Harmonizing Structural and Functional Brain Connectivity: A Graph Neural Network Approach
Poster No:
1710
Submission Type:
Abstract Submission
Authors:
Gang Qu1, Aiying Zhang2, Yu-Ping Wang3
Institutions:
1Tulane University, METAIRIE, LA, 2University of Virginia, Charlottesville, VA, 3Tulane University, New Orleans, LA
First Author:
Co-Author(s):
Introduction:
The fusion of multimodal neuroimaging data is essential for advancing neuroscience, offering a comprehensive perspective on the brain dynamics. Structural connectivity (SC) defines the anatomical network of neuronal pathways (Hagmann2010), while functional connectivity (FC) reflects the transient neural activity patterns associated with cognitive functions (Zhang2022). The concurrent analysis of these modalities is challenging, necessitating the reconciliation of SC's static precision with the variable nature of FC to obtain a holistic understanding of the brain's functionality(Damoiseaux2009,Yan2022). Harmonizing these modalities is further complicated by the high-dimensional nature of neuroimaging data, which necessitates sophisticated methods to maintain the integrity of the intricate topology of neural networks(Qu2021). By meticulously integrating SC and FC data within a unified topological context and enriching it with anatomical statistics (AS), our endeavor represents a substantial step forward in multi-modal brain network analysis, promising to enrich the understanding and interpretation of complex neural interactions.
Methods:
The HCP-D maps the connectome in individuals aged 5 to 21, encompassing 1,300 participants(Somerville2018). Scanning protocols, consistent across multiple sites, include sMRI, d-MRI, and rs-fMRI. We applied HCP minimal preprocessing to rs-fMRI data to employ smoothing and motion regression based on FD(Power2012). fMRI was assessed using Glasser360 atlas(Glasser2016), and d-MRI data were preprocessed with MRtrix, facilitating SC measurement with the same ROIs. AS from cortical parcellation complemented these data, offering a detailed view of brain structure and connectivity during development(Cruces2022).
To address the significant disparities in the values of SC and anatomical metrics, we employed normalization techniques. To address the significant disparities in the values, SC were standardized by the square root of the product of gray matter volumes in the interconnected regions.
To refine the representation of brain networks, we implemented a Graph Convolutional Network (GCN), operating on a sparsified binary adjacency matrix to capture the intricacies of the brain's network structure. FC and SC serve as primary features in the GCN, ensuring the preservation of topological accuracy and inter-regional connectivity patterns. AS are introduced as supplementary features for a comprehensive architectural analysis of the brain. Subsequent to the embedding of SC and FC, the embeddings are concatenated with AS, preceding the readout stage that entails graph pooling and final predictive model formulation, as illustrated in the Fig.1.
To address the significant disparities in the values of SC and anatomical metrics, we employed normalization techniques. To address the significant disparities in the values, SC were standardized by the square root of the product of gray matter volumes in the interconnected regions.
To refine the representation of brain networks, we implemented a Graph Convolutional Network (GCN), operating on a sparsified binary adjacency matrix to capture the intricacies of the brain's network structure. FC and SC serve as primary features in the GCN, ensuring the preservation of topological accuracy and inter-regional connectivity patterns. AS are introduced as supplementary features for a comprehensive architectural analysis of the brain. Subsequent to the embedding of SC and FC, the embeddings are concatenated with AS, preceding the readout stage that entails graph pooling and final predictive model formulation, as illustrated in the Fig.1.
Results:
The model was applied to estimate age-adjusted Crystal Cognition Composite (CCC) and Fluid Cognition Composite (FCC) Scores using multimodal neuroimaging data. Preliminary analyses assessed the efficacy of individual and combined data modalities. The model's predictive performance was then benchmarked against established models, including LR, MLP, and graph-based deep learning architectures such as GIN and GAT, with results delineated in Table 2. The integration of FC, SC, and AS markedly enhanced performance compared to single or dual-modality analyses. Additionally, the GCN architecture surpassed all comparison models in predictive accuracy, as quantified by RMSE with associated mean and standard deviation.
Conclusions:
Our integrative framework synthesizes functional and structural connectivity with AS, enhancing neuroimaging analytics. Validation on the HCP-D demonstrates a significant enhancement in cognitive score predictions, outstripping established analytical models. The empirical evidence attests to the GNN's adeptness in modeling the brain's elaborate topology, indicating its utility in augmenting the fidelity of multimodal neuroimaging investigations and contributing to a more nuanced comprehension of cerebral connectivity dynamics.
Higher Cognitive Functions:
Reasoning and Problem Solving
Modeling and Analysis Methods:
Classification and Predictive Modeling
Connectivity (eg. functional, effective, structural) 2
Diffusion MRI Modeling and Analysis
fMRI Connectivity and Network Modeling 1
Keywords:
Computational Neuroscience
Data analysis
FUNCTIONAL MRI
Machine Learning
STRUCTURAL MRI
1|2Indicates the priority used for review
·Illustration of the proposed framework.
Provide references using author date format
Damoiseaux, J. S. (2009). 'Greater than the sum of its parts: A review of studies combining structural connectivity and resting-state functional connectivity', Brain Structure and Function, vol. 213, pp. 525–533.
Glasser, M. F. (2016). 'A multi-modal parcellation of human cerebral cortex', Nature, vol. 536, no. 7615, pp. 171–178.
Hagmann, P. (2010). 'White matter maturation reshapes structural connectivity in the late developing human brain', Proceedings of the National Academy of Sciences, vol. 107, no. 44, pp. 19067–19072.
Power, J. D. (2012). 'Spurious but systematic correlations in functional connectivity MRI networks arise from subject motion', NeuroImage, vol. 59, no. 3, pp. 2142–2154.
Qu, G. (2021). 'Ensemble manifold regularized multi-modal graph convolutional network for cognitive ability prediction', IEEE Transactions on Biomedical Engineering, vol. 68, no. 12, pp. 3564–3573.
Somerville, L. H. (2018). 'The lifespan human connectome project in development: A large-scale study of brain connectivity development in 5–21 year olds', NeuroImage, vol. 183, pp. 456–468.
Yan, W. (2022). 'Deep learning in neuroimaging: Promises and challenges', IEEE Signal Processing Magazine, vol. 39, no. 2, pp. 87–98.
Zhang, A. (2022). 'Decoding age-specific changes in brain functional connectivity using a sliding-window based clustering method', Neuroscience. Preprint.