Structure-function coupling of the brain using Graph Neural Network

1405 

Abstract Submission 

YEONGJUN PARK¹, Mansu Kim², Bo-yong Park^3,4,5

¹Department of Computer Engineering, Inha University, Incheon, Republic of Korea, ²Graduate School of Artificial Intelligence, Gwangju Institute of Science and Technology, Gwangju, Republic of Korea, ³Department of Data Science, Inha Unversity, Incheon, Republic of Korea, ⁴Department of Statistics and Data Science, Inha Unversity, Incheon, Republic of Korea, ⁵Center for Neuroscience Imaging Research, Institute for Basic Science, Suwon, Republic of Korea

YEONGJUN PARK  
Department of Computer Engineering, Inha University
Incheon, Republic of Korea

Mansu Kim  
Graduate School of Artificial Intelligence, Gwangju Institute of Science and Technology
Gwangju, Republic of Korea

Bo-yong Park  
Department of Data Science, Inha Unversity|Department of Statistics and Data Science, Inha Unversity|Center for Neuroscience Imaging Research, Institute for Basic Science
Incheon, Republic of Korea|Incheon, Republic of Korea|Suwon, Republic of Korea

One of the core assumptions in neuroscience is that brain structure and function are strongly intertwined. Previous neuroimaging studies proposed methods for predicting functional connectivity from structural connectivity information based on eigenvector decomposition approaches [1] and neural network-based techniques [2]. However, the predictive performance has not been optimized at an individual subject level. In this work, we propose a model that integrates eigenvector decomposition and graph neural network (GNN) methodologies, which appropriately infers relationships of the latent features embedded based on the graph structure among different brain regions [3].

We obtained preprocessed diffusion-weighted imaging (DWI) and resting-state functional magnetic resonance imaging (rs-fMRI) of 974 young, healthy subjects from the Human Connectome Project database (age = 28.76 years; 54.93% female) [4], [5]. We constructed the structural connectivity via diffusion tractography and built the functional connectivity by calculating inter-regional correlations of functional time series, and the matrices were mapped onto the Schaefer atlas with 200 parcels [6]. We estimated the low-dimensional representations of the structural connectivity (i.e., structural gradients) using nonlinear dimensionality reduction techniques, and used these gradients for predicting the functional connectivity matrix. We built a prediction model using GNN, composed of three graph attention network (GAT) layers and one fully connected layer [7], to predict the functional connectivity matrix from the structural connectivity matrix (Fig. 1). The parameters of GAT layers were set as follows: the number of output channels = 30; the number of heads = 2. We randomly divided the data into training (n = 624), validation (n = 155), and test datasets (n = 195). The model was trained using the training data and validated using the validation dataset. The performance of the model was assessed using the test dataset based on the Pearson correlation between the upper triangular elements of the original and predicted matrices. Additionally, we re-performed the above processes based on the five-fold cross-validation instead of the holdout validation approach.

Our model showed high performance in predicting functional connectivity at an individual level. The mean ± standard deviation correlation coefficients were 0.843 ± 0.037 across individuals. It outperformed previous approaches based on the Riemannian optimization approach, which showed a performance of 0.776 ± 0.052 [1], as well as the previous deep learning-based approach, which showed 0.55 ± 0.1 [2]. When we performed the five-fold cross-validation, the performance was 0.843 ± 0.003 across the cross-validation folds, and it outperformed the Riemannian optimization approach based on the three-fold cross-validation (0.775 ± 0.049).

In this study, we proposed an integrated model for predicting the functional connectivity matrix from the structural connectome information. We confirmed that the proposed model outperformed previous state-of-the-art methods. Our approach may provide a new direction in structure-function coupling studies in neuroscience.

Acknowledgements
This study was funded by the National Research Foundation of Korea (NRF-2021R1F1A1052303; NRF-2022R1A5A7033499), Institute for Information and Communications Technology Planning and Evaluation (IITP) funded by the Korea Government (MSIT) (No. 2022-0-00448, Deep Total Recall: Continual Learning for Human-Like Recall of Artificial Neural Networks; No. RS-2022-00155915, Artificial Intelligence Convergence Innovation Human Resources Development (Inha University); No. 2021-0-02068, Artificial Intelligence Innovation Hub), and Institute for Basic Science (IBS-R015-D1).

Classification and Predictive Modeling ¹

Connectivity (eg. functional, effective, structural) ²

Methods Development

Data analysis

FUNCTIONAL MRI

Machine Learning

Modeling

Other - Structure-function coupling; Graph neural network; Deep learning