Neural Dynamics-Informed Pre-trained Framework for Personalized Brain Functional Network Construction

This paper proposes a neural dynamics-informed pre-trained framework that overcomes the limitations of traditional atlas-based methods by extracting personalized neural activity representations to guide brain parcellation and correlation estimation, thereby achieving superior performance in constructing personalized brain functional networks across heterogeneous scenarios.

The Gist

The Big Picture: Why We Need a New Map

Imagine your brain is a massive, bustling city. To understand how this city works (how you think, feel, and move), scientists try to draw a "map" of the traffic flow between different neighborhoods. This map is called a Brain Functional Network.

For a long time, scientists used a standard, pre-printed map (like a generic tourist guide) to study everyone's brain. They assumed that:

1. Everyone's neighborhoods are in the exact same spots.
2. The traffic rules (how neighborhoods talk to each other) are the same for everyone.

The Problem: This doesn't work well in the real world.

• The City Changes: A child's brain city looks different from an elderly person's. A brain with Parkinson's looks different from a healthy one. Even the way you take the "photo" of the city (the MRI machine settings) changes the view.
• The Old Map Fails: If you use a generic map for a specific person, you miss the unique traffic jams and shortcuts that make their brain special. It's like trying to navigate a specific neighborhood in Tokyo using a map of New York; the streets might look similar, but the actual flow is wrong.

The Solution: A "Smart GPS" for Every Brain

The authors of this paper built a new system called a "Neural Dynamics-Informed Pre-trained Framework." That's a mouthful, so let's call it the "Smart Brain GPS."

Instead of using a static, pre-printed map, this system builds a custom map for every single person, instantly adapting to their age, health, and the specific camera used to scan them.

Here is how it works, step-by-step:

1. The "Super-Student" Foundation (Pre-training)

First, the team taught a giant AI model by showing it thousands of brain scans from the UK Biobank.

• Analogy: Imagine a student who reads every library book in the world. They learn the general rules of how cities (brains) usually work. They know what a "traffic jam" looks like and how neighborhoods usually connect. This is the Foundation Model.

2. The "Physics Teacher" (Neural Dynamics)

Here is the secret sauce. The AI isn't just guessing; it's taught the laws of physics that govern how brain signals travel.

• Analogy: Think of the brain like a drum. If you hit one spot, the vibration ripples out in a specific wave pattern based on the drum's shape. The AI learns these "ripple rules" (Neural Dynamics). This ensures the map isn't just a random guess; it respects how brain signals actually physically move through space and time.

3. The "Custom Tailor" (Fine-Tuning)

When a new patient comes in (say, a teenager with ADHD), the AI takes its "Super-Student" knowledge and the "Physics Rules" and quickly tailors a map just for them.

• Analogy: It's like a high-end tailor who has a master pattern (the foundation model) but instantly adjusts the fabric to fit your exact body shape, posture, and style. It doesn't force your brain into a generic box; it molds the map to your brain.

What Did They Prove? (The Results)

The team tested this "Smart GPS" on 18 different datasets involving children, adults, the elderly, and people with various conditions (like Alzheimer's, depression, and autism). They compared it to the old "pre-printed map" methods.

The Results were impressive:

1. It's More Consistent: If you scan the same person twice in a row, the old method draws two slightly different maps. The new method draws the same reliable map every time. It's like a GPS that doesn't glitch when you drive over a bump.
2. Better Diagnosis: When trying to tell if someone has a brain disorder (like Autism or Parkinson's), the new method was much more accurate. It spotted the "traffic patterns" that indicate disease better than the old methods.
3. Better "Virtual Surgery": The researchers used the new maps to simulate "virtual neural modulation" (imagining what would happen if we stimulated a specific part of the brain with electricity).
   • Analogy: Imagine a doctor trying to fix a broken engine by guessing which wire to cut. The old method guessed randomly. The new method looked at the custom map, found the exact wire, and simulated the fix. The simulation showed that their method would actually help patients recover much faster than the old guesses.
4. Finding the "Culprit" Circuits: When looking for the specific brain circuits that go wrong in diseases, the new method found the same "bad circuits" even when the data came from different hospitals with different machines. The old method got confused by the different machines.

