Geometric constraints on individual brain function: a deep learning approach
Poster No:
1440
Submission Type:
Abstract Submission
Authors:
Mehul Gajwani1, Zachary Zendarski1, Stuart Oldham2, Trang Cao1, Mehrtash Harandi1, James Pang1, Alex Fornito1
Institutions:
1Monash University, Melbourne, Victoria, 2Murdoch Children's Research Institute, Melbourne, Victoria
First Author:
Co-Author(s):
Introduction:
Brain structure necessarily constrains brain function, but correlations between structural connectivity (SC; as measured by streamline count from diffusion MRI) and functional connectivity (FC; as measured by pairwise correlation of resting-state BOLD-fMRI time-series) are relatively modest. Deep learning models can reveal a tighter relationship, predicting FC from SC with a reasonable degree of accuracy (e.g. Pearson's r = 0.55 ± 0.1 for individuals and r = 0.9 ± 0.1 for group average estimates) (Sarwar et al., 2021).
Recent work (Pang et al., 2023) has shown that brain geometry may contain at least as much information about function as structural connectivity. Specifically, linear combinations of the eigenmodes of cortical surface geometry are capable of reconstructing group-averaged FC (e.g. group r = 0.7 when using 50 modes). Here, we move beyond these linear models to evaluate the relative performance of SC and cortical geometry in predicting FC through deep learning.
Recent work (Pang et al., 2023) has shown that brain geometry may contain at least as much information about function as structural connectivity. Specifically, linear combinations of the eigenmodes of cortical surface geometry are capable of reconstructing group-averaged FC (e.g. group r = 0.7 when using 50 modes). Here, we move beyond these linear models to evaluate the relative performance of SC and cortical geometry in predicting FC through deep learning.
Methods:
We used structural, functional, and diffusion MRI data from the Human Connectome Project as previously described (Glasser et al., 2013; Oldham & Ball, 2023; Pang et al., 2023). Cortical surface meshes, SC, and FC matrices passed quality control in 967 participants; hence, we used 767 were used for training, 100 for validation, and 100 for testing.
We used a fully-connected feed-forward neural network (Fig. 1B) developed in (Sarwar et al., 2021) with vectorized input to test three competing models using different input data from individuals (Fig. 1). The input of the first model input is the upper triangle of the SC matrix with 100 nodes (Fig. 1A). The input of the second model is the parcellated (Schaefer et al., 2018) first 50 geometric eigenmodes (GM) calculated in fsLR-32k space as previously described (Pang et al., 2023) (Fig. 1D and E). Finally, the input of the third model is 50 connectome eigenmodes (CM) (Naze et al., 2021) generated from the graph Laplacian of the SC matrix . After training (Fig. 1C and F), one run of each model was selected by matching the model inter-prediction similarity to the empirical distribution of similarity between individuals' FC.
We used a fully-connected feed-forward neural network (Fig. 1B) developed in (Sarwar et al., 2021) with vectorized input to test three competing models using different input data from individuals (Fig. 1). The input of the first model input is the upper triangle of the SC matrix with 100 nodes (Fig. 1A). The input of the second model is the parcellated (Schaefer et al., 2018) first 50 geometric eigenmodes (GM) calculated in fsLR-32k space as previously described (Pang et al., 2023) (Fig. 1D and E). Finally, the input of the third model is 50 connectome eigenmodes (CM) (Naze et al., 2021) generated from the graph Laplacian of the SC matrix . After training (Fig. 1C and F), one run of each model was selected by matching the model inter-prediction similarity to the empirical distribution of similarity between individuals' FC.
Results:
Model accuracy was assessed using Pearson correlation. Group level estimates were generated by averaging model predicted FC (pFC) across individuals and comparing it to the average empirical FC (eFC) across the same individuals. Across 100 individuals in the test set, the resulting correlations between the group pFC and eFC are r = 0.98 for the SC model, r = 0.99 for the GM model, and r = 0.99 for the CM model (Fig. 2A). Furthermore, all models accurately predict the group average even when using test sets with fewer individuals (Fig. 2B).
At the individual level, the models reconstruct individual FC matrices more accurately than the previous state-of-the-art(Sarwar et al., 2021): SC r = 0.63 ± 0.19; GM r = 0.64 ± 0.15; CM r = 0.65 ± 0.17. Results from five example individuals are shown in Fig. 2C. Models generally replicate large-scale patterns of FC, and individuals with erroneous reconstructions vary between models. Finally, paired-samples ANOVA showed no significant difference between the predictive accuracy of each model across 100 test individuals.
At the individual level, the models reconstruct individual FC matrices more accurately than the previous state-of-the-art(Sarwar et al., 2021): SC r = 0.63 ± 0.19; GM r = 0.64 ± 0.15; CM r = 0.65 ± 0.17. Results from five example individuals are shown in Fig. 2C. Models generally replicate large-scale patterns of FC, and individuals with erroneous reconstructions vary between models. Finally, paired-samples ANOVA showed no significant difference between the predictive accuracy of each model across 100 test individuals.
Conclusions:
Overall, we demonstrate that structural connectivity, geometric eigenmodes, and connectome eigenmodes can reconstruct functional connectivity with comparable accuracy. Notably, the geometric eigenmodes do not directly measure any information pertaining to inter-regional connectivity; instead, they represent local variations in shape and carry an implicit distance-dependence between points that is captured by an exponential-decay rule that is known to dominate many features of cortical organization. Together, these findings suggest that cortical geometry is as informative of large-scale brain functional connectivity as inter-regional structural connectivity.
Modeling and Analysis Methods:
Classification and Predictive Modeling 1
Connectivity (eg. functional, effective, structural)
Diffusion MRI Modeling and Analysis
fMRI Connectivity and Network Modeling 2
Task-Independent and Resting-State Analysis
Keywords:
Computational Neuroscience
Cortex
FUNCTIONAL MRI
Machine Learning
MRI
NORMAL HUMAN
STRUCTURAL MRI
Tractography
Other - Eigenmodes
1|2Indicates the priority used for review
Provide references using author date format
Naze, S. (2021). 'Robustness of connectome harmonics to local gray matter and long-range white matter connectivity changes'. NeuroImage, 224, 117364. https://doi.org/10.1016/j.neuroimage.2020.117364
Oldham, S. (2023). 'A phylogenetically-conserved axis of thalamocortical connectivity in the human brain'. Nature Communications, 14(1), Article 1. https://doi.org/10.1038/s41467-023-41722-8
Pang, J. C. (2023). 'Geometric constraints on human brain function'. Nature, 1–9. https://doi.org/10.1038/s41586-023-06098-1
Sarwar, T. (2021). 'Structure-function coupling in the human connectome: A machine learning approach'. NeuroImage, 226, 117609. https://doi.org/10.1016/j.neuroimage.2020.117609
Schaefer, A. (2018). 'Local-Global Parcellation of the Human Cerebral Cortex from Intrinsic Functional Connectivity MRI'. Cerebral Cortex, 28(9), 3095–3114. https://doi.org/10.1093/cercor/bhx179