Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

NeuroTree: Hierarchical Functional Brain Pathway Decoding for Mental Health Disorders

Jun-En Ding, Dongsheng Luo, Anna Zilverstand, Kaustubh Kulkarni, Feng Liu

TL;DR

NeuroTree addresses the need for interpretable, hierarchically organized brain-network representations in mental health by deconstructing dynamic fMRI connectivity into a trainable, age-aware tree of functional pathways. It combines a $k$-hop AGE-GCN with neural ODEs to model time-evolving brain states, and employs CMFC to sharpen region-specific FC patterns, while constructing a hierarchical brain tree that highlights high-order pathways. The approach achieves state-of-the-art performance on cannabis and schizophrenia datasets and provides insights into age-dependent brain network deterioration, linking network disruptions to known large-scale networks. By delivering both accurate predictions and anatomically meaningful explanations, NeuroTree advances the interpretability and clinical relevance of graph-based brain analyses for psychiatric disorders.

Abstract

Mental disorders are among the most widespread diseases globally. Analyzing functional brain networks through functional magnetic resonance imaging (fMRI) is crucial for understanding mental disorder behaviors. Although existing fMRI-based graph neural networks (GNNs) have demonstrated significant potential in brain network feature extraction, they often fail to characterize complex relationships between brain regions and demographic information in mental disorders. To overcome these limitations, we propose a learnable NeuroTree framework that integrates a k-hop AGE-GCN with neural ordinary differential equations (ODEs) and contrastive masked functional connectivity (CMFC) to enhance similarities and dissimilarities of brain region distance. Furthermore, NeuroTree effectively decodes fMRI network features into tree structures, which improves the capture of high-order brain regional pathway features and enables the identification of hierarchical neural behavioral patterns essential for understanding disease-related brain subnetworks. Our empirical evaluations demonstrate that NeuroTree achieves state-of-the-art performance across two distinct mental disorder datasets. It provides valuable insights into age-related deterioration patterns, elucidating their underlying neural mechanisms.

NeuroTree: Hierarchical Functional Brain Pathway Decoding for Mental Health Disorders

TL;DR

NeuroTree addresses the need for interpretable, hierarchically organized brain-network representations in mental health by deconstructing dynamic fMRI connectivity into a trainable, age-aware tree of functional pathways. It combines a -hop AGE-GCN with neural ODEs to model time-evolving brain states, and employs CMFC to sharpen region-specific FC patterns, while constructing a hierarchical brain tree that highlights high-order pathways. The approach achieves state-of-the-art performance on cannabis and schizophrenia datasets and provides insights into age-dependent brain network deterioration, linking network disruptions to known large-scale networks. By delivering both accurate predictions and anatomically meaningful explanations, NeuroTree advances the interpretability and clinical relevance of graph-based brain analyses for psychiatric disorders.

Abstract

Mental disorders are among the most widespread diseases globally. Analyzing functional brain networks through functional magnetic resonance imaging (fMRI) is crucial for understanding mental disorder behaviors. Although existing fMRI-based graph neural networks (GNNs) have demonstrated significant potential in brain network feature extraction, they often fail to characterize complex relationships between brain regions and demographic information in mental disorders. To overcome these limitations, we propose a learnable NeuroTree framework that integrates a k-hop AGE-GCN with neural ordinary differential equations (ODEs) and contrastive masked functional connectivity (CMFC) to enhance similarities and dissimilarities of brain region distance. Furthermore, NeuroTree effectively decodes fMRI network features into tree structures, which improves the capture of high-order brain regional pathway features and enables the identification of hierarchical neural behavioral patterns essential for understanding disease-related brain subnetworks. Our empirical evaluations demonstrate that NeuroTree achieves state-of-the-art performance across two distinct mental disorder datasets. It provides valuable insights into age-related deterioration patterns, elucidating their underlying neural mechanisms.

Paper Structure

This paper contains 46 sections, 6 theorems, 44 equations, 11 figures, 6 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Neural ODEs for Dynamic BOLD fMRI Signal Modeling.
  4. Path-Aware Learning in GNNs.
  5. The Proposed Architecture
  6. Preliminaries
  7. Graph Convolutional Networks
  8. $K$-hop ODE-Based Graph Embedding
  9. Optimization of Objective Function
  10. Hierarchical Brain Tree Construction
  11. Optimal Weighted Tree Path
  12. High-Order Tree Path Aggregation
  13. Experiments
  14. Datasets
  15. Qualitative Results
  16. ...and 31 more sections

Key Result

Theorem 4.2

Suppose the $l^2$-norm of the $k$-hop connectivity adjacency operator $\|\hat{A}_k(t)\|_{2}$ can be derived from Eq. eq:equation7. Then, as the $k$-hop approaches infinity ($k \to \infty$ ), the $\|\hat{A}_k(t)\|_{2}$ is bounded by the following inequality:

Figures (11)

  • Figure 1: Overview of NeuroTree framework.
  • Figure 2: Aggregation for hierarchical neighborhood paths. Fig. (a) illustrates the original graph structure with weighted connectivity before pruning, providing subnetworks that represent the differences between brain regions. In Fig. (b), the zero-order path is depicted; here, aggregation initiates at the highest-scoring node, which integrates its immediate neighborhood by combining edges $e_{ab}$ and $e_{ad}$, i.e., two direct paths connecting the highest-scoring node to its neighbors. Fig. (c) shows the aggregation of higher-order paths, where isolated nodes are connected along the shortest weighted paths, thereby capturing more complex connectivity patterns beyond immediate neighbors.
  • Figure 3: Convergence analysis of $\Phi_k(t)$ over $k$-hop The spectral norm of $\Phi_k(t)$ reveals differential convergence rates across varying $k$-orders among distinct mental disorders, notably demonstrating that cannabis exhibits a steeper convergence gradient compared to COBRE as $\lambda$ increases.
  • Figure 4: The visualization of the brain tree illustrates psychiatric disorders structured into three hierarchical trunk levels. Panels (a-1) and (b-1) mark the most significant nodes along the tree path. The $l$-1 pathways represent regional connectivity, corresponding to the level three brain maps on the right. Panels (a-2) and (b-2) depict the number of connections using color gradients across the hierarchical levels.
  • Figure 5: The scatterplot shows the gaps between fMRI-predicted brain age and chronological age for healthy control and mental disorder groups.
  • ...and 6 more figures

Theorems & Definitions (12)

  • Definition 4.1: $k$-hop graph convolution
  • Theorem 4.2
  • Proposition 4.3: Convergence and Uniqueness
  • Theorem 4.4: Discretization of Age-Aware Continuous-Time Graph Convolution tang2024interpretable
  • Definition 4.5: MST algorithm
  • Theorem 1.1: K-hop Connectivity
  • proof
  • Corollary 1.2: Asymmetric Property of $K$-hop Operator
  • proof
  • proof
  • ...and 2 more