Abstract
Discovering components that are shared across multiple datasets, alongside dataset-specific features, has significant potential for studying relationships between subjects or tasks in functional Magnetic Resonance Imaging (fMRI). While coupled matrix and tensor factorization approaches offer flexibility in data fusion, existing methods often require complex post-processing and hyperparameter selection to distinguish shared from distinct components. In this paper, we propose a tensor-based framework for multi-task fMRI data fusion using a partially constrained canonical polyadic (CP) decomposition model. Unlike previous approaches, our method directly recovers shared and dataset-specific components, leading to results that are immediately interpretable. We also introduce a strategy to select highly reproducible solutions. Evaluation on real fMRI data across three tasks demonstrates that the proposed method identifies meaningful components that clearly distinguish group differences between patients with schizophrenia and healthy controls.
Introduction
Functional Magnetic Resonance Imaging (fMRI) generates vast amounts of data across various tasks and subjects. A primary goal in neuroimaging is "data fusion"—combining these datasets to identify common neural signatures while isolating features unique to specific tasks. However, standard methods like Independent Vector Analysis (IVA) or basic matrix factorizations often struggle to cleanly separate shared and distinct components, typically requiring "hard coupling" or cumbersome post-processing.
This paper addresses the need for a mathematical framework that can automatically and reliably decouple these components in a single step. By using a partially constrained coupled CP decomposition, the proposed method provides a direct and interpretable way to identify biomarkers for disorders such as schizophrenia.
Methodology
The core of the proposed framework is the Partially Constrained Coupled CP Decomposition:
- The Model: fMRI data is represented as a third-order tensor (Subjects × Voxels × Task-features). Each task's data is modeled as the sum of two tensors: one for shared components (common across all tasks) and one for distinct components (unique to that task).
- Coupling Constraints: In the shared tensor, the "Subject" and "Voxel" (spatial map) factors are constrained to be identical across all tasks, while the "Task-feature" factor remains flexible.
- Optimization: A block coordinate descent approach solves the non-convex optimization problem. To ensure meaningful spatial maps, a regularization term is included to penalize high coherence between maps.
- Reproducibility: To mitigate sensitivity to initialization, the algorithm is run 200 times. A "pseudo-metric" is then used to select the most reproducible solution, ensuring the results are robust and not mathematical artifacts.
Experimental Setup
The methodology was evaluated using the MCIC collection, comprising 271 subjects (121 with schizophrenia, 150 healthy controls). Three tasks were analyzed:
- Auditory Oddball (AOD): Subjects listen for specific target tones.
- Sternberg Item Recognition Paradigm (SIRP): A memory task split into Encode (SIRP-E) and Probe (SIRP-P) phases.
Data was reduced to feature maps using voxel-wise linear regression before being organized into tensors for decomposition.
Results and Discussion
The model successfully isolated the Default Mode Network (DMN) as a shared component across all tasks, while correctly identifying auditory regions as distinct to the AOD task. A key insight was that fusing tasks with similar requirements (such as AOD and SIRP-P, which both involve motor responses) improved the estimation of specific components, leading to more statistically significant differences between healthy controls and patients.
Clinical Biomarkers and Task Specificity
Shared components, including the DMN, motor, and frontoparietal regions, showed highly significant group differences ($p < 0.05$). In the AOD task, distinct components clearly highlighted auditory and visual regions. The auditory component, in particular, provided some of the lowest p-values, making it a highly discriminative feature for identifying schizophrenia.
Despite the complexity of the model and the 200 reproducibility runs, the total execution time was only 157.5 seconds on a standard quad-core machine, demonstrating the efficiency of the SVD-based data compression used in the framework.
Conclusion
The proposed partially constrained coupled CP decomposition offers a powerful, efficient, and interpretable tool for multi-task fMRI data fusion. While challenges remain in automated rank selection and capturing non-linear neural interactions, this framework provides a robust foundation for identifying clinical biomarkers. Future work may extend this model to multimodal data, such as combining fMRI with EEG or structural MRI, to provide an even more comprehensive view of brain disorders.
Summary updated