Independent Vector Analysis with Multivariate Gaussian Model: a Scalable Method by Multilinear Regression

Ben Gabrielson, Mingyu Sun, Mohammad A. B. S. Akhonda, Vince D. Calhoun, Tülay AdaliICASSP 2023 · DOI

Abstract

Joint Blind Source Separation (JBSS) is a critical technique for unmixing overlapping brain signals across multiple subjects in fMRI studies. While Independent Vector Analysis (IVA) effectively captures individual subject variability, its computational complexity makes it intractable for large-scale datasets. Conversely, the industry-standard Group-ICA scales well but often obscures subject-specific nuances. This paper introduces regIVA-G, a scalable approach that leverages multilinear regression to maintain the detail of IVA with the efficiency of Group-ICA. By decoupling the optimization of additional datasets from a small base model, regIVA-G achieves linear computational growth. Our results demonstrate that regIVA-G provides a robust, high-performance solution for large-scale neuroimaging analysis, running orders of magnitude faster than traditional IVA-G.

Introduction

In medical imaging, researchers frequently face the "cocktail party problem": isolating distinct neural sources from raw data collected across hundreds of subjects. This task, known as Joint Blind Source Separation (JBSS), involves a fundamental trade-off between accuracy and scalability. Independent Vector Analysis (IVA) is highly accurate because it models the unique dependencies within each subject's data, but its computational cost grows quadratically or exponentially with the number of subjects. Group-ICA, while fast, assumes a common spatial map for all subjects, which can "blur" important individual differences.

To bridge this gap, we propose IVA-G regression (regIVA-G), a method designed to handle hundreds or thousands of datasets with ease while preserving subject-specific variability.

Methodology

The regIVA-G algorithm operates in two distinct stages:

  1. Base Model Estimation: A small, representative subset of subjects is processed using standard IVA-G. This produces a set of high-quality source templates, or "regressors."
  2. Multilinear Regression: For all remaining subjects, the algorithm performs a specialized regression. It identifies sources in the new data that maximize correlation with the base templates while minimizing cross-source interference.

Mathematically, this process is reduced to finding the principal eigenvector of a quadratic form for each subject. Because each subject is processed independently against the base model, the total execution time scales linearly ($O(K)$) with the number of datasets.

Performance and Scalability

We evaluated regIVA-G using both simulated environments and hybrid fMRI data from the MCIC collection. A primary advantage of regIVA-G is its computational efficiency. In simulations with up to 200 datasets, standard IVA-G exhibited a sharp spike in processing time, whereas regIVA-G maintained a near-linear trajectory similar to Group-ICA.

Figure 2: Execution time in minutes vs. the number of datasets (K). Note the exponential growth of traditional IVA-G (yellow) compared to the scalability of regIVA-G (green).

In hybrid fMRI tests, regIVA-G achieved a separation score (Joint-ISI) of 0.0073 in just 1.7 minutes. In contrast, traditional IVA-G took 23.7 minutes—nearly 14 times longer—to achieve a comparable level of accuracy.

Accuracy and Robustness

Accuracy was measured using the Joint Intersymbol Interference (Joint-ISI) metric across various levels of subject variability ($R_n$). Group-ICA typically hits a performance ceiling when subjects are highly diverse. In contrast, regIVA-G remains consistently accurate even as variability increases.

Figure 3: Log(joint-ISI) across different simulation scenarios. regIVA-G (green) consistently outperforms Group-ICA (purple/blue) and remains stable where traditional IVA-G (yellow) may degrade.

Notably, the "base model" does not need to be large to be effective. Even when the base model contains fewer datasets than the rank of the variability, it provides sufficient information to guide the regression accurately, allowing regIVA-G to break the performance limitations of standard group-based methods.

Conclusion

regIVA-G offers a powerful middle ground for large-scale fMRI studies, providing the subject-specific detail of IVA with the speed of Group-ICA. While the current model assumes linear dependencies and uses a one-way learning flow from the base model, it sets the stage for future work in non-linear brain dynamics and iterative model refinement. By making high-fidelity JBSS computationally feasible for thousands of subjects, regIVA-G enables more nuanced population-level neuroimaging research.

Summary updated

BibTeX
@inproceedings{gabrielson2023independent,
  author = {Ben Gabrielson and Mingyu Sun and Mohammad A. B. S. Akhonda and Vince D. Calhoun and Tülay Adali},
  title = {{Independent Vector Analysis with Multivariate Gaussian Model: a Scalable Method by Multilinear Regression}},
  booktitle = {IEEE International Conference on Acoustics, Speech and Signal Processing},
  year = {2023},
  doi = {10.1109/icassp49357.2023.10096698},
}

Loading related content...

Figure