Introduction
Functional magnetic resonance imaging (fMRI) provides a non-invasive window into neural activity, yet the resulting 4D signals are notoriously complex, with spatial and temporal patterns deeply intertwined. Traditional analysis methods often struggle with two primary challenges: a heavy reliance on prior knowledge of experimental design (e.g., General Linear Models) and a focus on group-average "common" networks that obscure individual-specific functional "fingerprints." To address these limitations, this research introduces a novel "Twin-Transformers" framework designed to disentangle these features using self-supervised learning, enabling the discovery of both universal and unique brain networks without prior task information.
Methodology
The proposed architecture treats fMRI data as a 2D matrix (Time × Space) and employs a dual-transformer approach to process these dimensions concurrently.
Twin-Transformers Architecture
The framework consists of two specialized components:
- Spatial Transformer: This module partitions the brain into spatial patches and utilizes self-attention mechanisms to identify voxels belonging to specific functional networks.
- Temporal Transformer: This module analyzes time-series data to extract the characteristic signal patterns associated with each identified network.
Spatial-Temporal Commonality-Variability (ST-CV) Loss
To facilitate learning across subjects, the model incorporates a multi-faceted loss function. It employs Self-Supervised Reconstruction to ensure no information is lost, Spatial Commonality to align maps across subjects, and Temporal Commonality to maximize correlation of shared network patterns. Crucially, the model reserves a subset of network components to capture individual-specific patterns, allowing for the identification of neural signatures that deviate from the group template.
Experimental Setup
The framework was evaluated using the Human Connectome Project (HCP) 3T motor task-fMRI dataset, comprising 917 subjects. The motor task involved specific movements (tongue, fingers, toes) prompted by visual cues. Data was downsampled to 4mm resolution to manage computational overhead, resulting in over 28,000 voxels per subject.
Results and Discussion
The Twin-Transformers model successfully identified task-related and resting-state networks without access to the experimental design.
Task-Free Network Discovery
The model consistently recovered three major task-related networks: Tongue, Right Foot, and Left Foot. The temporal patterns identified for these networks showed a high Pearson Correlation Coefficient (PCC = 0.68) with the actual experimental task design, demonstrating the model's ability to extract meaningful neural structures from raw signals.
Resting-State and Individual Networks
Beyond task-specific activity, the model identified 9 common resting-state networks (e.g., Default Mode Network) active during task performance. Furthermore, it successfully isolated individual-specific networks—regions that synchronize within a single subject but do not align with the group average or task timing. Reproducibility tests yielded Jaccard similarity scores between 0.74 and 0.81, confirming the robustness of the findings.
Conclusion and Future Work
The Twin-Transformers framework represents a significant step toward personalized neuroscience by disentangling the universal and individual components of brain function. Future work will focus on improving computational efficiency to handle higher-resolution data and exploring the clinical utility of individual-specific networks in predicting behavioral traits or diagnosing neurological conditions.
Summary updated