Parallel Group ICA + ICA: Joint Estimation of Linked Functional Network Variability and Structural Covariation with Application to Schizophrenia

Shile Qi, Jing Sui, Jiayu Chen, Jingyu Liu, Rongtao Jiang, Rogers F. Silva, Armin Iraji, Eswar Damaraju, Mustafa S. Salman, Dongdong Lin, Zening Fu, Dongmei Zhi, Jessica A. Turner, Juan Bustillo, Judith M. Ford, Daniel H. Mathalon, James Voyvodic, Sarah McEwen, Adrian Preda, Ayşenil Belger, Steven G. Potkin, Bryon A. Mueller, Tülay Adalı, Vince D. CalhounbioRxiv 2019 · DOI

TL;DR

This paper introduces "parallel Group ICA + ICA," a novel method for fusing fMRI and structural MRI data that preserves the rich temporal information from fMRI. This approach successfully identifies linked functional and structural brain patterns associated with schizophrenia and demonstrates their potential as biomarkers for predicting cognitive deficits.

What Problem Does This Paper Address?

The study tackles a key challenge in neuroimaging: how to effectively combine information from different brain imaging modalities (like fMRI and structural MRI) to gain a more comprehensive understanding of brain disorders. Current methods often reduce 4D fMRI data (which includes temporal information) to 3D spatial maps, losing valuable dynamic information. This loss of temporal detail can hinder the detection of subtle, yet important, relationships between brain structure and function, particularly in complex conditions like schizophrenia. The authors aim to bridge this gap by developing a fusion technique that leverages the full temporal information of fMRI alongside structural data.

Methods

The core innovation is the "parallel Group ICA + ICA" (pGICA+ICA) algorithm. It builds upon Group Independent Component Analysis (GICA) for fMRI and Parallel Independent Component Analysis (pICA) for structural MRI.

  • GICA for fMRI: This part processes the 4D fMRI data to extract group-level functional networks (components) and their subject-specific variations, preserving temporal dynamics.
  • pICA for sMRI: This part processes the structural MRI data to identify components related to structural variations across subjects.
  • The "Parallel" Fusion: The key novelty lies in how these two modalities are linked. pGICA+ICA introduces a "variability matrix" for fMRI that captures how individual subject's functional networks deviate from the group average. This matrix is then directly correlated with the structural MRI components. This allows for a joint optimization that maximizes both the independence within each modality and the correlation between the modalities' variability. This direct linkage of first-level fMRI features with structural MRI features is a significant departure from previous methods.

Datasets

The study utilizes two real human brain datasets:

  • fBIRN (Function Biomedical Informatics Research Network) Phase III: A discovery cohort of 311 subjects (162 healthy controls, 149 schizophrenia patients).
  • COBRE (Center for Biomedical Research Excellence): A replication cohort of 177 subjects (94 schizophrenia patients, 83 healthy controls).

Simulated fMRI and sMRI data were also generated to rigorously test the method's performance under controlled conditions.

Key Insights and Ideas

  • Temporal Information is Crucial: The paper strongly advocates for the importance of retaining temporal information from fMRI in multimodal fusion, arguing that its omission in prior methods leads to a loss of valuable insights.
  • Linking Variability: The core idea of linking subject-specific functional network variability (how it deviates from the group norm) with structural variations is a powerful way to uncover disease-related associations.
  • Biomarker Potential: The identified linked fMRI-sMRI components show promise as biomarkers, not only for distinguishing between schizophrenia patients and controls but also for predicting cognitive performance.
  • Robustness to Noise and Association Strength: The method demonstrates effectiveness in detecting these links even under noisy conditions and when the underlying association between modalities is weak, which is often the case in real-world data.

Key Results

  • Simulation: pGICA+ICA outperformed separate GICA and ICA in accurately detecting inter-modality linkages, both for strong (r=0.49) and weak (r=0.28) associations, and was less susceptible to inflating links artificially.
  • Real Data (fBIRN): A significant linked fMRI-sMRI component pair (fMRI_IC3 and sMRI_IC19) was identified, showing group differences between schizophrenia patients and controls. This pair was associated with dysfunctional brain regions previously implicated in schizophrenia.
  • Cognitive Prediction (fBIRN): Features derived from this linked component pair could predict multiple cognitive scores (working memory, verbal learning, composite cognition) with significant accuracy (e.g., r=0.266 for working memory).
  • Cross-Cohort Prediction (COBRE): The predictive power of these features extended to an independent cohort (COBRE), demonstrating their generalizability.
  • Post-Fusion FNC Analysis: The method allows for functional network connectivity (FNC) analysis on the fMRI components after fusion, revealing group differences in network interactions.

Limitations and Future Work

  • Parameter Tuning: The paper mentions that learning rates (λci and ηi) for the independence and association constraints are tuned adaptively, suggesting that careful parameter selection is important for optimal performance.
  • Generalizability to Other Disorders: While the authors suggest the method's applicability to other brain disorders (bipolar disorder, depression, ADHD), this was not directly tested in the current study.
  • Other Modalities: The authors propose that the method can be extended to fuse fMRI with other modalities like diffusion MRI (dMRI), which would be a natural next step.
  • Dynamic FNC: The current study focused on static FNC analysis; incorporating dynamic FNC analysis based on the parallel GICA+ICA results is suggested for future work.
  • Alternative Variability Measures: The authors note that other measures of subject variability (beyond L2-distance) could be incorporated into the method.

Summary updated

Abstract

ABSTRACT There is growing evidence that rather than using a single brain imaging modality to study its association with physiological or symptomatic features, the field is paying more attention to fusion of multimodal information. However, most current multimodal fusion approaches that incorporate functional magnetic resonance imaging (fMRI) are restricted to second-level 3D features, rather than the original 4D fMRI data. This trade-off is that the valuable temporal information is not utilized during the fusion step. Here we are motivated to propose a novel approach called “parallel group ICA+ICA” that incorporates temporal fMRI information from GICA into a parallel ICA framework, aiming to enable direct fusion of first-level fMRI features with other modalities ( e.g. structural MRI), which thus can detect linked functional network variability and structural covariations. Simulation results show that the proposed method yields accurate inter-modality linkage detection regardless of whether it is strong or weak. When applied to real data, we identified one pair of significantly associated fMRI-sMRI components that show group difference between schizophrenia and controls in both modalities. Finally, multiple cognitive domain scores can be predicted by the features identified in the linked component pair by our proposed method. We also show these multimodal brain features can predict multiple cognitive scores in an independent cohort. Overall, results demonstrate the ability of parallel GICA+ICA to estimate joint information from 4D and 3D data without discarding much of the available information up front, and the potential for using this approach to identify imaging biomarkers to study brain disorders.
BibTeX
@misc{qi2019parallel,
  author = {Shile Qi and Jing Sui and Jiayu Chen and Jingyu Liu and Rongtao Jiang and Rogers F. Silva and Armin Iraji and Eswar Damaraju and Mustafa S. Salman and Dongdong Lin and Zening Fu and Dongmei Zhi and Jessica A. Turner and Juan Bustillo and Judith M. Ford and Daniel H. Mathalon and James Voyvodic and Sarah McEwen and Adrian Preda and Ayşenil Belger and Steven G. Potkin and Bryon A. Mueller and Tülay Adalı and Vince D. Calhoun},
  title = {{Parallel Group ICA + ICA: Joint Estimation of Linked Functional Network Variability and Structural Covariation with Application to Schizophrenia}},
  year = {2019},
  doi = {10.1101/595017},
  howpublished = {bioRxiv},
}
Figure