Source-linked AI summary
Group ICA 2.0: Closing the Gap Between Subjects and Group Latent Decomposition with Copula-Linked Group ICA (CoLiG-ICA)
Oktay Agcaoglu
TL;DR
Conventional gICA primarily identifies shared components, limiting recovery of networks present only in individuals or subject subsets. CoLiG-ICA links subject decompositions to templates while jointly estimating additional cohort-only and subject-only sources. It showed improved component independence and motion-related variance reduction, and identified three additional networks in a schizophrenia-only analysis.
Problem
Conventional gICA primarily identifies components shared across subjects, limiting recovery of networks present only in individuals or subject subsets.
Method
CoLiG-ICA combines template coupling with ICA, copula-based dependence modeling, and deep learning optimization to estimate template-linked, cohort-only, and subject-only components.
Results
CoLiG-ICA showed significantly lower intercomponent spatial dependence and reduced motion-related variance than MOO-ICAR, while identifying three additional networks beyond 53 NeuroMark components.
Takeaways & Limitations
CoLiG-ICA represents cohort- and subject-level variability beyond shared template-linked networks while preserving large-scale functional network structure.
Takeaways & Limitations
Subject-level free components showed a mixture of nuisance-related and potentially meaningful patterns, motivating systematic evaluation of their recovery and interpretability.
Abstract
from arXiv · showhide
Group Independent Component Analysis (gICA) is widely used to decompose high-dimensional functional MRI data into interpretable brain networks. However, conventional gICA primarily identifies components shared across subjects. This group-level assumption can limit the recovery of networks present only in individuals or subject subsets, reducing sensitivity to intersubject heterogeneity in clinical neuroimaging datasets. We introduce Copula-Linked Group ICA (CoLiG-ICA), an algorithm in the Group ICA 2.0 framework that jointly estimates template-linked, cohort-only, and subject-only brain networks within a unified model. CoLiG-ICA combines ICA-based spatial decomposition, copula-based dependence modeling, and deep learning optimization to preserve the consistency and interpretability of template-constrained ICA while enabling free components beyond the reference networks. By linking subject decompositions to shared templates and jointly estimating cohort-only and subject-only sources, CoLiG-ICA represents individual variability not captured by conventional group priors. We evaluate CoLiG-ICA using resting-state fMRI data from the UCLA-CNP dataset and compare it with conventional constrained ICA in estimating template-linked components, discovering additional free components, improving component independence, and capturing subject-level variability beyond the shared group prior. Compared with MOO-ICAR, CoLiG-ICA showed significantly lower intercomponent spatial dependence, indicating improved subject-level component independence, and significantly reduced motion-related variance in the template-linked components. Additionally, in a schizophrenia-only group analysis, CoLiG-ICA identified three additional resting-state networks beyond the 53 template-linked NeuroMark components: one sensorimotor and two visual networks.
1 INTRODUCTION
Conventional group ICA simplifies cross-subject analysis by estimating shared components, but its cohort-level assumptions limit sensitivity to subject-specific variability. CoLiG-ICA extends template-constrained decomposition with cohort- and subject-specific networks through Group ICA 2.0.
- Conventional group ICA: Conventional group ICA reduces and concatenates subject data before estimating common group-level spatial components, then uses back-reconstruction for subject-specific maps and time courses.This framework simplifies cross-subject component interpretation.
- Limitations of conventional gICA: gICA can be cohort-specific, requiring re-estimation across datasets and producing components that differ across studies, complicating between-group comparisons.Differences may involve intrinsic networks and artifact components.
- Related approaches: Independent Vector Analysis preserves subject-level variability while aligning components, but its multiset optimization can become computationally and memory intensive as subject numbers increase.Published multi-subject resting-state fMRI applications remain limited in scale.
- Group ICA 2.0: Group ICA 2.0 combines deep learning optimization and copula-based dependence modeling for flexible multimodal linkage and group- and subject-level decompositions.CoLiG-ICA is introduced as a full methodological formulation within this framework.
- CoLiG-ICA: CoLiG-ICA integrates ICA, copula-based dependence modeling, and deep learning optimization while retaining template-constrained estimation benefits.The method estimates flexible group and individual decompositions.
- CoLiG-ICA: CoLiG-ICA estimates template-linked, cohort-specific, and subject-specific networks while maximizing independence and allowing different model orders across subjects.Different model orders support multi-scale analysis across individuals and cohorts.
2 METHOD
CoLiG-ICA links subject-level decompositions to shared templates while retaining unconstrained components that can capture cohort-only and subject-only patterns. It combines template coupling, copula modeling, and subject-specific optimization within one decomposition framework.
- Model overview: CoLiG-ICA couples each subject’s functional networks to provided templates while allowing additional cohort-only and subject-only components.This design retains consistent component identity across subjects while extending decomposition beyond group-level priors.
- Model overview: Template-linked components are coupled to reference maps with Gaussian copulas, whereas free components remain unconstrained.The block diagram identifies the copula-linked and unconstrained branches of the model.
