Assessing atypical brain functional connectivity development: An approach based on generative adversarial networks.

Area of Science:

• Neuroscience
• Machine Learning
• Medical Imaging

Background:

• Characterizing atypical neurodevelopment is crucial for diagnosing and predicting psychiatric disorders.
• Functional connectivity (FC) patterns from resting-state fMRI offer insights into brain function.
• Machine learning, particularly Generative Adversarial Networks (GANs), presents novel analytical capabilities.

Purpose of the Study:

• To investigate the potential of GANs combined with FC measures for predicting neurotypicality.
• To develop a predictive neurotypicality score using brain imaging data.
• To explore the utility of GANs in identifying potential biomarkers for psychiatric disorders.

Main Methods:

• Utilized ROI-to-ROI analysis of resting-state fMRI data from a cohort of children and adolescents (377 neurotypical, 126 atypical).
• Trained GAN models on neurotypical FC data to capture typical patterns.
• Employed discriminator subnetworks within GANs to differentiate neurotypical from atypical FC patterns, using ensemble methods for improved performance.
• Applied the LIME algorithm for model interpretability and identification of local hubs.

Main Results:

• GAN-based models demonstrated the ability to discriminate between neurotypical and atypical functional connectivity patterns.
• Ensemble approaches enhanced the discriminative power of the GAN discriminator models.
• The Local Interpretable Model-Agnostic (LIME) algorithm provided explanations for model predictions, highlighting local hubs.

Conclusions:

• The combination of GANs and FC measures is a promising strategy for building predictive neurotypicality scores.
• This approach holds potential for developing novel biomarkers based on functional connectivity for psychiatric disorders.
• Interpretable AI methods like LIME are valuable for understanding complex neuroimaging models.