Abstract
Neuroimaging analysis often faces a trade-off between group-level models that ignore individual functional fingerprints and subject-specific models that fail to scale to whole-brain data. We introduce a "Decomposed" manifold learning framework that addresses this gap. By utilizing a matrix decomposition strategy inspired by Singular Value Decomposition (SVD), our model captures subject-specific spatial weights with significantly fewer parameters. Our results across simulated and real-world fMRI datasets show that this framework is both scalable and sensitive to individual differences, enabling unsupervised discovery of clinically relevant patterns in disorders such as schizophrenia.
Introduction
Functional magnetic resonance imaging (fMRI) data is high-dimensional and noisy, leading researchers to hypothesize that brain activity resides on a low-dimensional manifold—a simpler geometric structure hidden within complex data. However, existing machine learning approaches typically fall into two traps: the "Average Trap," where models assume spatial identity across subjects, and the "Scaling Trap," where models learn independent parameters for each individual. While Group models are computationally efficient, they lose the unique functional "fingerprint" of the individual. Conversely, Subject models suffer from massive memory overhead, requiring billions of parameters for whole-brain datasets.
The Decomposed Framework
To overcome these limitations, we propose the Decomposed model. This architecture avoids learning a massive, unique weight matrix ($W_i$) for every subject. Instead, it decomposes the subject-specific weight matrix into three components: $W_i = U S_i V^T$.
In this formulation, $U$ and $V$ are shared across all subjects, representing a "shared vocabulary" of brain patterns. The diagonal matrix $S_i$ contains singular values unique to each subject. This approach reduces the parameter count by approximately three orders of magnitude, making voxelwise whole-brain analysis computationally feasible.
Experimental Validation
The framework was validated across a diverse range of data, including a simulated "Half Moons" toy dataset and naturalistic fMRI data (Sherlock and Forrest Gump). In the simulation, "subjects" were generated by applying random rotations to a 2D distribution to test if the model could "un-rotate" them.
The Decomposed model successfully recovered these subject-specific rotations. When the learned subject-specific weights ($S_i$) were projected into a 2D space using PCA, they formed a perfect circle, mathematically recovering the exact rotation used to generate the data. This demonstrates that the model captures the true underlying variance between individuals.
Key Findings
- Scalability and Performance: The Decomposed model significantly outperformed standard Group models in reconstructing brain activity and classifying cognitive states (e.g., identifying if a movie scene was "indoor" or "outdoor"). Unlike individual Subject models, it handled whole-brain data without memory failure.
- Clinical Discovery: When applied to the fBIRN resting-state fMRI dataset, the model—trained as a Variational Autoencoder (VAE) without diagnostic labels—learned subject weights that allowed a simple classifier to identify schizophrenia patients with ~79% accuracy.
- Zero-Shot Adaptation: The model can adapt to a brand-new subject using as little as a single time-point (TR) of fMRI data by fine-tuning only the $S_i$ weights while keeping the shared brain "vocabulary" frozen.
Conclusion
The Decomposed manifold learning framework provides a "Goldilocks" solution for neuroimaging, balancing individual precision with computational efficiency. By mapping minds rather than averages, this approach paves the way for scalable fMRI foundation models and personalized clinical diagnostics. Future work will explore incorporating subject-specific temporal dynamics and further improving the biological interpretability of the learned weights.
Summary updated