brain-dit-fmri-foundation-model-v5

name

brain-dit-fmri-foundation-model-v5

description

Brain-DiT v5 universal multi-state fMRI foundation model with pre-training and fine-tuning for zero-shot and few-shot brain decoding across multiple states. Supports cross-task, cross-subject, and cross-dataset fMRI analysis using diffusion transformer architecture. Use when: fMRI foundation models, brain decoding, diffusion transformers for neuroimaging, cross-subject fMRI analysis, zero-shot brain state prediction, multi-task fMRI modeling, neural state decoding, fMRI pre-training. Activation: Brain-DiT, fMRI foundation model, diffusion transformer brain, multi-state fMRI, brain decoding, cross-subject fMRI, fMRI pre-training, neural state prediction, zero-shot brain analysis.

Brain-DiT v5: Universal Multi-State fMRI Foundation Model

Overview

Brain-DiT is a diffusion transformer-based foundation model for fMRI data that supports:

• Pre-training on large-scale fMRI datasets
• Zero-shot and few-shot fine-tuning for downstream tasks
• Cross-subject, cross-task, and cross-dataset generalization
• Multi-state brain activity modeling

Source Paper: Brain-DiT v5 (arXiv:2505.00936, 2025-05-01)

Core Architecture

┌──────────────────────────────────────────────┐
│              Brain-DiT Architecture           │
├──────────────────────────────────────────────┤
│  fMRI Input → Patch Embedding                 │
│       ↓                                       │
│  DiT Blocks (diffusion transformer layers)    │
│       ↓                                       │
│  State Conditioning (task/stimulus labels)    │
│       ↓                                       │
│  Output: Brain state reconstruction/prediction │
└──────────────────────────────────────────────┘

Key Innovations

1. Universal Multi-State Modeling: Single model handles multiple brain states/tasks
2. Diffusion-Based Generation: Uses diffusion process for robust fMRI prediction
3. Cross-Subject Generalization: Learns subject-invariant representations
4. Foundation Model Pre-training: Scales to large fMRI corpora

Usage Pattern

# Fine-tuning for new task
model = BrainDiT(pretrained="base")
model.fine_tune(
    dataset=new_fMRI_dataset,
    task="classification",  # or "reconstruction", "prediction"
    n_shots=5,  # few-shot learning
    epochs=50
)

# Zero-shot inference
predictions = model.predict(new_fMRI_data, task="unknown_task")

Applications

• Brain-computer interfaces: Decode intent from fMRI signals
• Clinical diagnosis: Detect neurological disorders from brain activity patterns
• Cognitive neuroscience: Understand multi-task brain organization
• Cross-study analysis: Harmonize fMRI data across different studies

Related Skills

• brain-dit-fmri-foundation-model — Brain-DiT overview
• brain-dit-universal-multi-state — Brain-DiT multi-state modeling
• meta-learning-in-context-brain-decoding — Cross-subject brain decoding