Diagnosis and Pathogenic Analysis of Autism Spectrum Disorder Using Fused Brain Connection Graph

TL;DR

This work proposes a model for diagnosing Autism spectrum disorder (ASD) using multimodal magnetic resonance imaging (MRI) data, employing graph neural networks (GNNs) for fused graph classification and introduces a loss function that maximizes inter-class and minimizes intra-class margins.

Abstract

We propose a model for diagnosing Autism spectrum disorder (ASD) using multimodal magnetic resonance imaging (MRI) data. Our approach integrates brain connectivity data from diffusion tensor imaging (DTI) and functional MRI (fMRI), employing graph neural networks (GNNs) for fused graph classification. To improve diagnostic accuracy, we introduce a loss function that maximizes inter-class and minimizes intra-class margins. We also analyze network node centrality, calculating degree, subgraph, and eigenvector centralities on a bimodal fused brain graph to identify pathological regions linked to ASD. Two non-parametric tests assess the statistical significance of these centralities between ASD patients and healthy controls. Our results reveal consistency between the tests, yet the identified regions differ significantly across centralities, suggesting distinct physiological interpretations. These findings enhance our understanding of ASD's neurobiological basis and offer new directions for clinical diagnosis.

Figures (5)

• Figure 1: (a)--(b) Visualizations of the brain region adjacency matrices $\mathbf{A}^{\rm DTI}$ and $\mathbf{A}^{\rm fMRI}$ obtained from DTI (a) and fMRI (b) modalities, respectively. In panel (a), the nodes represent the 90 brain regions, while in panel (b), the columns/rows correspond to the same set of brain regions. The workflow of the proposed model with three modules panel (c)--(e).
• Figure 2: (a) Visualizations of ASD classification on TNR (left panel), TPR (middle panel), and AUC (right panel) (with standard deviation) of baseline methods and the proposed models, averaged over five cross-validation folds. "(+ $\mathcal{R}_{g}$)" denotes the proposed regularization-based GNN model. (b) Comparisons on two fusion manners, ${\rm GNN}(\mathbf{A}=\mathbf{A}^{\rm DTI},\mathbf{X}=\mathbf{A}^{\rm fMRI})$ and ${\rm GNN}(\mathbf{A}=\mathbf{A}^{\rm fMRI},\mathbf{X}=\mathbf{A}^{\rm DTI})$, namely, exchange the roles of $\mathbf{A}=\mathbf{A}^{\rm DTI},\mathbf{X}=\mathbf{A}^{\rm fMRI}$ in GNNs, where ${\rm GNN}$ can be selected from {GCN, GAT, GraphSAGE, ChebyNet}. (c) ROC curves of the four baselines (blue) and the corresponding proposed methods (green). (d) Searching for the penalty tuning parameter $\alpha$ in Eq. (\ref{['eq:overallLoss']}) of the four proposed models.
• Figure 3: The visualization depicts three distinct centralities of the Karate Club graph: degree centrality (left panel), eigenvector centrality (middle panel), and subgraph centrality (right panel). The color intensity and node size correspond to normalized centrality values, with larger and darker nodes indicating higher centrality levels.
• Figure 4: The visualization portrays the distributions of the three types of network node centrality within the ASD group (comprising 67 patients denoted by red curves) and the control group (consisting of 71 samples denoted by blue curves). The top 15 distributions of brain regions are displayed based on the ranking of network node centrality. The panels, from top to bottom, correspond to degree centrality, eigenvector centrality, and subgraph centrality, respectively.
• Figure 5: Visualization of ASD-related brain regions and functional connectivity related to ASD, analyzed through two statistical methods: (a) MWU test (rows 1 and 2) and (b) MMD test (rows 3 and 4). Three centrality measures are displayed for each testing method: degree centrality (column 1), eigenvector centrality (column 2), and subgraph centrality (column 3). Rows 1 and 2 show brain regions and functional connectivity for both left and right hemispheres of the brain. Each sub-figure, for example, the figure of row 1 and column 1, shows significant regions of the left and right halves of the brain from lateral and dorsal views, and the medial view visualizes the top 15 ASD-related regions. The brain is divided into 7 regions, each represented by a different color in the subgraphs: Central Region, Frontal Lobe, Temporal Lobe, Parietal Lobe, Occipital Lobe, Limbic Lobe, and Sub Cortical Gray Nuclei. The overlapping areas between different centrality measures are indicated in the rightmost column while different test methods are indicated at the bottom row in light blue. The dark blue area highlights the final intersection of regions that consistently overlap across all centrality measures and statistical tests.