Source-linked AI summary
MSR-IVA: Masked Structural Residual Independent Vector Analysis for State-Aware Fusion of Structural MRI and Dynamic Functional Network Connectivity
Victor Solomon, Zening Fu, Rafal Angryk, Vince D. Calhoun, Jingyu Liu
TL;DR
The paper addresses how to fuse one structural representation with multiple dynamic functional states when independent decompositions lose correspondence and some subjects lack particular states. It proposes MSR-IVA, which combines shared structural components, state-specific residuals, and masks. Relative to independent pairwise IVA, MSR-IVA improved matched source coupling by 6.5% and reduced unmatched source dependence by 15.7%, while producing intermediate cross-state structural similarity.
Problem
Existing fusion approaches do not directly preserve structural correspondence across multiple state-specific dFNC representations or accommodate subjects missing individual states.
Method
MSR-IVA combines a shared structural component with state-specific residual adaptations and masks that restrict each objective to subjects with valid observations.
Results
6.5% matched source coupling improvement and 15.7% unmatched source dependence reduction were achieved relative to IP-IVA.
Takeaways & Limitations
MSR-IVA supports controllable structural sharing that preserves cross-state source correspondence while allowing state-specific adaptation.
Takeaways & Limitations
The evaluation focused on States 1 and 2 because State 3 had near-complete availability and offered limited opportunity to assess masking under missing state expression.
Abstract
from arXiv · showhide
Multimodal fusion of structural MRI (sMRI) and dynamic functional network connectivity (dFNC) can reveal how brain structure relates to changing functional states. When the same structural latent representation is coupled with multiple states, applying independent vector analysis (IVA) separately to each state can produce unrelated structural decompositions, while forcing identical decompositions may suppress state-specific relationships. In addition, not every subject expresses every dynamic state. We propose masked structural residual IVA (MSR-IVA), a state-aware framework that combines a shared structural representation with state-specific residual adaptations and masks for incomplete state expression. On an Alzheimer's Disease Neuroimaging Initiative cohort, MSR-IVA improved matched source coupling by 6.5% and reduced unmatched dependence by 15.7% relative to the independent pairwise IVA baseline. Among subjects expressing both states, mean absolute cross-state structural source correlation was 0.9177 for MSR-IVA versus 0.2978 for no sharing, demonstrating controlled structural sharing that preserves source correspondence while allowing state-specific adaptation.
1. INTRODUCTION
Existing IVA fusion does not preserve structural correspondence across multiple dynamic states, while complete-case requirements exclude subjects lacking a state. MSR-IVA addresses both issues by combining shared structural components, state-specific residuals, and state masks.
- Independent pairwise IVA produces separate structural decompositions across states, whereas identical sharing can suppress state-specific structural–functional relationships.
- Incomplete state expression means some subjects lack observations for particular dynamic states, making complete-case analysis discard otherwise valid state-specific information.
- MSR-IVA fuses one structural modality with multiple partially observed dynamic functional states using shared structural components and residual adaptations.
- State masks ensure each structural–functional fusion objective uses only subjects with valid observations for that state.
- The framework provides soft sharing between fully independent and fully shared structural decompositions while preserving cross-state correspondence and allowing state-specific adaptation.
2. DATA AND PREPROCESSING
The study matched structural and resting-state functional MRI data from an ADNI cohort and derived subject-level representations for two dynamic states. Because state expression was incomplete, masked objectives retained subjects separately within each available state.
- The matched ADNI cohort contained 573 subjects spanning cognitively normal, mild cognitive impairment, and dementia groups.
- A complete-case analysis would retain 124 common subjects, discarding 239 State 1 subjects and 95 State 2 subjects.
- The masked formulation retained all subjects with valid observations within each state-specific objective.
- Dynamic functional connectivity processing used template-guided ICA, 21-time-point sliding windows, and 1378 connectivity features per window.
- State 1 was expressed by 363 subjects, State 2 by 219, and only 124 subjects expressed both selected states.
3. STATE-AWARE IVA FRAMEWORK
The framework extends IVA for multimodal fusion with state-specific dFNC representations by handling incomplete state expression and balancing shared versus state-adapted structural decompositions. It evaluates source coupling, dependence, structural sharing, and robustness against pairwise and sharing baselines.
