Research map / Architectures and efficiency / Attention and transformers
$\textit{NeuroPath}$: A Neural Pathway Transformer for Joining the Dots of Human Connectomes
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 NeuroPath, a graph Transformer whose multi-head self-attention is masked by structural-connectivity multi-hop detour adjacency (plus FC masks) to model how SC pathways support FC links, evaluated on HCPA, UKB, ADNI, and OASIS.
Problem
Predicting cognitive tasks and disease status from paired structural and functional brain connectivity while capturing how SC pathways physically support FC links, which existing deep models do not model well.
Approach
Introduces the 'topological detour' concept: multi-hop SC paths that support a direct FC link, implemented as a twin-branch multi-head self-attention where one branch is masked by h-hop detour adjacency and the other by FC adjacency, with a consistency loss aligning the two branches.
Key results
- NeuroPath achieves 99.22% accuracy on UKB CORR static, the best among compared models, and 99.59% on UKB BOLD static.
- On ADNI CORR static, NeuroPath reaches 85.56% accuracy and 83.29% F1, the best among compared methods, and it is the only model above 90% accuracy on OASIS.
- In zero-shot HCPA->UKB static resting/tasking classification, NeuroPath reaches 91.29 F1 versus 74.49 for the next best (NAGphormer).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add both TD-MHSA and FC-MHSA branches (twin branch) instead of no branch (ablation) | NeuroPath · ADNI static BOLD | accuracy: 82.42 -> 85.56 | Helped |
| Add both TD-MHSA and FC-MHSA branches instead of using TD-MHSA only (ablation) | NeuroPath · HCPA static BOLD | accuracy: 97.33 -> 98.23 | Helped |
| Add both TD-MHSA and FC-MHSA branches instead of using FC-MHSA only (ablation) | NeuroPath · UKB static BOLD | accuracy: 99.25 -> 99.59 | Helped |
| Add both branches instead of FC-MHSA only (ablation) | NeuroPath · OASIS static BOLD | F1 score: 86.58 -> 86.37 | Hurt |
| Increase pathway head number to H=7 (ablation) | NeuroPath · HCPA and UKB | F1 score | Helped |
| Increase pathway head number from 1 to 8 (ablation) | NeuroPath · ADNI and OASIS | F1 score | No clear effect |
| Use deeper model (16 layers rather than 4 or 8) (ablation) | NeuroPath · HCPA, UKB, ADNI, and OASIS | F1 score: HCPA: 97.76, 97.72, 96.60; UKB: 99.59, 99.61, 99.44; ADNI: 78.93, 78.42, 78.32; OASIS: 86.16, 86.77, 85.78 | Mixed |
| Change FC threshold among 0.3, 0.5, and 0.7 (ablation) | NeuroPath · HCPA, UKB, ADNI, and OASIS | F1 score: HCPA: 97.57, 95.09, 97.32; UKB: 99.27, 99.59, 99.15; ADNI: 78.36, 77.35, 79.49; OASIS: 86.59, 87.02, 86.13 | Mixed |
| Train on HCPA and test on UKB rather than using a same-dataset setting | NeuroPath · HCPA→UKB static BOLD zero-shot | F1 score: 91.29 vs 74.49 | Helped |
| Use NeuroPath instead of the strongest baseline for ADNI→OASIS dynamic zero-shot | NeuroPath · ADNI→OASIS dynamic BOLD zero-shot | F1 score: 79.65 vs 83.55 | Hurt |
Surprising findings
- A plain GCN ties or beats many sophisticated brain models on UKB, which the authors interpret as those models rarely capturing neuroscience-relevant features.
- Graphormer collapses on HCPA and UKB with deeper layers (below 50 F1), while NeuroPath stays stable.
Limitations
- Top pathway representation covers only about half of neural pathways in theory, and performance is lower on dynamic graphs with higher degree.
- Disease datasets (ADNI, OASIS) are small.
- Zero-shot performance still lags the fully supervised results.
Experimental setup
Datasets: HCPA UK Biobank ADNI OASIS
Benchmarks: HCPA UK Biobank ADNI OASIS
Metrics: accuracy F1 score
Baselines: MLP GCN BrainGNN BNT BolT Graphormer NAGphormer
Builds on: Graphormer Transformer NAGphormer
Closest papers by meaning
- Functional Connectomes of Neural Networks (AAAI 2025 · similarity 0.89)
- Exploring the Enigma of Neural Dynamics Through A Scattering-Transform Mixer Landscape for Riemannian Manifold (ICML 2024 · similarity 0.87)
- NeuroTree: Hierarchical Functional Brain Pathway Decoding for Mental Health Disorders (ICML 2025 · similarity 0.87)
- BrainMAP: Learning Multiple Activation Pathways in Brain Networks (AAAI 2025 · similarity 0.87)
- Learning dynamic representations of the functional connectome in neurobiological networks (ICLR 2024 · similarity 0.86)
- Beyond Pairwise Connections: Extracting High-Order Functional Brain Network Structures under Global Constraints (NeurIPS 2025 · similarity 0.86)
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.