UC Riverside

Functional magnetic resonance imaging (fMRI) measures brain activity through the blood-oxygen-level-dependent (BOLD) signal, which has been widely used to study the human brain in neuroimaging studies. Resting-state fMRI has been acquired to study functional connectivity between brain regions without any specific tasks being involved during the scan. Raw fMRI data, a time series data with three-dimensional images in stereotaxic space, provide rich information related to neural activities. Based on recurrent neural networks (RNNs), the temporal dynamics behind the brain network has been characterized. To address the functional connectivity between spatially distant regions within/across functional networks relying on regions-of-interest (ROIs) based fMRI data, we propose two deep learning architectures for fMRI analysis. First, Convolutional RNN (ConvRNN) has been introduced with convolutions for spatial feature extraction on ROI-based fMRI data. Local interactions among ROIs from the same functional network have been extracted, and temporal features have been processed by the RNN-based architecture. Better classification performance has been achieved over previous studies. The in-place visualization based on ConvRNN reveals the informative regions related to individual identification, leading to the same conclusions from previous studies. Second, we introduce a connectivity-based graph convolution network (cGCN) architecture for fMRI analysis. fMRI data are represented as the k-nearest neighbors graph based on the group functional connectivity, and spatial features are extracted from connectomic neighborhoods through Graph Convolution Networks (GCNs). We have demonstrated our cGCN architecture on two scenarios with improved classification accuracy. cGCN on the graph-represented data can be extended to fMRI data in other data representations, which provides a promising deep learning architecture for fMRI analysis.