Constrained CPD of Complex-Valued Multi-Subject fMRI Data via Alternating Rank-R and Rank-1 Least Squares

Li‐Dan Kuang, Qiu‐Hua Lin, Xiao‐Feng Gong, Jianming Zhang, Wenjun Li, Feng Li, Vince D. CalhounIEEE Transactions on Neural Systems and Rehabilitation Engineering 2022 · DOI

TL;DR

This paper introduces a novel method, SARRR1LS-PO, for analyzing complex-valued fMRI data that improves upon existing techniques by better preserving information often lost in traditional filtering. It achieves this by using an alternating rank-R and rank-1 least squares optimization, incorporating spatial phase sparsity and orthonormality constraints, leading to more accurate identification of brain networks.

What Problem Does This Paper Address?

Functional Magnetic Resonance Imaging (fMRI) data is inherently noisy and exhibits significant variability across individuals and over time. Analyzing this data to understand brain function requires robust decomposition techniques. While complex-valued fMRI data offers richer information than magnitude-only data, existing methods like Canonical Polyadic Decomposition (CPD) struggle with this complexity. Previous approaches, such as pcsCPD, have used band-pass filtering to suppress noise, but this often leads to the loss of valuable information. This paper aims to address this by developing a method that can effectively decompose complex-valued fMRI data while minimizing information loss and handling noise and inter-subject variability.

Methods

The core of this work is a novel optimization strategy called Alternating Rank-R and Rank-1 Least Squares (ARRR1LS). This method relaxes the strict assumptions of traditional CPD by iteratively updating different components of the fMRI data. Specifically, it involves:

  1. Rank-R Least Squares for Shared Spatial Maps (SMs): Updates SMs while considering spatial phase sparsity.
  2. Orthonormality Constraint: Minimizes crosstalk between shared SMs.
  3. Rank-R Least Squares for Aggregating Mixing Matrix: Updates the matrix that combines information from shared time courses and subject intensities.
  4. Shift-Invariant Rank-1 Least Squares: Updates shared time courses, subject-specific time delays, and intensities.

Building on ARRR1LS, the authors propose SARRR1LS-PO, which further incorporates:

  • Shift-Invariance: Accounts for temporal delays across subjects.
  • Spatial Phase Sparsity: Leverages the characteristic phase behavior of BOLD-related voxels to improve separation.
  • Spatial Orthonormality: Reduces correlation and crosstalk between SM components.

This combination aims to comprehensively address the noise and spatiotemporal variability in complex-valued multi-subject fMRI data.

Datasets

The paper evaluates its methods using:

  • Simulated Complex-Valued fMRI Data: Extended from existing real-valued simulated datasets, modified to incorporate complex-valued properties and realistic noise levels.
  • Experimental Complex-Valued fMRI Data: Collected from 16 subjects performing a finger-tapping task, with data also analyzed after applying an 8-150 mHz band-pass filter (as used in prior work).

Key Insights and Ideas

  • The Value of Complex-Valued Data: The study strongly suggests that complex-valued fMRI data contains crucial information beyond magnitude, particularly in its phase component, which is often discarded or poorly utilized by traditional methods.
  • Filtering is a Double-Edged Sword: While filtering reduces noise, it can also remove meaningful neural signals. The proposed method aims to achieve noise reduction without extensive filtering, preserving more of the original signal.
  • Phase Sparsity is Key: The unique phase characteristics of BOLD signals in activated voxels can be exploited as a powerful constraint for more accurate component separation.
  • Combining Constraints is Synergistic: The combination of shift-invariance, spatial phase sparsity, and spatial orthonormality constraints leads to superior performance compared to using any single constraint or less sophisticated methods.

Key Results

  • Simulated Data: The proposed SARRR1LS-PO consistently outperformed T-sICA and pcsCPD across various Signal-to-Noise Ratio (SNR) levels, showing higher average correlation coefficients (p-values) for both magnitude and phase of shared SMs and TCs. The performance gap was particularly pronounced for shared TCs.
  • Experimental Data:
    • In both raw and filtered fMRI data, SARRR1LS-PO achieved the highest average p-values for sensorimotor shared SMs and TCs, indicating better separation accuracy.
  • The performance advantage of SARRR1LS-PO was more evident in raw fMRI data compared to filtered data, highlighting its robustness to noise and its ability to extract information lost by filtering.
  • SARRR1LS-PO demonstrated a higher number of correctly estimated time delays compared to pcsCPD.
  • A comparison with real-valued ICA-sCPD showed that SARRR1LS-PO extracted approximately 30.7% more activated voxels, underscoring the benefit of complex-valued analysis.

Limitations and Future Work

  • Parameter Tuning: While the paper mentions setting parameters like $\sigma_{th}$ and $\lambda$, the optimal selection process and sensitivity analysis could be further explored.
  • Computational Complexity: Although ARRR1LS is noted as being faster than classical ALS, the computational cost of SARRR1LS-PO, especially with a large number of components or subjects, might still be a consideration.
  • Subject-Specific Components: The current work focuses on shared components. The authors suggest future work to extend the method to extract subject-specific components, which are often crucial in fMRI analysis.
  • Application to Other fMRI Paradigms: The proposed method was validated on task-based fMRI data. Future work could explore its application to resting-state fMRI data for classifying healthy controls and patients with brain diseases.
  • Accurate Phase De-ambiguity: The paper proposes an improved phase de-ambiguity method, but further validation and comparison with other methods could be beneficial.

Summary updated

Abstract

Complex-valued shift-invariant canonical polyadic decomposition (CPD) under a spatial phase sparsity constraint (pcsCPD) shows excellent separation performance when applied to band-pass filtered complex-valued multi-subject fMRI data. However, some useful information may also be eliminated when using a band-pass filter to suppress unwanted noise. As such, we propose an alternating rank- R and rank-1 least squares optimization to relax the CPD model. Based upon this optimization method, we present a novel constrained CPD algorithm with temporal shift-invariance and spatial sparsity and orthonormality constraints. More specifically, four steps are conducted until convergence for each iteration of the proposed algorithm: 1) use rank- R least-squares fit under spatial phase sparsity constraint to update shared spatial maps after phase de-ambiguity; 2) use orthonormality constraint to minimize the cross-talk between shared spatial maps; 3) update the aggregating mixing matrix using rank- R least-squares fit; 4) utilize shift-invariant rank-1 least-squares on a series of rank-1 matrices reconstructed by each column of the aggregating mixing matrix to update shared time courses, and subject-specific time delays and intensities. The experimental results of simulated and actual complex-valued fMRI data show that the proposed algorithm improves the estimates for task-related sensorimotor and auditory networks, compared to pcsCPD and tensorial spatial ICA. The proposed alternating rank- R and rank-1 least squares optimization is also flexible to improve CPD-related algorithm using alternating least squares.
BibTeX
@article{kuang2022constrained,
  author = {Li‐Dan Kuang and Qiu‐Hua Lin and Xiao‐Feng Gong and Jianming Zhang and Wenjun Li and Feng Li and Vince D. Calhoun},
  title = {{Constrained CPD of Complex-Valued Multi-Subject fMRI Data via Alternating Rank-<i>R</i> and Rank-1 Least Squares}},
  journal = {IEEE Transactions on Neural Systems and Rehabilitation Engineering},
  year = {2022},
  doi = {10.1109/tnsre.2022.3198679},
}
Figure