Why Does This Matter?

This paper challenges the old way of doing things. It says: "Stop using a one-size-fits-all map for a world of unique brains."

By combining Big Data (learning from thousands of brains) with Physics (understanding how signals move), they created a tool that can:

• Diagnose brain diseases more accurately.
• Predict how a patient will respond to treatment.
• Help doctors design better therapies (like brain stimulation) that are tailored to the individual, not the average.

In short: They moved brain science from using a generic street atlas to using a real-time, personalized GPS that knows exactly how your unique brain city flows.

Technical Summary

1. Problem Statement

Current methods for constructing brain functional networks (BFNs) rely heavily on pre-defined brain atlases (e.g., Desikan-Killiany, Schaefer) and linear assumptions (e.g., Pearson correlation) to estimate neural activity correlations. These dominant approaches face critical limitations in heterogeneous scenarios (varying age groups, brain disorders, races, and imaging acquisition strategies) because:

• Static Atlases: Pre-defined atlases are often derived from ex vivo specimens or group averages, failing to capture subject-specific variations in functional boundaries caused by age, pathology, or linguistic differences.
• Linear Assumptions: Linear correlation metrics fail to characterize the complex, non-linear, and dynamic variations in neural activity correlations across different states and subjects.
• Consequence: This leads to BFNs with poor consistency, low generalizability, and an inability to accurately reflect personalized neural activity patterns, limiting their utility in clinical diagnosis, neural modulation, and abnormal circuit identification.

2. Methodology

The authors propose a Neural Dynamics-Informed Pre-trained Framework that follows a "Pre-training $\rightarrow$ Fine-tuning $\rightarrow$ Personalized Construction" paradigm. The framework consists of three core stages:

A. Neural Dynamics-Informed Pre-trained Foundation Model

• Architecture: A latent diffusion model serves as the core, comprising an fMRI encoder ($E$), a latent diffusion model, and two decoders ($D$ for reconstruction, $D_1$ for mapping to cortical space).
• Pre-training Strategy:
  1. Stage 1: Train $E$ and $D$ using an autoencoder approach to reconstruct fMRI signals from large-scale data (UK Biobank).
  2. Stage 2: Pre-train the diffusion model in the latent space to learn the probability distribution of neural signals.
  3. Stage 3: Train decoder $D_1$ to map personalized representations back to the cortical space.
• Neural Dynamics Constraint (Fine-tuning): To embed intrinsic brain dynamics, the foundation model is fine-tuned using a neural wave equation. The model minimizes a physics-informed loss ($L_{phy}$) that ensures the extracted personalized representations ($Z_{pattern}$) satisfy the constraints of neural signal propagation (spatial diffusion and temporal damping) on the cortical surface.

B. Personalized Representation-Guided Adaptive Brain Parcellation

• Instead of using fixed atlases, the framework uses the personalized representation $Z_{pattern}$ to guide brain parcellation.
• Mechanism: The representation is mapped to the cortical space via $D_1$. A self-supervised clustering approach initializes the labels, which are then optimized using a Markov Random Field (MRF) model.
• Optimization: The process minimizes a Gibbs energy function containing unary potentials (functional consistency via vMF distribution) and pairwise potentials (spatial aggregation via gradient-weighted Potts model), ensuring anatomical plausibility while adapting to individual functional boundaries.

C. Personalized Representation-Guided Correlation Estimation

• Architecture: A Transformer-based correlation encoder.
• Mechanism:
  1. Intra-attention: Captures correlations within the region-level neural activity matrix derived from the adaptive parcellation.
  2. Cross-attention: Fuses the intra-attention features with the personalized representation ($Z_{pattern}$). This allows the model to dynamically retrieve and integrate prior patterns specific to the subject's scenario, enabling the estimation of non-linear and varying correlation patterns.
• Output: The final personalized BFN is constructed by calculating correlation coefficients from the fused features.

