A Comparative Study: Enhancing Conditional Generative Adversarial Networks for Functional Connectivity Synthesis in Major Depressive Disorder

Ji-Hye Oh, Eunjung Jo, Tae-Eui Kam

Abstract: Major Depressive Disorder (MDD) is a prevalent mental health condition, affecting a significant number of individuals globally and representing a critical health challenge. The use of functional connectivity (FC), obtained from resting-state functional Magnetic Resonance Imaging (rs-fMRI), is vital in identifying patterns linked to MDD, thereby aiding in its precise diagnosis. Nevertheless, the scarcity of FC poses a significant hurdle in the effective diagnosis of MDD. To overcome this issue, various studies have utilized Conditional Generative Adversarial Networks (cGAN) to create synthetic FC. This synthetic FC is then used as additional training data for MDD diagnosis. However, most previous research has primarily focused on using cGAN validated in fields such as natural image processing, which may not be sufficient for generating realistic synthetic FC given the limited quantity and complex patterns of FC. Therefore, we introduce the utilization of three existing techniques, i.e., class-wise scaling loss, pre-trained autoencoder, and label embedding projection, aimed at enhancing cGAN performance, enabling cGAN to generate synthetic FC with more accurate and improved representations. We assessed the methods on the rs-fMRI dataset available to the public and the results indicate that employing these methods with the cGAN provides significant assistance in synthetic FC generation.