LAUSR

Functional magnetic resonance imaging (fMRI) is an emerging neuroimaging modality that is widely used to study brain function and disorders due to its advantages of noninvasiveness, no radiation damage, and high spatial resolution. Existing studies have focused on fMRI-based recognition models to help diagnose brain disorders. However, due to the high cost of fMRI data acquisition and labeling, the amount of fMRI data is usually small, which largely limits the performance of recognition models. In addition, cross-site classification is always a challenge in fMRI studies, because the heterogeneity of data collection at different sites increases the complexity and diversity of the data distribution, making the cross-site classification less robust than site-specific classification. In this article, we propose three data augmentation methods based on functional connectivity networks (FCNs) of fMRI data, aided by a deep feature fusion method, for automatic disease identification. First, the Gaussian noise method, Mixup method, and sliding window method are proposed to effectively augment FCN data, as these can balance the variability of the sample distribution. Second, the convolution neural network and graph attention network (GAT) are separately employed to extract local and global features from FCNs. Finally, the two kinds of features are integrated to classify subjects. The experimental results on the attention deficit and hyperactivity disorder (ADHD)-200 dataset indicate that: 1) the data augmentation methods can effectively improve identification performance, in particular, the sliding window method performs best; 2) the cross-site ADHD classification is improved by combining the data augmentation method of sliding window and deep feature fusion method; and 3) the rationality of data augmentation for FCNs is explained by visualizing the hidden fused features with t-stochastic neighborhood embedding (t-SNE) algorithm.