- IVA foundation: IVA separates distinct source component vectors while allowing statistical dependence among corresponding sources across related datasets.This dependence structure supports linked latent sources in multimodal fusion.
- State-aware formulation: State masks restrict each state-specific objective to subjects with valid observations, preventing absent states from contributing.The valid subject set Ψr determines which observations enter state r.
- Baseline: IP-IVA estimates separate sMRI–dFNC models independently, yielding independently estimated structural decompositions for the two states.Each state has its own structural and functional demixing matrices, without explicit cross-state structural connection.
- Proposed MSR-IVA: MSR-IVA parameterizes each structural demixing matrix as a shared component plus a state-specific residual, while retaining separate functional transformations.The residual penalty regulates state-specific structural adaptation, and α controls the strength of sharing.
- Evaluation: Structural sharing was quantified on the 124-subject common set using one-to-one source assignment, while permutation testing assessed SCV correspondence with family-wise error control.Cstruct characterizes sharing degree rather than serving as a higher-is-better performance metric.
- Evaluation: Moderate residual regularization increased matched source coupling and reduced cross-SCV dependence relative to no sharing, whereas stronger regularization eventually constrained performance.The sensitivity analysis evaluated α across positive values and selected α = 0.1 as a representative moderate setting.
4. RESULTS
Across repeated initializations, MSR-IVA improved source coupling and reduced cross-SCV dependence relative to IP-IVA, while producing intermediate cross-state structural similarity with aligned source correspondence.
- Sensitivity to Structural Residual Regularization: Moderate residual regularization improved matched source coupling and reduced cross-SCV dependence, whereas stronger regularization eventually constrained structural representations too tightly.At α = 0.1, the method was selected for repeated-run comparison.
- Robustness Across Initializations: 6.5% higher mean within-SCV coupling and 15.7% lower mean cross-SCV dependence were achieved by MSR-IVA than IP-IVA across five random initializations.Averaged across states, MSR-IVA achieved Cwithin = 0.9540 ± 0.0116 and Ccross = 0.0504 ± 0.0019.
- Source Coupling Analysis: 0.9177 ± 0.0081 cross-state structural similarity placed MSR-IVA between no sharing at 0.2978 ± 0.0075 and hard sharing at 1.0000.The comparison used 124 subjects expressing both states.
- Source Coupling Analysis: All 20 structural SCVs retained same-index correspondence under optimal matching across five initializations, despite nonidentical state-level representations.This indicates that the observed adaptation was not caused by arbitrary source permutation.
- Source Coupling Analysis: 0.9584 versus 0.8966 mean within-SCV coupling, 0.0485 versus 0.0619 mean cross-SCV dependence, and 0.9099 versus 0.8348 mean coupling contrast favored MSR-IVA over IP-IVA.These state-specific metrics were summarized for the seed-42 comparison.
- Source Coupling Analysis: 0.9389 versus 0.0491 ± 0.0083 in State 1 and 0.9780 versus 0.0482 ± 0.0079 in State 2 showed diagonal coupling above mismatched-SCV null distributions.Both states reached a global empirical p = 1/10001 ≈ 1.0 × 10^-4 with 10,000 permutations and plus-one correction.
5. DISCUSSION
MSR-IVA achieved soft structural sharing: it preserved cross-state source correspondence without forcing identical structural representations and retained state-specific variation while using available observations.
- Controlled Structural Sharing: 0.9177 cross-state structural similarity placed MSR-IVA between no sharing at 0.2978 and hard sharing at 1.0000.The intermediate value characterizes sharing rather than a higher-is-better performance metric.
- Incomplete State Expression: The masking formulation lets each state use available observations instead of restricting fusion to the common-subject intersection.This preserves matched structural–functional relationships while retaining state-specific variation.
6. CONCLUSION
MSR-IVA combines shared structural components, state-specific residuals, and state masks to support controllable sharing across partially observed dFNC states.
- Conclusion: MSR-IVA combines a shared structural component with state-specific residuals and state masks for fusion with multiple partially observed dFNC states.The framework retains available observations for each state while allowing controlled structural sharing.
- Conclusion: 6.5% improved matched source coupling and 15.7% reduced unmatched source dependence were observed relative to IP-IVA across five random initializations.Structural similarity further supported an intermediate regime between independent and identical decompositions.