Deep Neural Generative Adversarial Model based on VAE + GAN for Disorder Diagnosis

Functional magnetic resonance imaging (fMRI) plays a significant role in the study and analysis of brain cognitive function. Although the existing deep learning methods have shown good performance in fMRI disease classification, these methods have been hindered by insufficient training samples, and they also have the risk of overfitting in classification tasks. In this paper, inspired by the adversarial learning strategy, we propose a data augmentation method for brain functional connectivity (FC). The accuracy of the neural network classifier can be effectively improved by using the augmented data set. We developed a brain functional connectivity generative adversarial network (FC-GAN) for generating the brain functional connectivity patterns. We also briefly trained a FC-GAN to separately generate the control group and the experimental group of a disorder for augmenting the data set. The augmented data set was then used to train a neural network for a classification task. To test the validity of our framework, we performed experiments on three data sets, two diseases - attention deficit hyperactivity disorder (ADHD) and autism spectrum disorder (ASD) - and finally evaluated the performance by classification. Using our augmented data set, the classifier achieved an improved performance in accuracy, sensitivity, specificity, and area under receiver operating characteristic curve compared to the classification using the original data set. Compared with the currently used generation methods, our method achieved competitive performance. These findings suggest that FC-GAN approaches could augment the data set of brain functional connection in classification task.