Abstract
Autism spectrum disorder (ASD) is a neurodevelopmental condition characterized by social communication challenges and restricted, repetitive behaviors. While resting-state functional MRI (rs-fMRI) offers a window into the brain's functional organization, the inherent biological heterogeneity of ASD makes consistent identification difficult for traditional machine learning models. This paper presents the Functional Graph Discriminative Network (FGDN), a novel deep learning approach that models the brain as a structured graph. By leveraging group-level templates for both ASD and healthy controls (HC), the FGDN captures subtle connectivity patterns that distinguish the two groups. Evaluated on the multisite ABIDE dataset, the FGDN achieves state-of-the-art classification performance and identifies key neurological biomarkers, offering a robust tool for both clinical diagnostics and neuroscience research.
Introduction
The search for objective biological markers for Autism Spectrum Disorder (ASD) is a primary goal in modern neuroscience. Functional connectivity (FC) analysis, derived from resting-state fMRI (rs-fMRI), has shown significant potential in this area. However, most existing methods treat connectivity as a flat vector of correlations, ignoring the complex, non-Euclidean network topology of the human brain. Furthermore, the high variance across different imaging sites and individual patients often degrades the performance of standard classifiers like Support Vector Machines (SVMs).
Methodology: The Functional Graph Discriminative Network
To overcome these challenges, the authors developed the Functional Graph Discriminative Network (FGDN). This model treats brain regions as nodes and their functional correlations as edges, utilizing graph neural networks (GNNs) to learn hierarchical representations of brain activity.
Graph Construction and Template Anchoring
The FGDN pipeline begins by parcellating the brain into regions of interest (ROIs) using atlases such as MODL or AAL. The core innovation is the use of graph templates. The researchers calculate average connectivity matrices for the ASD and Healthy Control (HC) groups. These templates act as "anchors," providing the model with a reference for typical connectivity patterns in each group.
Dual-Graph Architecture and Spectral Convolution
For every subject, the model constructs two separate graphs: one using the ASD template and one using the HC template. Both graphs share the subject's specific node features (time-series data) but have different edge structures defined by the templates.
These dual graphs are processed through two layers of Spectral Graph Convolutions using Chebyshev polynomials. This mathematical approach allows the network to perform "convolutions" on irregular graph structures, effectively aggregating information from neighboring brain regions. The resulting features are then flattened and passed through a fully connected layer to produce a final classification.
Experimental Results
The FGDN was rigorously tested using the Autism Brain Imaging Data Exchange (ABIDE) dataset, which includes 866 subjects from multiple international sites.
State-of-the-Art Performance
The model achieved a peak accuracy of 72.5% and an Area Under the Curve (AUC) of 77.8%. In comparative tests, the FGDN consistently outperformed traditional machine learning baselines (Gaussian Naive Bayes, Ridge Regression) and other deep learning models, including LSTMs (68.5%) and standard Graph Convolutional Networks (70.4%).
Generalization and Biomarker Discovery
A key strength of the FGDN is its inter-site robustness. In "leave-one-site-out" cross-validation, the model maintained its performance, proving its ability to generalize across different scanners and imaging protocols.
Additionally, the model serves as a tool for discovery. By systematically masking specific brain regions and measuring the impact on accuracy, the researchers identified the superior temporal gyrus (R82) as the most discriminative region for ASD. This finding is highly consistent with existing neuroscience literature that links this region to social perception and language processing—core areas of impairment in ASD.
Conclusion and Future Directions
The Functional Graph Discriminative Network provides a powerful framework for identifying ASD from complex neuroimaging data. By combining the structural strengths of graph neural networks with group-level templates, the model effectively manages the biological heterogeneity of the disorder.
While the current accuracy is a significant achievement, future research will focus on making the graph templates more dynamic and incorporating phenotypic data (such as age and IQ) to further improve diagnostic precision. The FGDN stands as a promising technique for advancing our understanding of ASD and other complex brain disorders.
Summary updated