A two-stage DBN-based method to exploring functional brain networks in naturalistic paradigm FMRI

FMRI using naturalistic paradigms such as watching movies has gained increasing interest in recent neuroimaging studies. Data-driven blind source separation (BSS) methods such as independent component analysis (ICA) are widely used to extract meaningful features in fMRI data. Recent studies have shown that BSS based on deep neural networks (DNNs) such as restricted Boltzmann machine (RBM) and deep belief network (DBN) outperform ICA. Those DNN-based methods interpret spatially aggregated fMRI time series of multiple subjects in group analysis to reduce model complexity, and only brain networks with group-wise temporal consistency can be identified. However, fMRI activities to naturalistic paradigm exhibit both group-wise consistency and inter-subject difference. To address this problem, we propose a two-stage DBN-based BSS method. In the first stage, a DBN model is trained using temporally aggregated fMRI time series of multiple subjects. In the second stage, subject-specific DBN models are initialized using model parameters learned in the first stage and are trained to converge using individual fMRI data to refine brain network identification. We use an fMRI dataset acquired using a movie stimulus to validate the proposed method.