nerve-brain-fc-tokenization

name: nerve-brain-fc-tokenization description: > NERVE (Network-Aware Representations of Brain Functional Connectivity via Bilinear Tokenization) - self-supervised learning framework for brain functional connectivity (FC) representation learning. Redefines FC matrix tokenization by partitioning into intra/inter-network connectivity blocks, using structured bilinear factorization for heterogeneous patch sizes. Use when: building brain network ML models, self-supervised fMRI/FC representation learning, masked autoencoder for brain data, brain-behavior prediction, or developmental neuroimaging analysis. Keywords: NERVE, brain functional connectivity, bilinear tokenization, masked autoencoder, self-supervised learning, brain network, FC representation, MAE brain, connectome tokenization.

NERVE: Network-Aware Brain FC Tokenization

arXiv: 2605.14048 | Milecki et al. (May 2026)

Core Problem

Standard masked autoencoders (MAEs) for brain functional connectivity (FC) treat FC matrices as structurally homogeneous elements, ignoring the large-scale modular organization of brain networks. Fixed-size patch tokenization (from vision MAEs) is fundamentally misaligned with brain network structure.

NERVE Solution

Bilinear Tokenization

• Partition FC matrices into patches defined by network pairs (intra-network and inter-network blocks)
• Each patch corresponds to a distinct functional role (e.g., DMN-visual, fronto-parietal)
• Patches are heterogeneous in size (unlike image patches)

Structured Bilinear Factorization

• Embed FC patches through low-rank bilinear decomposition
• Preserves network identity information
• Reduces parameter complexity from quadratic to linear in number of networks
• Key insight: W = A ⊗ B where A captures source-network factors, B captures target-network factors

Architecture

FC Matrix (N×N)
    ↓ Network-aware partitioning
Patches (heterogeneous sizes, network-pair defined)
    ↓ Bilinear factorization
Low-dim embeddings preserving network identity
    ↓ Masked Autoencoder
Self-supervised representation
    ↓ Downstream: behavior/psychopathology prediction

Key Advantages

1. Network-aware: Respects known brain network organization (DMN, FPN, visual, etc.)
2. Parameter efficient: Linear scaling vs. quadratic in number of networks
3. Transferable: Stable representations across developmental cohorts
4. Interpretable: Patches map to known functional circuits

Evaluated On

• ABCD (Adolescent Brain Cognitive Development)
• PNC (Philadelphia Neurodevelopmental Cohort)
• CCNP (Connectomes Related to Human Brain Development)

Applications

• Behavior prediction from resting-state fMRI
• Psychopathology screening
• Developmental trajectory analysis
• Transfer learning across cohorts

When to Use NERVE Approach

• Self-supervised pre-training on FC matrices
• Brain-behavior prediction tasks
• Multi-cohort transfer learning
• When interpretability of FC representations matters

Activation

• NERVE, bilinear tokenization, brain FC
• 脑功能连接, 脑网络表示学习
• Masked autoencoder for brain data
• FC matrix tokenization, self-supervised fMRI