Exploring brain structure-function interactions associated with fluid cognition based on GCN
Poster No:
2399
Submission Type:
Abstract Submission
Authors:
Jing Xia1, Yi Hao Chan1, Deepank Girish1, Jagath Rajapakse1
Institutions:
1School of Computer Science and Engineering, Nanyang Technological University, Singapore
First Author:
Co-Author(s):
Deepank Girish
School of Computer Science and Engineering, Nanyang Technological University
Singapore
School of Computer Science and Engineering, Nanyang Technological University
Singapore
Jagath Rajapakse
School of Computer Science and Engineering, Nanyang Technological University
Singapore
School of Computer Science and Engineering, Nanyang Technological University
Singapore
Introduction:
Brain structural connectivity (SC) and functional connectivity (FC) not only have individual markers for fluid cognition [1, 2], but their interaction may also contribute to cognition [3], given the inherent linkage between neural function and structure. Incorporating the interaction information between FC and SC could advance the prediction of fluid cognition and enhance our understanding of the fundamental neuroanatomical and neurophysiological bases of cognition. However, most graph neural network (GNN)-based studies on brain connectivity focus on a single imaging modality [4, 5] and they do not capture the interaction between FC and SC. To this end, we propose an explainable structure-function interaction model based on the graph convolution network (GCN) to predict fluid cognition and identify the significant edges and nodes for prediction, as shown in Fig. 1.
Methods:
We use a dataset of 838 participants from HCP [6], including corresponding T1w, resting-state functional MRI (rs-fMRI), and DTI images, to evaluate the proposed method. Rs-fMRI and DTI images are preprocessed by the DPARSF 5.1 advanced toolkit and FSL software to construct the FC and SC with 116 regions of interest (ROIs), following the Anatomical Automatic Labeling protocol [7]. Fluid cognition scores are extracted from the phenotype list, ranging from 87 to 147. We develop an explainable structure-function interaction model to predict fluid cognition (Fig. 1A). Specifically, SC and FC graphs are modeled with brain ROIs as nodes and the connectivity between these ROIs as edges. A bottleneck MLP is employed on the concatenation matrix of structural and functional node features to learn the interactive weights between SC and FC (Fig. 1C). Then, a joint graph is constructed by inserting the learned interactive weights as edges between paired regions of FC and SC graphs. Finally, GCN is applied to the joint graph to predict fluid cognition scores. We adopt GNNExplainer [8] to elucidate salient aspects of the model, identifying a salient subgraph structure, and a salient subset of node features crucial to the prediction. Specifically, we identify the significant inter-edges between paired nodes of FC and SC, as well as key intra-regions within FC and SC contributing to fluid cognition (Fig. 1B).
Results:
Pearson's correlation (r) between the actual and predicted fluid cognition scores was used to evaluate model performance. Our structure-function interaction model (r=0.3) outperformed other state-of-the-art multi-modal fusion approaches, including MV-GCN [9] (r=0.28) and Joint-GCN [10] (r=0.29).
The most salient inter-edges found by GNNExplainer are located in the inferior frontal, precentral, postcentral, superior temporal, and cingulate cortices (Fig. 2). Fluid cognition involves the ability to executive function, such as planning, solving problems, and learning quickly. The structure-function coupling in the frontal, parietal, and temporal regions is mainly associated with executive functional networks engaged in high-level cognitive processes [3]. Therefore, the importance map from GNNExplainer identifies meaningful coupling regions related to fluid cognition. Moreover, Pearson correlation between the important weights and the learned interactive weights is 0.693 (p<0.001), indicating that our model can capture meaningful interactions for prediction. Furthermore, salient intra-regions related to fluid cognition within FC and SC are in the prefrontal, parietal, and occipital cortices, aligning with existing studies on brain structural and functional markers for fluid cognition [2].
The most salient inter-edges found by GNNExplainer are located in the inferior frontal, precentral, postcentral, superior temporal, and cingulate cortices (Fig. 2). Fluid cognition involves the ability to executive function, such as planning, solving problems, and learning quickly. The structure-function coupling in the frontal, parietal, and temporal regions is mainly associated with executive functional networks engaged in high-level cognitive processes [3]. Therefore, the importance map from GNNExplainer identifies meaningful coupling regions related to fluid cognition. Moreover, Pearson correlation between the important weights and the learned interactive weights is 0.693 (p<0.001), indicating that our model can capture meaningful interactions for prediction. Furthermore, salient intra-regions related to fluid cognition within FC and SC are in the prefrontal, parietal, and occipital cortices, aligning with existing studies on brain structural and functional markers for fluid cognition [2].
Conclusions:
We develop an explainable structure-function interaction framework to predict fluid cognition while identifying significant coupling regions between structure and function. This study provides a solution for fusing multi-model features and an explainable approach to explore biomarkers.
Higher Cognitive Functions:
Higher Cognitive Functions Other 2
Modeling and Analysis Methods:
fMRI Connectivity and Network Modeling
Novel Imaging Acquisition Methods:
Multi-Modal Imaging 1
Keywords:
Cognition
FUNCTIONAL MRI
Machine Learning
WHITE MATTER IMAGING - DTI, HARDI, DSI, ETC
1|2Indicates the priority used for review
Provide references using author date format
[2] Dhamala, E., Jamison, K.W., Jaywant, A., et al. (2021), “Distinct functional and structural connections predict crystallised and fluid cognition in healthy adults”, Human brain mapping, vol. 42, no. 10, pp. 3102–3118.
[3] Baum, G.L., Cui, Z., Roalf, D.R., et al. (2020), “Development of structure–function coupling in human
brain networks during youth”, Proceedings of the National Academy of Sciences, vol. 117, no. 1, pp. 771–
778.
[4] He, T., Kong, R., Holmes, A.J., et al. (2020), “Deep neural networks and kernel regression achieve comparable accuracies for functional connectivity prediction of behavior and demographics”, NeuroImage, vol. 206, pp. 116276.
[5] Wen, X., Cao, Q., Zhang, D. (2022), “A multi-scale multi-hop graph convolution network for predicting fluid intelligence via functional connectivity”, in 2022 IEEE International Conference on Bioinformatics and Biomedicine (BIBM). IEEE, pp. 1262–1267.
[6] Van Essen, D.C., Smith, S.M., Barch, D.M., et al. (2013), “The wu-minn human connectome project: an overview”, Neuroimage, vol. 80, pp. 62–79.
[7] Tzourio-Mazoyer, N., Landeau, B., Papathanassiou, D., et al. (2002), “Automated anatomical labeling of activations in spm using a macroscopic anatomical parcellation of the mni mri single-subject brain”, Neuroimage, vol. 15, no. 1, pp. 273–289.
[8] Ying, Z., Bourgeois, D., You, J., et al. (2019), “Gnnexplainer: Generating explanations for graph neural networks”, Advances in neural information processing systems 32.
[9] Liu, J., Ma, G., Jiang, F., et al. (2019), “Community-preserving graph convolutions for structural and functional joint embedding of brain networks”, in 2019 IEEE International Conference on Big Data (Big Data). IEEE, pp. 1163–1168.
[10] Li, Y., Wei, Q., Adeli, E., et al. (2022), “Joint graph convolution for analyzing brain structural and functional connectomes,", in International Conference on Medical Image Computing and Computer-Assisted Intervention. Springer, pp. 231–240.