Functional network connectivity (FNC)-based generative adversarial network (GAN) and its applications in classification of mental disorders

• Published in: Journal of neuroscience methods Vol. 341; p. 108756
• Main Authors: Zhao, Jianlong, Huang, Jinjie, Zhi, Dongmei, Yan, Weizheng, Ma, Xiaohong, Yang, Xiao, Li, Xianbin, Ke, Qing, Jiang, Tianzi, Calhoun, Vince D., Sui, Jing
• Format: Journal Article
• Language: English
• Published: Netherlands Elsevier B.V 15.07.2020
• Subjects: Classification; Deep learning; Depressive Disorder, Major - diagnostic imaging; Generative adversarial networks (GAN); Humans; Image Processing, Computer-Assisted; Magnetic Resonance Imaging; Major depressive disorders; Neural Networks, Computer; Resting-state fMRI; Schizophrenia; Schizophrenia - diagnostic imaging; Deep learning; Resting-state fMRI; Generative adversarial networks (GAN); Schizophrenia; Classification; Major depressive disorders
• ISSN: 0165-0270; 1872-678X; 1872-678X
• DOI: 10.1016/j.jneumeth.2020.108756

•We developed a novel GAN model using resting-state fMRI to classify mental disorders from HCs in large, multi-site datasets.•The proposed GAN model used adversarial learning and feature matching to improve the classification performance.•We tried on two kinds of disorders: 1) MDD vs HC (269 MDD, 286 HCs); 2) SZ vs HC (558 SZ, 542 HCs).•To increase result interpretability, leave-one-FNC-out looping was adopted to identify the most contributing FNCs. As a popular deep learning method, generative adversarial networks (GAN) have achieved outstanding performance in multiple classifications and segmentation tasks. However, the application of GANs to fMRI data is relatively rare. In this work, we proposed a functional network connectivity (FNC) based GAN for classifying psychotic disorders from healthy controls (HCs), in which FNC matrices were calculated by correlation of time courses derived from non-artefactual fMRI independent components (ICs). The proposed GAN model consisted of one discriminator (real FNCs) and one generator (fake FNCs), each has four fully-connected layers. The generator was trained to match the discriminator in the intermediate layers while simultaneously a new objective loss was determined for the generator to improve the whole classification performance. In a case for classifying 269 major depressive disorder (MDD) patients from 286 HCs, an average accuracy of 70.1% was achieved in 10-fold cross-validation, with at least 6% higher compared to the other 6 popular classification approaches (54.5–64.2%). In another application to discriminating 558 schizophrenia patients from 542 HCs from 7 sites, the proposed GAN model achieved 80.7% accuracy in leave-one-site-out prediction, outperforming support vector machine (SVM) and deep neural net (DNN) by 3%–6%. More importantly, we are able to identify the most contributing FNC nodes and edges with the strategy of leave-one-FNC-out recursively. To the best of our knowledge, this is the first attempt to apply the GAN model on the FNC-based classification of mental disorders. Such a framework promises wide utility and great potential in neuroimaging biomarker identification.