Memory Efficient PCA Methods for Large Group ICA.

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Data Science

Background:

• Principal Component Analysis (PCA) is crucial for data reduction in group Independent Component Analysis (ICA) of fMRI data.
• Current methods often involve computing group-level PCA on temporally concatenated datasets before ICA.
• Handling very high-dimensional fMRI datasets poses significant memory and computational challenges.

Purpose of the Study:

• To develop an efficient method for reducing very high-dimensional, temporally concatenated fMRI datasets into their group PCA space.
• To optimize existing randomized PCA methods for minimal dataloads and memory requirements.
• To enable scalable group-level PCA for large fMRI datasets.

Main Methods:

• Extended existing randomized PCA methods to create Multi Power Iteration (MPOWIT) for efficient PCA computation with minimal dataloads.
• MPOWIT estimates a larger subspace while monitoring convergence of a smaller subset, reducing iterations and dataloads.
• Introduced efficient subsampled eigenvalue decomposition for PCA subspace approximation and initialization of randomized methods.

Main Results:

• MPOWIT significantly reduces the number of iterations and dataloads, accelerating convergence without sacrificing accuracy.
• The memory required for MPOWIT is independent of the number of subjects, enabling analysis of large cohorts.
• Successfully performed a 1600-subject group-level PCA of fMRI on a desktop computer with 4 GB RAM in a few hours.

Conclusions:

• MPOWIT offers a highly scalable and memory-efficient solution for group-level PCA of fMRI data, even for thousands of subjects.
• The method is highly parallelizable, facilitating fast, distributed implementations for big data analysis.
• MPOWIT and related techniques are available in the open-source GIFT software, providing practical tools for neuroimaging research.