Coupled CP Tensor Decomposition with Shared and Distinct Components for Multi-Task Fmri Data Fusion

R. A. Borsoi, Isabell Lehmann, Mohammad A. B. S. Akhonda, Vince D. Calhoun, Konstantin Usevich, David Brie, Tülay Adali

Introduction

Functional Magnetic Resonance Imaging (fMRI) researchers often face a significant challenge when analyzing data across multiple tasks, such as memory and auditory paradigms. This "fusion" challenge stems from the fact that brain networks exhibit both shared and distinct behaviors. For instance, the Default Mode Network (DMN) is typically active across various tasks, while other networks are highly specialized to specific stimuli.

Existing data fusion methods, such as Independent Vector Analysis (IVA) or standard matrix factorizations, often struggle to cleanly separate these features. They typically require complex post-processing steps to distinguish shared components from task-specific ones, which increases hyperparameter tuning requirements and can reduce reproducibility. This paper introduces a novel tensor-based framework that simultaneously identifies shared and unique brain activity patterns in a single, interpretable step.

Methodology

The researchers propose a Partially Constrained Coupled CP (Canonical Polyadic) Tensor Decomposition. The framework represents fMRI data as third-order tensors (Subjects × Voxels × Task-features) and employs an explicit decoupling strategy. Unlike standard models, this approach splits the decomposition into two parallel paths: one for components shared across all task datasets and another for components unique to each task.

To ensure biological and mathematical validity, the model enforces specific constraints: shared components must utilize the same spatial maps and subject factors across tasks, while distinct components are permitted to vary. The optimization is handled via a block coordinate descent approach, including a regularization term to penalize high coherence between spatial maps, preventing degenerate results. To address local minima, the authors utilize a reproducibility strategy, running the algorithm multiple times and selecting the most stable result via a similarity pseudo-metric.

Experimental Setup

The study utilized the MCIC collection, a neuroimaging dataset comprising 271 subjects (121 diagnosed with schizophrenia and 150 healthy controls). The analysis focused on two primary tasks:

  1. Auditory Oddball (AOD): Listening for specific target tones.
  2. Sternberg Item Recognition Paradigm (SIRP): A working memory task, analyzed during both Encoding (E) and Probe (P) phases.

Raw fMRI data were converted into regression coefficient maps (feature maps) to isolate task-related activations for the decomposition.

Key Results and Discussion

The proposed method demonstrated significant clinical and computational advantages. By forcing the data to explain itself through either a shared or task-specific lens, the model discovered biomarkers without the need for secondary sorting algorithms.

Clinical Discrimination and Shared Components

The method successfully identified a shared component involving the DMN and motor regions. This component showed highly significant differences ($p < 0.05$) between schizophrenia patients and healthy controls, suggesting that the fusion of tasks with similar demands (like AOD and SIRP-P) allows tasks to "inform" each other, improving the accuracy of motor component estimation.

Shared brain components identified by the model across tasks, showing highly significant clinical discrimination (p = 1.47e-6).

Task Specificity

In the AOD task, the distinct auditory component was identified as the most discriminative feature for identifying patients. The spatial maps produced were notably "clean," showing clear activations in expected functional networks such as the visual and motor systems.

Distinct components for the Auditory Oddball (AOD) task, highlighting task-specific biomarkers with significant p-values.

Efficiency and Limitations

The entire decomposition for the three-task fusion was completed in approximately 157.5 seconds on a standard quad-core computer, proving its practical utility for large datasets. However, the researchers noted that users must still manually select the "rank" (number of components). Additionally, while the results were comparable to IVA, the p-values were occasionally less extreme, suggesting a trade-off between direct interpretability and statistical sensitivity.

Conclusion

This partially constrained Coupled CP decomposition provides a robust, single-step solution for multi-task fMRI fusion. By grounding the mathematical model in the biological reality of shared and distinct brain networks, the framework offers a more direct path to identifying interpretable biomarkers for neurological disorders like schizophrenia. Future work may explore "soft couplings" to further account for complex data variability between tasks.

Summary updated

Abstract

Discovering components that are shared in multiple datasets, next to dataset-specific features, has great potential for studying the relationships between different subjects or tasks in functional Magnetic Resonance Imaging (fMRI) data. Coupled matrix and tensor factorization approaches have been useful for flexible data fusion, or decomposition to extract features that can be used in multiple ways. However, existing methods do not directly recover shared and dataset-specific components, which requires post-processing steps involving additional hyperparameter selection. In this paper, we propose a tensor-based framework for multi-task fMRI data fusion, using a partially constrained canonical polyadic (CP) decomposition model. Differently from previous approaches, the proposed method directly recovers shared and dataset-specific components, leading to results that are directly interpretable. A strategy to select a highly reproducible solution to the decomposition is also proposed. We evaluate the proposed methodology on real fMRI data of three tasks, and show that the proposed method finds meaningful components that clearly identify group differences between patients with schizophrenia and healthy controls.
BibTeX
@inproceedings{borsoi2023coupled,
  author = {R. A. Borsoi and Isabell Lehmann and Mohammad A. B. S. Akhonda and Vince D. Calhoun and Konstantin Usevich and David Brie and Tülay Adali},
  title = {{Coupled CP Tensor Decomposition with Shared and Distinct Components for Multi-Task Fmri Data Fusion}},
  booktitle = {IEEE International Conference on Acoustics, Speech and Signal Processing},
  year = {2023},
  doi = {10.1109/icassp49357.2023.10096241},
}
Figure