brain-dit-universal-multi-state - SKILL.md Agent Skill

name: brain-dit-universal-multi-state description: "Brain-DiT universal multi-state fMRI foundation model methodology. Integrates diffusion transformer architecture with fMRI data for generative modeling and brain state analysis. Covers multi-state fMRI generation, brain state transition modeling, and foundation model fine-tuning for neuroscience applications. Use when working with fMRI foundation models, brain state generation, diffusion models for neuroimaging, or multi-state neural dynamics simulation." version: 1.0.0 author: Research Synthesis license: MIT metadata: hermes: tags: [fmri, foundation-model, diffusion-transformer, brain-state, generative-model, neural-dynamics] source_paper: "Brain-DiT: Universal Multi-State fMRI Foundation Model" citations: 0

Brain-DiT: Universal Multi-State fMRI Foundation Model

Overview

Brain-DiT is a diffusion transformer architecture designed as a universal foundation model for fMRI data. It enables generative modeling of multiple brain states, brain state transitions, and transfer learning across neuroscience tasks.

Core Architecture

Diffusion Transformer (DiT) Backbone

Transformer-based diffusion model for fMRI spatiotemporal data
Multi-state conditioning for different cognitive/clinical states
Scalable architecture supporting various fMRI resolutions
Pre-training on large-scale fMRI datasets

Multi-State Conditioning

State embeddings for different brain conditions (rest, task, disease)
Cross-attention mechanisms for state-specific generation
Continuous state space for interpolating between conditions
Temporal dynamics modeling for state transitions

Implementation Patterns

# Brain-DiT multi-state fMRI generation
def generate_brain_state(model, target_state, n_samples=1, guidance_scale=7.5):
    """Generate fMRI data for a specific brain state."""
    # 1. Encode target state
    state_embedding = model.encode_state(target_state)
    
    # 2. Conditional diffusion sampling
    generated_fmri = model.sample(
        condition=state_embedding,
        n_samples=n_samples,
        guidance_scale=guidance_scale,
        steps=50
    )
    
    return generated_fmri

# Brain state transition modeling
def model_state_transition(model, from_state, to_state, n_steps=20):
    """Model transitions between brain states."""
    # Interpolate in state space
    path = model.interpolate_states(from_state, to_state, n_steps)
    transitions = [model.sample(condition=s) for s in path]
    return transitions

Key Methodologies

1. Foundation Model Pre-training

Large-scale fMRI dataset collection and preprocessing
Self-supervised learning for neural representation
Multi-site harmonization for scanner effects
Cross-task and cross-population generalization

2. State-Specific Fine-tuning

Adapter-based fine-tuning for specific tasks
Low-rank adaptation for efficient transfer
Few-shot learning for rare brain states
Domain adaptation across populations

3. Brain State Analysis

Latent space analysis for neural representations
State similarity and clustering
Trajectory analysis for dynamical systems
Biomarker extraction from latent features

Applications

Synthetic fMRI Generation: Data augmentation for small datasets
Brain State Classification: Leveraging pre-trained features
Disease Modeling: Generating pathological brain states
Treatment Simulation: Modeling intervention effects
Cross-Site Harmonization: Reducing scanner variability

Pitfalls

fMRI data requires careful preprocessing (motion correction, normalization)
Diffusion models are computationally expensive for large volumes
Multi-site data needs harmonization before training
State conditioning requires careful definition and validation
Foundation models may learn scanner-specific artifacts

Verification Steps

Validate generated fMRI against real data distributions
Check state classification accuracy using generated data
Verify temporal consistency in state transitions
Assess cross-site generalization performance
Evaluate clinical/biological plausibility of generated states