A telescopic independent component analysis on functional magnetic resonance imaging dataset

Shiva Mirzaeian, Ashkan Faghiri, Vince D. Calhoun, Armin Iraji

Abstract

Brain function can be modeled as dynamic interactions between functional sources at different spatial scales, and each spatial scale can contain its functional sources with unique information, thus using a single scale may provide an incomplete view of brain function. This paper introduces a novel approach, termed "telescopic independent component analysis (TICA)," designed to construct spatial functional hierarchies and estimate functional sources across multiple spatial scales using fMRI data. The method employs a recursive independent component analysis (ICA) strategy, leveraging information from a larger network to guide the extraction of information about smaller networks. We apply our model to the default mode network (DMN), visual network (VN), and right frontoparietal network (RFPN). We investigate further on the DMN by evaluating the difference between healthy people and individuals with schizophrenia. We show that the TICA approach can detect the spatial hierarchy of the DMN, VN, and RFPN. In addition, the TICA revealed DMN-associated group differences between cohorts that may not be captured if we focus on a single-scale ICA. In sum, our proposed approach represents a promising new tool for studying functional sources.

Introduction

Functional Magnetic Resonance Imaging (fMRI) is a cornerstone of modern neuroscience, allowing researchers to observe brain activity by tracking blood oxygenation levels. A common analytical approach is Independent Component Analysis (ICA), which identifies Intrinsic Connectivity Networks (ICNs)—spatially distinct regions that exhibit synchronized activity. However, the brain's functional architecture is inherently hierarchical. Large-scale networks, such as the Default Mode Network (DMN), are composed of smaller, specialized sub-units.

Traditional ICA methods often require a fixed "model order," forcing a choice between broad, low-resolution networks and fragmented, high-resolution components. This approach fails to capture the mathematical relationship between scales, potentially obscuring pathological changes that only manifest at specific levels of the functional hierarchy. This paper introduces Telescopic Independent Component Analysis (TICA), a recursive method designed to bridge these scales and provide a more comprehensive view of brain function.

Methodology: The Telescopic Approach

TICA treats the discovery of brain networks as a multi-step, recursive process that uses large-scale domains as a "lens" to focus on finer sub-units.

  1. Scale 1 (Global Domain Identification): A standard group-level ICA is performed at a low model order (e.g., 20) to identify broad functional domains like the DMN, Visual Network (VN), or Right Frontoparietal Network (RFPN).
  2. Recursive Weighting: The spatial map generated at Scale 1 is used as a voxel-wise weight for the original BOLD signal. By multiplying the signal by this mask, the algorithm effectively filters out "noise" from the rest of the brain, focusing exclusively on the activity within the target network.
  3. Scale 2 (Sub-network Extraction): A second ICA is performed on this weighted signal at a higher model order. This "zoomed-in" view allows for the precise identification of sub-components that are mathematically tied to their "parent" network.

The study validated this approach using data from the Human Connectome Project (HCP) for reliability testing (1,019 subjects) and the Functional Imaging Biomedical Informatics Research Network (FBIRN) for clinical application in schizophrenia (310 individuals).

Results and Key Insights

The TICA method demonstrated several advantages over traditional single-scale ICA:

Enhanced Precision and Reliability

Contrary to the expectation that finer scales would introduce more noise, Scale 2 sub-networks showed higher subject-level similarity and test-retest reliability compared to Scale 1. Pearson correlation coefficients for Scale 2 components ranged from 0.80 to 0.91. The recursive weighting step appears to improve the signal-to-noise ratio for individual features by isolating the relevant functional context.

Clinical Sensitivity in Schizophrenia

The most significant finding involved the DMN in individuals with schizophrenia. While Scale 1 analysis showed some differences, TICA’s Scale 2 revealed much more pronounced diagnostic effects. Specific regions, including the superior frontal gyrus and medial orbital cortex, exhibited significant group differences in the "zoomed" view that were entirely absent in standard wide-view ICA.

Structural Consistency

TICA successfully decomposed the DMN, VN, and RFPN into replicable sub-components. Group difference maps remained highly stable, maintaining a similarity of >0.90 across multiple iterations of the dataset.

Discussion and Future Directions

TICA provides a mathematical bridge between functional scales, preserving the relationship between a network fragment and the whole system. This context is crucial for understanding how localized changes contribute to systemic brain dysfunction.

While this study focused on two scales, the "telescope" framework can theoretically be extended to three or more levels for even higher resolution. Future work should also explore dynamic functional network connectivity (dFNC) within this hierarchical framework to observe how these multiscale interactions evolve over time. Ultimately, TICA represents a promising tool for creating a comprehensive multiscale atlas of both healthy and pathological brain states.

Summary updated

Abstract

Brain function can be modeled as dynamic interactions between functional sources at different spatial scales, and each spatial scale can contain its functional sources with unique information, thus using a single scale may provide an incomplete view of brain function. This paper introduces a novel approach, termed "telescopic independent component analysis (TICA)," designed to construct spatial functional hierarchies and estimate functional sources across multiple spatial scales using fMRI data. The method employs a recursive independent component analysis (ICA) strategy, leveraging information from a larger network to guide the extraction of information about smaller networks. We apply our model to the default mode network (DMN), visual network (VN), and right frontoparietal network (RFPN). We investigate further on the DMN by evaluating the difference between healthy people and individuals with schizophrenia. We show that the TICA approach can detect the spatial hierarchy of the DMN, VN, and RFPN. In addition, the TICA revealed DMN-associated group differences between cohorts that may not be captured if we focus on a single-scale ICA. In sum, our proposed approach represents a promising new tool for studying functional sources.
BibTeX
@inproceedings{mirzaeian2024telescopic,
  author = {Shiva Mirzaeian and Ashkan Faghiri and Vince D. Calhoun and Armin Iraji},
  title = {{A telescopic independent component analysis on functional magnetic resonance imaging dataset}},
  booktitle = {Network},
  year = {2024},
  doi = {10.1162/netn\_a\_00421},
}
Figure