Low-Rank Tucker-2 Model for Multi-Subject fMRI Data Decomposition With Spatial Sparsity Constraint

Yue Han, Qiu-Hua Lin, Li-Dan Kuang, Xiao-Feng Gong, Fengyu Cong, Yu-Ping Wang, Vince D. CalhounIEEE TMI 2022 · DOI

TL;DR

This paper introduces a novel low-rank Tucker-2 model with a spatial sparsity constraint (slcTKD) to analyze multi-subject fMRI data. By incorporating spatial sparsity, the method effectively extracts shared spatial and temporal patterns, leading to improved performance in identifying brain activity and classifying subjects compared to existing tensor decomposition techniques.

What Problem Does This Paper Address?

Functional Magnetic Resonance Imaging (fMRI) generates complex, high-dimensional data that requires sophisticated methods for analysis. Multi-subject fMRI data, in particular, presents challenges due to inherent noise and distinct characteristics across individuals. Traditional tensor decomposition methods like Canonical Polyadic Decomposition (CPD) and Tucker Decomposition (TKD) have been used to extract meaningful components (spatial maps, time courses, subject-specific intensities) from fMRI data. However, these methods often struggle to capture common spatial and temporal patterns across subjects due to noise and the assumption of orthogonality, which may not hold true for fMRI data. The authors aim to address this by proposing a Tucker-2 model that explicitly accounts for the spatial sparsity of fMRI activations and imposes low-rank constraints on factor matrices, leading to more robust and informative decomposition.

Methods

The core of this paper is the proposed slcTKD (sparsity-low-rank-constraints Tucker-2) model. It builds upon the Tucker-2 decomposition framework, which represents fMRI data as a product of a core tensor and factor matrices. The key innovations are:

  • Spatial Sparsity Constraint on Spatial Maps (SMs): Recognizing that fMRI activations are inherently sparse (only a small percentage of neurons are active at any given time), the authors impose an $\ell_p$ norm constraint ($0 < p \le 1$) on the spatial maps. This encourages the model to identify the most relevant active regions.
  • Low-Rank Constraints on Factor Matrices: The factor matrices representing shared spatial maps (S) and time courses (TCs) are constrained to be low-rank using the Frobenius norm. This helps in extracting principal components shared across subjects.
  • Sparsity Constraint on Core and Residual Tensors: Sparsity constraints are also applied to the core tensor (G) and the residual tensor (E) using the $\ell_1$ norm to improve model uniqueness and performance.
  • Optimization Algorithm: The model is solved using an alternating direction method of multipliers (ADMM) and half quadratic splitting (HQS) for efficient computation.
  • Feature Extraction: Beyond standard decomposition, the authors extract novel spatial and temporal features from the core tensor, which are then used for subject classification.

Datasets

The authors evaluate their proposed method using two types of datasets:

  • Simulated fMRI Data: Generated to mimic real fMRI data, including eight simulated fMRI-like sources with varying noise levels (SNR from -10 dB to 10 dB). This allows for controlled testing of the algorithm's performance under different conditions.
  • Experimental fMRI Data: A real fMRI dataset collected from 10 subjects performing a finger-tapping motor task while receiving auditory instructions. The data was preprocessed using standard neuroimaging software.

Key Insights and Ideas

  • Sparsity is Crucial for fMRI: The paper strongly emphasizes that fMRI spatial maps are significantly sparser than typical images, and enforcing this sparsity is key to successful decomposition, especially in noisy data.
  • Beyond Orthogonality: Traditional TKD methods often rely on orthogonality constraints, which are not always suitable for fMRI. The proposed sparsity and low-rank constraints offer a more flexible and effective approach.
  • Core Tensor as a Rich Information Source: The core tensor, often seen as a bottleneck in other methods, is highlighted as a rich source of information, containing not just subject-specific intensities but also novel spatial and temporal features.
  • Intra-group Differences Matter: The extracted spatial and temporal features can capture subtle differences within groups of subjects, which is crucial for understanding individual variability in brain function.

Key Results

  • Superior Performance: The slcTKD method consistently outperformed other compared algorithms (RKCA, TKD, CPD, BTD) in terms of Pearson correlation coefficients (|pc|) for both simulated and experimental fMRI data, especially under lower SNR conditions.
  • Improved Spatial Map Estimation: The proposed method extracted a larger number of activated voxels within the spatial reference for both task-related and DMN components in experimental data, indicating more accurate and contiguous spatial maps.
  • Effective Subject Classification: The extracted spatial and temporal features from the core tensor were successfully used to cluster subjects into distinct subgroups, revealing meaningful differences in task-related activations and temporal responses.
  • Robustness to Noise: The slcTKD algorithm demonstrated better performance and less degradation compared to other methods as noise levels increased.
  • Model Order Independence: The proposed method showed more consistent performance across different model orders compared to other algorithms.

Limitations and Future Work

  • Parameter Tuning: While the paper provides guidance on parameter selection, fine-tuning these parameters (especially for new datasets) might still be necessary.
  • Computational Complexity: Although ADMM and HQS are used, the computational complexity can still be significant for very large datasets.
  • Future Directions: The authors suggest several avenues for future research:
  • Improving the Tucker-2 model to extract individual time delays.
  • Extending the method to complex-valued fMRI data.
  • Developing adaptive sparsity parameter selection methods.
  • Applying the method to resting-state fMRI data or for analyzing patients with neurological disorders.

Summary updated

BibTeX
@article{han2022low,
  author = {Yue Han and Qiu-Hua Lin and Li-Dan Kuang and Xiao-Feng Gong and Fengyu Cong and Yu-Ping Wang and Vince D. Calhoun},
  title = {{Low-Rank Tucker-2 Model for Multi-Subject fMRI Data Decomposition With Spatial Sparsity Constraint}},
  journal = {IEEE Trans. Medical Imaging},
  year = {2022},
  doi = {10.1109/tmi.2021.3122226},
}
Figure