Multimodal Dynamic Fusion Captures Flexible, Time-Sensitive Structure-Function Linkages in MRI Data
Poster No:
1977
Submission Type:
Abstract Submission
Authors:
Marlena Duda1, Vince Calhoun2
Institutions:
1Georgia State University, Atlanta, GA, 2GSU/GATech/Emory, Decatur, GA
First Author:
Co-Author:
Introduction:
A major goal in the field of neuroimaging centers on capturing the relationship between brain structure and time-varying functional activity/connectivity. To this end, recent approaches have estimated a single structural basis set of the brain and then sought to represent functional activation time series as a linear combination of these structural manifolds [1,2], a technique which can be categorized as asymmetric data fusion. Here, we introduce dynamic fusion, an ICA-based symmetric fusion approach which captures unique structural basis sets with respect to changing functional manifolds derived from dynamic functional connectivity (dFNC) states.
Methods:
We analyzed functional and structural MRI data from N = 833, 310 subjects in the HCP [3] (healthy controls) and FBIRN (schizophrenia [SZ]/control) [4] datasets, respectively. We extracted dFNC states from resting-state fMRI using a sliding window Pearson correlation approach in HCP [5] and via filter bank connectivity (FBC) approach [6,7] in FBIRN, of which the latter enables both time- and frequency-resolved dFNC states. Subject-average connectomes for each dFNC state were separately fused with subject gray matter volume (GMV) maps in independent fusion experiments via the mCCA + jICA framework [8], and cross-fusion correlation was computed for both GMV and dFNC components. In FBIRN, we assessed SZ/control group differences in component loading parameters via two-sample t-tests. We also applied dynamic fusion in a sliding window manner (w = 10 TR) across task boundaries in HCP working memory (WM) task fMRI to investigate the adaptive structural basis set in real time across changing functional contexts.
Results:
We observed low cross-fusion correspondence overall for the dFNC components, which was expected as the functional input differed across each independent fusion; however, even though the GMV inputs were identical across all fusion experiments, only a few GMV components exhibited high cross-fusion correspondence, followed by a steep drop-off of cross-fusion correspondence (Fig 1A-B). This suggests most structural components are highly influenced by the joint relationship with the functional data, i.e., "dynamic", and only a few GMV components appear consistently regardless of the dFNC inputs, i.e., "static" structural components. Of the static components, one was identified across all experiments in both HCP and FBIRN datasets, suggesting a "global static" structural component marked by higher GMV in visual/cerebellar regions, with some evident state-specific variations (Fig 1C). We observed significant group differences in GMV component loadings for static and dynamic components alike, with the strongest group differences found for dynamic components (Fig 1D). In the HCP WM sliding window experiment, we observed the emergence of structural stability in different components in clear relation to known task boundaries (Fig 2).
·Figure 1
·Figure 2
Conclusions:
Here we propose a new approach for investigating the link between brain structure and time-varying brain function, termed dynamic fusion. Our approach is fully data driven and allows both modalities to contribute to the fusion equally (i.e. symmetric fusion), thus enforcing fewer assumptions and enabling a broader spectrum of flexibility than recent works in structural dynamics. We show that dynamic fusion identifies distinct structural basis sets that are specific to each dFNC state and are not observed when FNC from the full time series is considered at once, which challenges the notion that a single structural manifold is sufficient or appropriate for representing every time point in a rs-fMRI time series. Our results also suggest that dynamic components, which are driven by the changing linkage to the varying functional manifolds, capture stronger SZ/control group differences than static components, indicating they may encode unique aspects of clinically-relevant pathophysiology that are missed with traditional fusion approaches.
Modeling and Analysis Methods:
Activation (eg. BOLD task-fMRI)
Connectivity (eg. functional, effective, structural)
Methods Development 2
Multivariate Approaches 1
Task-Independent and Resting-State Analysis
Keywords:
Data analysis
FUNCTIONAL MRI
Machine Learning
Multivariate
Schizophrenia
STRUCTURAL MRI
Other - Multimodal Data Fusion
1|2Indicates the priority used for review
Provide references using author date format
[2] Pang, J.C. et al. (2023), “Geometric constraints on human brain function,” Nature, vol. 618, no. 7965, Art. no. 7965.
[3] Van Essen, D.C. et al. (2013), “The WU-Minn Human Connectome Project: An overview,” NeuroImage, vol. 80, pp. 62-79.
[4] Keator, D.B. et al. (2015), “The Function Biomedical Informatics Research Network Data Repository,” NeuroImage, vol. 124, pp. 1074-1079.
[5] Abrol, A. et al. (2017), “Schizophrenia Shows Disrupted Links between Brain Volume and Dynamic Functional Connectivity,” Front. Neurosci., vol. 11.
[6] Duda, M. et al. (2023), “Alterations in grey matter structure linked to frequency-specific cortico-subcortical connectivity in schizophrenia via multimodal data fusion,” bioRxiv 2023.07.05.547840
[7] Faghiri, A. et al. (2021), “A unified approach for characterizing static/dynamic connectivity frequency profiles using filter banks,” Netw. Neurosci., vol. 5, no. 1, pp. 56-82.
[8] Sui, J. et al. (2013), “Three-way (N-way) fusion of brain imaging based on mCCA+jICA and its application to discriminating schizophrenia,” NeuroImage, vol. 66, pp. 119-132.