Graph-based Conditional Generative Adversarial Networks for Major Depressive Disorder Diagnosis with Synthetic Functional Brain Network Generation

Neurology Information Science Artificial Intelligence Medicine Psychiatry Medical Imaging
Major Depressive Disorder Generative Adversarial Networks Graph Convolutional Networks Resting-State Functional MRI Synthetic Functional Connectivity

Graph-Based Conditional Generative Adversarial Network

Graph-Based Conditional Generative Adversarial Network for Generating Synthetic Functional Brain Networks to Diagnose Major Depressive Disorder

Research Background:

Major Depressive Disorder (MDD) is a widespread mental disorder that affects millions of people’s lives and poses a significant threat to global health. Studies have shown that functional connectivity (FC) extracted from resting-state functional magnetic resonance imaging (rs-fMRI) can reveal functional connectivity patterns associated with MDD, playing a crucial role in precise diagnosis. However, the limited availability of relevant data presents a challenge for robust diagnosis of MDD. In response to this challenge, recent research has explored the use of Deep Neural Networks (DNN) architectures to construct Generative Adversarial Networks (GAN) to generate synthetic FC data, but these methods often overlook the inherent topological properties of FC.

To overcome these difficulties, the authors propose a novel method combining Graph Convolutional Networks (GCN) with a conditional GAN and a class-aware discriminator, dubbed GC-GAN. By applying GCN in both the generator and discriminator, GC-GAN captures the complex FC patterns between brain regions. The class-aware discriminator ensures the diversity and quality of the generated synthetic FC. Additionally, this paper introduces a topological optimization technique to enhance MDD diagnosis performance using an expanded FC dataset.

Source and Authors:

This paper was authored by Ji-Hye Oh, Deok-Joong Lee, Chang-Hoon Ji (IEEE student member), Dong-Hee Shin, Ji-Wung Han, Young-Han Son (IEEE graduate student member), and Tae-Eui Kam. The authors are affiliated with the Department of Artificial Intelligence at Korea University. The paper was published in the March 2024 issue of the IEEE Journal of Biomedical and Health Informatics (Vol. 28, No. 3).

Detailed Description of the Research Workflow:

a) Research Workflow:

  1. Data Acquisition and Preprocessing:

    • This study used the largest public MDD rs-fMRI dataset, REST-META-MDD, provided by the Depression Imaging Research Consortium (DIRECT).
    • Data from 25 different sites were used, focusing primarily on the largest site, Site 20, which includes data from 249 MDD subjects and 228 normal controls (NC).
    • Data collection was performed using a Siemens Tim Trio 3T scanner, followed by a series of preprocessing steps, including slice timing correction, head motion correction, band-pass filtering, and confound removal.

  2. Pretraining the GCN-Based Classifier:

    • Each rs-fMRI-derived FC is represented as an undirected graph, with nodes representing regions of interest (ROIs) and edges representing connections between nodes.
    • The Minimum Redundancy Maximum Relevance (mRMR) feature selection algorithm was used on the real FC data to define topology, capturing salient connectivity patterns between MDD and NC.
    • Spectral graph convolution with Chebyshev polynomial approximation and the evolved Laplacian matrix was used to extract informative features from the FC graph.

  3. Generating Synthetic Functional Connectivity using GC-GAN:

    • Both the generator and discriminator in GC-GAN use a GCN architecture. The discriminator classifies both the realness of the data (whether it is real FC) and the class of the data (MDD/NC).
    • The generator starts from a Gaussian random noise matrix and real class labels and generates synthetic FC based on feedback from the discriminator.
    • Cross-entropy loss and Mean Squared Error (MSE) loss were optimized to ensure that the generated FC closely resembles real FC both in recognition and numerically.

  4. Topological Optimization and Robust MDD Diagnosis:

    • The synthetic FC generated by GC-GAN was used to enhance the GCN classifier, enabling more accurate MDD diagnosis using the expanded FC dataset.
    • The mRMR algorithm was applied to feature selection on the expanded dataset to optimize topology, and a new GCN classifier was trained for robust MDD diagnosis.

b) Main Research Results:

  • Experimental Results on Synthetic Functional Connectivity Data Generated by GC-GAN:

    • Comparing various GAN architectures, such as SSGAN, WGAN-GP, and ACGAN, GC-GAN demonstrated superior performance in generating high-fidelity synthetic FC and enhancing data diversity.
    • In MDD diagnosis experiments, GC-GAN achieved a classification accuracy of 66.84%, sensitivity of 70.24%, specificity of 63.14%, and F1 score of 68.72%.

  • Classifier and Disease Diagnosis Performance:

    • Comparative experiments using various GNN models (e.g., GAT, GraphSAGE, and Ensemble) confirmed the broad applicability of the GC-GAN method across different classifiers.
    • The proposed method showed performance improvements in multiple classifiers, especially in terms of Geometric Mean Accuracy (GAA) and Balanced Accuracy between classes.

  • Model Validation on Different Datasets:

    • Experiments on different site datasets demonstrated the reliability and applicability of GC-GAN across different datasets, with domain validation on sites 20, 1, and 21.

c) Research Conclusion and Application Value:

The GC-GAN model proposed in this paper not only enhances the accuracy of MDD diagnosis but also extends the application of GANs in studying brain functional connectivity. Additionally, the introduced topological optimization technique further improves diagnostic performance through data augmentation, providing new perspectives and more efficient methods for diagnosing and treating brain diseases.

d) Research Highlights:

  • Model Innovation:
    • This is the first introduction of GCN into GAN, significantly enhancing the authenticity and diversity of synthetic data through the combination of conditional GAN and class-aware discriminator.
  • High Practicality:
    • The topological optimization technique improves diagnostic accuracy, offering a new technical approach for diagnosing psychiatric disorders based on functional connectivity.
  • Data Adaptability:
    • Extensive validation on different classifiers and datasets highlights the method’s applicability and robustness.

e) Other Valuable Information:

  • Model Scalability:
    • The proposed method is not only applicable to MDD but can also be used for diagnostic studies of other neurological diseases such as Alzheimer’s and autism.
  • Public Datasets and Open Source Code:
    • The dataset and model code used in the research are publicly available, facilitating reproduction and further research by other researchers.

By introducing GC-GAN and topological optimization techniques, this study provides an innovative and effective method for diagnosing Major Depressive Disorder using functional connectivity, demonstrating significant improvements in feature extraction and diagnostic performance, with considerable scientific value and application prospects.