitation of estimating diverse and complex hemodynamic response patterns,a unsupervised framework of Deep Recurrent Autoencoder(DRAE)mode is proposed.The DRAE model combines the deep recurrent neural network(DRNN)and autoencoder to model brain response activities without any prior knowledge,estimates the optimal temporal patterns automatically,and characterizes diverse corresponding hemodynamic response patterns in tfMRI data simultaneously.