Modeling task fMRI data via mixture of deep expert networks

Abstract

Task-based fMRI (tfMRI) has been a powerful noninvasive tool to study cognitive behaviors of the human brain. Computational modeling of tfMRI data involves two important aspects: the development of algorithms to learn a wide variety of meaningful patterns from specific task fMRI signals and the development of models to discriminate among different tasks. Although both aspects are important, most previous fMRI studies focused either on learning meaningful patterns from single tasks or learning discriminations across different tasks, and as a consequence, a unified approach to studying both aspects is rarely explored. To bridge this knowledge gap, in this study, we adopted the basic idea from convolutional neural network (CNN) and proposed a new variant of CNN called deep expert network (DEN) to model tfMRI data in a hierarchical manner. At the same time, by mixing these DEN models, we are able to achieve discriminations across different tasks. To validate the effectiveness and efficiency of the proposed mixture of DENs (MoDEN), we applied them on three Human Connectome Project (HCP) task fMRI datasets (language, social and working memory tasks). Our experiments achieved promising results and demonstrated the superior ability of MoDEN in modeling task-based fMRI data.