Function MRI Representation Learning via Self-supervised Transformer for Automated Brain Disorder Analysis

Major depressive disorder (MDD) is a prevalent mental health disorder whose neuropathophysiology remains unclear. Resting-state functional magnetic resonance imaging (rs-fMRI) has been used to capture abnormality or dysfunction functional connectivity networks for automated MDD detection. A functional connectivity network (FCN) of each subject derived from rs-fMRI data can be modeled as a graph consisting of nodes and edges. Graph neural networks (GNNs) play an important role in learning representations of graph-structured data by gradually updating and aggregating node features for brain disorder analysis. However, using one single GNN layer focuses on local graph structure around each node and stacking multiple GNN layers usually leads to the over-smoothing problem. To this end, we propose a transformer-based functional MRI representation learning (TRL) framework to encode global spatial information of FCNs for MDD diagnosis. Experimental results on 282 MDD patients and 251 healthy control (HC) subjects demonstrate that our method outperforms several competing methods in MDD identification based on rs-fMRI data. Besides, based on our learned fully connected graphs, we can detect discriminative functional connectivities in MDD vs. HC classification, providing potential fMRI biomarkers for MDD analysis.