EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

Enabling Sparse CCA for Mix Source Separation in High-Dimensional Data

Muhammad Usman Khalid

IEEE Access · 2026 · 14 · 37795-37812

DOI: 10.1109/access.2026.3672074 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Muhammad Usman Khalid. Die Arbeit nennt 69 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: Deanship of Scientific Research at Imam Mohammad Ibn Saud Islamic University.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Sparse canonical correlation analysis is a powerful method for capturing the relationships between two multidimensional datasets and can be extended for mix source separation (MSS) by leveraging the autocorrelation matrix. However, its application to whole-brain functional magnetic resonance imaging (fMRI) datasets is challenging due to high-dimensional nature of the cross-correlation matrix, loss of spatial information, multicollinearity, and low signal-to-noise ratio. To overcome these limitations, two novel dissociative canonical correlation analysis methods named DCCAPMD and DCCAPCA are introduced in this paper. The DCCAPMD algorithm employs penalized matrix decomposition and optimizes at least one dimension of the covariance matrix by utilizing divide and conquer clustering. In contrast, DCCAPCA employs dissociative principal component analysis to eliminate reliance on the covariance matrix, focusing entirely on the error matrix. Both algorithms preserve the spatial structure using box filtering, mitigate multicollinearity by eliminating redundant information through a mechanism named sparse contribution based component pruning, and amplify signal power by deploying singular spectrum analysis. The primary advantage of these methods is their ability to manage or circumvent large cross-correlation matrices, rendering the computations tractable for whole-brain fMRI data, while also enhancing source recovery, signal strength, and the sensitivity of spatial maps, making them ideal for MSS. At the individual level, these methods accurately capture within-subject autocorrelations, thereby facilitating the extraction of intrinsic and functional temporal components and their corresponding networks and activation maps from fMRI data. The performance of the proposed algorithms is validated across three distinct types of fMRI datasets, including synthetic, block design, and event related. The proposed algorithms significantly outperformed existing state-of-the-art methods, including PCA+PMD, SPCA+PMD, ICA+PMD, and SICA+PMD. The proposed methods achieved a substantial memory reduction of approximately 99% for the covariance matrices of real fMRI datasets (e.g., from $236115 \times 236115$ to $236115 \times 100$ ) and a substantial decrease for the error matrices (e.g., from $236115 \times 236115$ to $284 \times 236115$ ), while simultaneously enhancing performance, as demonstrated by a overall correlation value increase of approximately 16% compared to SICA+PMD. The source code for all methods (including the proposed) used in this study is publicly available at the following GitHub repository https://github.com/usmankhalid06/DCCA

Förderung

  • Deanship of Scientific Research at Imam Mohammad Ibn Saud Islamic University

Zitieren

APA

Khalid, M. U. (2026). Enabling Sparse CCA for Mix Source Separation in High-Dimensional Data. IEEE Access, 14, 37795-37812. https://doi.org/10.1109/access.2026.3672074

Vancouver

Khalid MU. Enabling Sparse CCA for Mix Source Separation in High-Dimensional Data. IEEE Access. 2026;14:37795-37812. doi:10.1109/access.2026.3672074

BibTeX
@article{Khalid2026,
  author = {Khalid, Muhammad Usman},
  title = {Enabling Sparse CCA for Mix Source Separation in High-Dimensional Data},
  journal = {IEEE Access},
  year = {2026},
  volume = {14},
  pages = {37795--37812},
  doi = {10.1109/access.2026.3672074}
}
Kostenlos starten →