Research map / Graph learning / Graph self-supervised learning
NeuroTree: Hierarchical Functional Brain Pathway Decoding for Mental Health Disorders
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes NeuroTree, a k-hop ODE-based age-aware GCN with contrastive masked FC loss that decodes fMRI graphs into hierarchical brain pathway trees for classifying cannabis use disorder and schizophrenia and estimating brain age.
Problem
Existing fMRI graph neural networks fail to capture complex relations between brain regions and demographic factors (e.g., age) and ignore hierarchical organization of brain networks in mental disorders.
Approach
Combines a k-hop AGE-GCN (age-modulated graph convolution discretized from neural ODEs) with a node-wise contrastive masked functional connectivity loss, then builds a weighted hierarchical brain tree via MST pruning and high-order path aggregation.
Key results
- On the Cannabis dataset NeuroTree reports AUC 0.80 and accuracy 0.73 under five-fold cross-validation, versus 0.74 AUC for the best ablation without CMFC loss.
- On COBRE NeuroTree reports AUC 0.71 and accuracy 0.65, versus MixHop at AUC 0.69 and accuracy 0.61.
- Brain age prediction MSE is lowest for the 18-25 group (9.51 on Cannabis, 4.48 on COBRE).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add NeuroTree's hierarchical age-aware k-hop ODE-GCN and CMFC framework | graph classification · Cannabis | AUC: 0.73 -> 0.80 | Helped |
| add NeuroTree's hierarchical age-aware k-hop ODE-GCN and CMFC framework | graph classification · COBRE | AUC: 0.69 -> 0.71 | Helped |
| add age parameter θ (ablation) | NeuroTree · Cannabis | AUC: 0.49 -> 0.80 | Helped |
| add age parameter θ (ablation) | NeuroTree · Cannabis | accuracy: 0.60 -> 0.73 | Helped |
| add age parameter θ (ablation) | NeuroTree · COBRE | AUC: 0.50 -> 0.71 | Helped |
| add age parameter θ (ablation) | NeuroTree · COBRE | accuracy: 0.47 -> 0.65 | Helped |
| add CMFC loss (ablation) | NeuroTree · Cannabis | AUC: 0.74 -> 0.80 | Helped |
| add CMFC loss (ablation) | NeuroTree · Cannabis | accuracy: 0.73 -> 0.73 | No clear effect |
| add CMFC loss (ablation) | NeuroTree · COBRE | AUC: 0.69 -> 0.71 | Helped |
| add CMFC loss (ablation) | NeuroTree · COBRE | accuracy: 0.63 -> 0.65 | Helped |
| set λ to 0.2 instead of other tested values (ablation) | NeuroTree · Cannabis | accuracy: 0.8 | Mixed |
| set λ to 0.8 instead of other tested values (ablation) | NeuroTree · COBRE | accuracy: 0.8 | Mixed |
Surprising findings
- The variant without the age parameter (w/o θ) performs near chance on Cannabis (AUC 0.49), and PathNN underperforms MixHop on COBRE.
Limitations
- Reported improvements are small relative to standard deviations on small cohorts (COBRE n=142, Cannabis n=323).
- Dataset release is promised only upon acceptance, and the reported results may vary across GPU configurations.
- Interpretability claims about age-related pathways are qualitative and not validated against external clinical measures.
Experimental setup
Datasets: Cannabis fMRI dataset COBRE
Benchmarks: Cannabis COBRE
Metrics: AUC accuracy precision recall MSE Pearson r
Baselines: Pearson GCN k-NN GCN GAT BrainGNN BrainUSL BrainGSL MixHop GPC-GCN PathNN
Builds on: Graph Convolutional Networks STE-ODE PathNN MixHop BrainGNN
Closest papers by meaning
- BrainMAP: Learning Multiple Activation Pathways in Brain Networks (AAAI 2025 · similarity 0.87)
- $\textit{NeuroPath}$: A Neural Pathway Transformer for Joining the Dots of Human Connectomes (NeurIPS 2024 · similarity 0.87)
- BrainOOD: Out-of-distribution Generalizable Brain Network Analysis (ICLR 2025 · similarity 0.85)
- NeuroBOLT: Resting-state EEG-to-fMRI Synthesis with Multi-dimensional Feature Mapping (NeurIPS 2024 · similarity 0.84)
- Functional Connectomes of Neural Networks (AAAI 2025 · similarity 0.84)
- Beyond Pairwise Connections: Extracting High-Order Functional Brain Network Structures under Global Constraints (NeurIPS 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.