magnet-brain-structure-function-gnn

name	magnet-brain-structure-function-gnn
description	Multi-Scale Adaptive Graph Network (MAGNet) for learning structural-functional brain representations. Models structure-function coupling for cognitive insight.
version	1.0.0
author	Research Synthesis
license	MIT
metadata	{"hermes":{"tags":["brain-network","structure-function","graph-attention","cognitive-insight","gnn","neuroimaging"],"source_paper":"Learning Structural-Functional Brain Representations through Multi-Scale Adaptive Graph Attention for Cognitive Insight (arXiv:2603.29967v1)"}}

MAGNet: Multi-Scale Adaptive Graph for Brain Structure-Function

Overview

MAGNet (Multi-scale Adaptive Graph Network) learns joint structural-functional brain representations through adaptive graph attention. By modeling how structural connectivity constrains and enables functional dynamics across multiple spatial scales, the framework provides cognitive insights into brain organization.

Core Concepts

Multi-Scale Architecture

• Local Scale: Regional connectivity patterns within brain modules
• Mesoscale: Inter-module connectivity and hub regions
• Global Scale: Whole-brain integration and small-world topology

Adaptive Graph Attention

• Attention weights adapt based on both structural and functional features
• Learns which structural connections are most relevant for functional prediction
• Dynamic re-weighting across scales

Structure-Function Coupling

• Structural connectome (from DTI) as graph scaffold
• Functional connectome (from fMRI) as node dynamics
• Coupling strength varies by region and cognitive state

Implementation Pattern

class MAGNet(nn.Module):
    def __init__(self, n_regions, n_scales=3, hidden_dim=128):
        super().__init__()
        self.scales = nn.ModuleList([
            ScaleAttention(n_regions, hidden_dim, scale_factor=2**i)
            for i in range(n_scales)
        ])
        self.fusion = nn.Linear(hidden_dim * n_scales, hidden_dim)
        self.decoder = nn.Linear(hidden_dim, n_regions)
    
    def forward(self, structural_conn, functional_signals):
        scale_features = []
        for scale_module in self.scales:
            feat = scale_module(structural_conn, functional_signals)
            scale_features.append(feat)
        fused = torch.cat(scale_features, dim=-1)
        fused = self.fusion(fused).relu()
        return self.decoder(fused)

Applications

• Brain structure-function coupling analysis
• Cognitive trait prediction from neuroimaging
• Brain network biomarker discovery
• Personalized neuroscience

Activation Keywords

• structure-function brain coupling, multi-scale brain network, adaptive graph attention brain, MAGNet brain model, structural connectome analysis, 脑结构功能耦合, 多尺度脑网络

References

• Learning Structural-Functional Brain Representations through Multi-Scale Adaptive Graph Attention for Cognitive Insight
• Authors: Badhan Mazumder, Sir-Lord Wiafe, Aline Kotoski, Vince D. Calhoun, Dong Hye Ye
• Published: 2026-03-31
• arXiv: https://arxiv.org/abs/2603.29967v1