Shift-Invariant Canonical Polyadic Decomposition of Complex-Valued Multi-Subject fMRI Data With a Phase Sparsity Constraint

Li‐Dan Kuang, Qiu‐Hua Lin, Xiao‐Feng Gong, Fengyu Cong, Yu‐Ping Wang, Vince D. Calhoun

Introduction

Functional Magnetic Resonance Imaging (fMRI) is a cornerstone of modern neuroscience, providing a window into the brain's dynamic activity. However, analyzing multi-subject fMRI data is fraught with challenges. Traditional methods often discard the phase information of the complex-valued signal, focusing only on magnitude. Furthermore, standard group analysis techniques like Canonical Polyadic Decomposition (CPD) often fail to account for the inherent variability in hemodynamic delays and spatial activation patterns across different individuals.

To address these limitations, researchers have developed pcsCPD (Phase Sparsity Constrained Shift-Invariant CPD), a novel tensor decomposition framework designed to leverage the full complex-valued signal while robustly handling subject-specific timing differences and noise.

The pcsCPD Framework

The proposed method introduces three critical innovations to the standard CPD model:

  1. Shift-Invariance: Recognizing that neural responses are not perfectly synchronized across subjects, pcsCPD estimates subject-specific time delays in the frequency domain. This allows the model to "align" temporal components across the group, preventing the "smearing" of signals that occurs in rigid models.
  2. Phase De-ambiguity: Because complex-valued signals can have arbitrary phase rotations, the algorithm includes a rotation step to ensure that the spatial source phase remains consistent and interpretable across all components.
  3. Phase Sparsity Constraint: This is the method's core denoising mechanism. In BOLD (Blood-Oxygen-Level-Dependent) imaging, voxels associated with actual brain activity typically exhibit very small phase changes. In contrast, noise and physiological artifacts often manifest as large phase values. By imposing a smoothed $\ell_0$ norm constraint, pcsCPD penalizes voxels with large phases, effectively "pruning" noise from the shared spatial maps.

Performance Evaluation

The researchers validated pcsCPD using both synthetic simulations and real-world experimental data.

Simulation Results

In simulations involving 10 subjects and 8 distinct signal sources, pcsCPD demonstrated remarkable resilience to noise and temporal shifts. As shown in Figure 4, the proposed method consistently achieved higher correlation with ground-truth spatial maps (SM) and time courses (TC) compared to standard CPD, shift-invariant CPD (csCPD), and tensor-based spatial ICA (T-sICA).

Performance comparison of pcsCPD against competing methods across various Signal-to-Noise Ratios (SNR) and temporal/spatial shifts.

The robustness of the method is particularly evident at low Signal-to-Noise Ratios. Figure 6 illustrates that while the performance of other complex-valued algorithms degrades significantly as SNR drops below 0 dB, pcsCPD maintains a high correlation with the underlying neural signal.

Correlation of estimated time courses with ground truth as a function of SNR, highlighting the superior noise robustness of the pcsCPD method.

Experimental Validation

When applied to real-world fMRI data from a finger-tapping motor task, pcsCPD outperformed traditional real-valued approaches. Specifically, it detected 178.7% more contiguous task-related activations than a hybrid method combining shift-invariant CPD and ICA. Qualitatively, the resulting spatial maps were cleaner and more biologically plausible, accurately localizing activity in the primary and supplementary motor areas without the "speckled" noise characteristic of less constrained methods.

Conclusion

By integrating shift-invariance with a physically-motivated phase sparsity constraint, pcsCPD offers a powerful new tool for multi-subject fMRI analysis. It successfully bridges the gap between the high information content of complex-valued data and the practical need for noise robustness and subject alignment. Future work will explore the application of this framework to resting-state fMRI, where the absence of external stimuli presents an even greater challenge for temporal alignment.

Summary updated

Abstract

Canonical polyadic decomposition (CPD) of multi-subject complex-valued fMRI data can be used to provide spatially and temporally shared components among groups with both magnitude and phase information. However, the CPD model is not well formulated due to the large subject variability in the spatial and temporal modalities, as well as the high noise level in complex-valued fMRI data. Considering that the shift-invariant CPD can model temporal variability across subjects, we propose to further impose a phase sparsity constraint on the shared spatial maps to denoise the complex-valued components and to model the inter-subject spatial variability as well. More precisely, subject-specific time delays are first estimated for the complex-valued shared time courses in the framework of real-valued shift-invariant CPD. Source phase sparsity is then imposed on the complex-valued shared spatial maps. A smoothed l<sub>0</sub> norm is specifically used to reduce voxels with large phase values after phase de-ambiguity based on the small phase characteristic of BOLD-related voxels. The results from both the simulated and experimental fMRI data demonstrate improvements of the proposed method over three complex-valued algorithms, namely, tensor-based spatial ICA, shift-invariant CPD and CPD without spatiotemporal constraints. When comparing with a real-valued algorithm combining shift-invariant CPD and ICA, the proposed method detects 178.7% more contiguous task-related activations.
BibTeX
@article{kuang2019shift,
  author = {Li‐Dan Kuang and Qiu‐Hua Lin and Xiao‐Feng Gong and Fengyu Cong and Yu‐Ping Wang and Vince D. Calhoun},
  title = {{Shift-Invariant Canonical Polyadic Decomposition of Complex-Valued Multi-Subject fMRI Data With a Phase Sparsity Constraint}},
  journal = {IEEE Transactions on Medical Imaging},
  year = {2019},
  doi = {10.1109/tmi.2019.2936046},
}
Figure