- Subject-level decomposition: The spatial ICA model represents each subject’s data as X_s = A_s M_s, with component time courses and spatial maps jointly organized by the model order.The model order is partitioned into template-linked and free components.
- Subject-level decomposition: Subject-specific ICA unmixing matrices are optimized separately for each subject after PCA projection reduces and whitens the fMRI data.The component maps are obtained as M_s = W_s Z_s, with U_s = W_s P_s relating them directly to the original data.
- Template coupling: The likelihood uses a copula density to link each subject component with its matched template component, and optimization minimizes the negative log-likelihood.For the Gaussian copula, the dependence parameter satisfies |ρ_i| < 1.
- Template coupling: Only the first C_T components receive copula constraints; the remaining C_F_s components lack template coupling and can capture patterns absent from the template.Component-template assignments are initialized before optimization and may be updated early until they stabilize.
3 ANALYSIS
The analysis evaluated CoLiG-ICA on quality-controlled UCLA-CNP resting-state fMRI and compared it with MOO-ICAR using the NeuroMark reference. A schizophrenia-only group analysis additionally estimated free components beyond the template-linked networks.
- Dataset and preprocessing: The UCLA-CNP resting-state fMRI data were preprocessed with a standard SPM12 pipeline after discarding the first eight functional volumes.The preprocessing included slice-timing correction, realignment, coregistration, tissue segmentation, and normalization.
- Dataset and preprocessing: After quality control, 190 UCLA-CNP participants were retained: 99 healthy controls, 30 with ADHD, 33 with bipolar disorder, and 28 with schizophrenia.Scans were flagged for excessive framewise displacement, translation, rotation, or insufficient remaining volumes.
- Comparative analysis: CoLiG-ICA was compared with MOO-ICAR after voxelwise temporal mean removal, using the 53-component NeuroMark template and model order 100.This configuration produced 53 template-linked and 47 additional free components for CoLiG-ICA.
- Comparative analysis: MOO-ICAR estimates only the 53 reference-template components, whereas CoLiG-ICA estimates additional components outside that template.The comparison therefore tests both template-linked decomposition and recovery of free components.
- Functional connectivity: Mean functional network connectivity matrices were obtained for both CoLiG-ICA and MOO-ICAR across seven functional domains.The domains were subcortical, auditory, sensorimotor, visual, cognitive control, default mode, and cerebellar.
- Schizophrenia-only analysis: In the schizophrenia-only group analysis, CoLiG-ICA used two-level PCA and estimated 53 template-linked plus 47 free components.This analysis was designed to identify additional cohort-specific components.
4 RESULTS
CoLiG-ICA reduced motion-related variance and intercomponent spatial dependence in template-linked components relative to MOO-ICAR, while identifying additional cohort-specific networks. These free components showed resting-state spectral characteristics and aligned with the existing modular FNC organization, although they were not systematically evaluated.
- Motion-related variance: CoLiG-ICA significantly reduced motion-related R^2 values in template-linked components under both six-parameter and Friston 24-parameter motion models.The reductions were 0.0199 and 0.0198, respectively.
- Spatial dependence: 0.06463 lower mean absolute off-diagonal spatial correlation was observed for CoLiG-ICA than MOO-ICAR across 190 subjects.The difference was highly significant by paired t-test and confirmed by a Wilcoxon signed-rank test.
- Cohort-specific components: Three additional cohort-specific components beyond the 53 NeuroMark components were classified as one sensorimotor and two visual resting-state networks.The visual components formed a lateralized pair, predominantly involving the left and right hemispheres.
- Cohort-specific components: The three additional components exhibited greater low-frequency power and reduced high-frequency power, consistent with resting-state network spectra.Spectra were calculated per participant and then averaged across participants.
- Functional network connectivity: The additional components’ connectivity patterns aligned well with the existing modular organization of the FNC matrix.The schizophrenia FNC analysis included 28 participants.
- Interpretive boundary: The free components were not systematically evaluated, despite reductions in motion-related variance and intercomponent spatial dependence among NeuroMark-linked components.Preliminary visual inspection suggested a mixture of nuisance-related and potentially meaningful network patterns.
5 CONCLUSION
CoLiG-ICA jointly estimates template-linked and free components, improving component independence and reducing motion-related variance while revealing additional resting-state networks. Larger, more heterogeneous clinical cohorts remain important for evaluating its broader utility.
- CoLiG-ICA jointly estimates template-linked and additional free components while maximizing independence among all estimated sources.
- In schizophrenia-only analysis, CoLiG-ICA identified three additional resting-state components beyond 53 template-linked NeuroMark components: one sensorimotor and two lateralized visual components.
- CoLiG-ICA showed significantly lower intercomponent spatial dependence and reduced motion-related variance in template-linked components compared with MOO-ICAR.
- Future work will test CoLiG-ICA in larger, more heterogeneous clinical cohorts and populations with structural abnormalities that may challenge template-guided approaches.