D. Joint Fine-Tuning

The three modules (representation extraction, parcellation, and correlation estimation) are jointly fine-tuned using a combined loss function:
$$L_{total} = L_{task} + \lambda_1 L_{seg}$$
Where $L_{task}$ is the downstream task loss (e.g., classification) and $L_{seg}$ is the segmentation loss derived from the parcellation energy function.

3. Key Contributions

1. Novel Framework: Proposes the first framework to construct personalized BFNs by integrating neural dynamics (wave equations) into a pre-trained foundation model, moving beyond static atlases and linear assumptions.
2. Adaptive Parcellation & Correlation: Introduces a mechanism where personalized representations dynamically guide both brain region parcellation and correlation estimation, capturing subject-specific functional boundaries and non-linear interactions.
3. Open-Source Tool: Releases a Python tool allowing users to contribute or customize personalized parcellation and correlation methods, fostering a flexible ecosystem for BFN construction.
4. Comprehensive Validation: Validated across 18 diverse datasets covering multiple age groups (children to elderly), five major brain disorders (AD, PD, MDD, ADHD, ASD), and varying acquisition strategies (scan duration, resolution).

4. Experimental Results

The framework was evaluated against the baseline Micapipe method across multiple dimensions:

• Consistency (Portrait Divergence):

  • Personalized BFNs showed significantly higher consistency (lower Portrait Divergence) across short-term sessions and heterogeneous scenarios compared to the baseline.
  • Median "1 - PDiv" values were 0.7–0.8 for the proposed method vs. 0.5–0.7 for the baseline.
  • High consistency was achieved without sacrificing the ability to differentiate between distinct disorders (verified via t-SNE).

• Downstream Task Performance:

  • Diagnosis: Outperformed baselines in diagnosing AD, PD, MDD, ADHD, and ASD, achieving median accuracies of 0.73–0.90 (vs. <0.73 for baseline).
  • Prediction: Significantly improved prediction of physiological indices (Age $R^2 \approx 0.90$; BMI) and motor imagery decoding accuracy (>0.80).
  • Individual Identification: Achieved superior accuracy in identifying unique subjects.

• Virtual Neural Modulation:

  • Identified modulation targets (e.g., Somatomotor Network for Parkinson's) that aligned better with empirical clinical evidence than the baseline.
  • Achieved higher recovery rates in virtual perturbation simulations (e.g., +8% improvement for PD, +6.3% for ASD).
  • Successfully normalized pathological brain activity patterns toward healthy states in MDD patients.

• Abnormal Circuit Identification:

  • Identified abnormal neural circuits with significantly higher cosine similarity across datasets with different acquisition strategies (e.g., different TRs and resolutions) compared to the baseline.
  • Demonstrated robust consistency in complex network metrics (characteristic path length, clustering coefficient).

5. Significance and Future Outlook

• Paradigm Shift: Challenges the dominance of static, atlas-based BFN construction by establishing a personalized, dynamics-informed paradigm.
• Clinical Impact: Provides a robust basis for precision medicine, enabling more accurate diagnosis, effective neural modulation target localization, and reliable identification of pathological circuits across diverse patient populations.
• Limitations & Future Work:
  • Current pre-training relies on UK Biobank data (biased toward English speakers, ages 40-69), which may limit generalization to children or non-native speakers.
  • Currently uses only a wave model for neural dynamics; future work aims to integrate multi-perspective models (neural mass, hemodynamic) and expand pre-training data to cover broader demographics and acquisition strategies.
  • The authors propose a future "Foundation Model-Personalized Adaptation" paradigm to further decouple subject-specific patterns and enhance precision diagnosis and treatment.