Multidimensional Comparisons Between Constrained ICA/IVA Algorithms for Multi-Subject fMRI Data Analysis
Lucas Gois1, Hanlu Yang2, Trung Vu2
1Center for Engineering, Modeling and Applied Social Sciences, Federal University of ABC, Santo André 09280-560, Brazil.
IEEE Access : Practical Innovations, Open Solutions
|March 23, 2026
Summary
This study compares three constrained algorithms for analyzing functional magnetic resonance imaging (fMRI) data to identify brain networks. Results show each method has unique strengths for reproducibility, biomarker detection, and scalability, guiding algorithm selection for fMRI research.
Area of Science:
- Neuroimaging
- Computational Neuroscience
- Data Science
Background:
- Large-scale functional magnetic resonance imaging (fMRI) datasets are crucial for brain health research.
- Independent Component Analysis (ICA) and Independent Vector Analysis (IVA) are data-driven techniques for analyzing multi-subject fMRI data to extract functional connectivity networks.
- Constrained versions of ICA and IVA improve performance and interpretability, but their comparative advantages are unclear.
Purpose of the Study:
- To comprehensively compare three state-of-the-art constrained algorithms: threshold-free constrained IVA (tf-cIVA), adaptive-reverse constrained IVA (ar-cIVA), and adaptive-reverse constrained ICA (ar-cEBM).
- To evaluate their performance using metrics like reproducibility, scalability, alignment with references, connectivity, and consistency on a large multi-site fMRI dataset.
- To provide practical guidance on algorithm selection based on specific research questions in fMRI analysis.
Main Methods:
- A multidimensional comparison of tf-cIVA, ar-cIVA, and ar-cEBM was conducted.
- Methods were evaluated on a multi-site fMRI dataset comprising 429 subjects.
- Key metrics included reproducibility, scalability, alignment with references, connectivity, and consistency.
Main Results:
- All three methods demonstrated replicability in spatial correlation with references and biomarker identification.
- tf-cIVA excelled in reproducibility and produced structured temporal functional network connectivity (FNC), suitable for dynamic analyses.
- ar-cIVA showed high sensitivity to group differences in spatial FNC for biomarker detection, while ar-cEBM offered superior computational scalability and stable spatial maps.
Conclusions:
- The choice of constrained algorithm (tf-cIVA, ar-cIVA, ar-cEBM) for fMRI analysis depends on the specific research goals, balancing reproducibility, biomarker sensitivity, and computational scalability.
- ar-cEBM's subject-wise approach yielded stable spatial maps, suggesting flexible density matching is critical for group consistency.
- This study provides essential practical guidance for researchers utilizing large-scale fMRI datasets and advanced analytical techniques.
