A Common Space for Human-Macaque Brain Comparison Constructed by Graph Convolutional Neural Network
Poster No:
1767
Submission Type:
Abstract Submission
Authors:
Haiyan Wang1, Yuheng Lu1, Yumeng Xin2, Luqi Cheng3, Yufan Wang1, Weiyang Shi1, Deying Li1, Congying Chu1, Lingzhong Fan1, Ning Liu2, Tianzi Jiang1
Institutions:
1Brainnetome Center, Institute of Automation, Chinese Academy of Sciences, Beijing, China, 2Institute of Biophysics, Chinese Academy of Sciences, Beijing, China, 3School of Life and Environmental Sciences, Guilin University of Electronic Technology, Guilin, China
First Author:
Co-Author(s):
Luqi Cheng
School of Life and Environmental Sciences, Guilin University of Electronic Technology
Guilin, China
School of Life and Environmental Sciences, Guilin University of Electronic Technology
Guilin, China
Congying Chu
Brainnetome Center, Institute of Automation, Chinese Academy of Sciences
Beijing, China
Brainnetome Center, Institute of Automation, Chinese Academy of Sciences
Beijing, China
Lingzhong Fan
Brainnetome Center, Institute of Automation, Chinese Academy of Sciences
Beijing, China
Brainnetome Center, Institute of Automation, Chinese Academy of Sciences
Beijing, China
Tianzi Jiang
Brainnetome Center, Institute of Automation, Chinese Academy of Sciences
Beijing, China
Brainnetome Center, Institute of Automation, Chinese Academy of Sciences
Beijing, China
Introduction:
Cross-species comparative research on primates plays a crucial role in understanding the unique high-order cognitive functions that distinguish humans from non-human primates. Given the variations in brain scale, morphology, and function across different species, it is essential to define a common space when conducting cross-species comparisons[1]. Currently, the common space definition is mainly based on known homologous landmarks and relies on single-modality images. In this study, we proposed a novel data-driven approach to create a common space using a graph convolutional neural network (GCN), effectively integrating both structural connectivity and resting-state functional connectivity information.
Methods:
For humans, data from Human Connectome Project[2] was used. For macaques, data was collected in a 3T Siemens Prisma MRI scanner at Beijing MRI Center for Brain Research (N=8, Anesthetic: T1, T2, and DTI, Awake: resting-state fMRI). For macaques, structural data was processed with HCP-NHPpipelines[3], resting-state fMRI data was processed using afni_proc.py. Vertex-to-vertex functional connectivity was calculated for both humans and macaques. The preprocessing of DTI data was conducted using FSL for both humans and macaques, and each vertex's blueprint was calculated using XTRACT[4]. Each of the blueprint feature was homogenous between humans and macaques[5]. The data of humans and macaques was separately mapped to a 10k surface.
In this study, we constructed a common space for cross-species comparison between humans and macaques using GCN (Fig. 1). Specifically, we used cortical vertices as the nodes of the graph. The edges of the graph were derived from the triangular facets of the surface or the vertex-to-vertex functional connection. We selected the 10, 20, 30, and 40 strongest nodes associated with each cortical vertex for the functional connections. The blueprints were used as features for the graph nodes. The model consisted of two Chebyshev convolutional layers and a linear layer. Since each convolutional layer integrated the features of connected nodes, the final output of the linear layer formed a feature space that combined both structural and functional connections. Finally, an attention mechanism was employed to map from human to macaque.
In this study, we constructed a common space for cross-species comparison between humans and macaques using GCN (Fig. 1). Specifically, we used cortical vertices as the nodes of the graph. The edges of the graph were derived from the triangular facets of the surface or the vertex-to-vertex functional connection. We selected the 10, 20, 30, and 40 strongest nodes associated with each cortical vertex for the functional connections. The blueprints were used as features for the graph nodes. The model consisted of two Chebyshev convolutional layers and a linear layer. Since each convolutional layer integrated the features of connected nodes, the final output of the linear layer formed a feature space that combined both structural and functional connections. Finally, an attention mechanism was employed to map from human to macaque.
Results:
Taking myelin map projection as an example, Fig. 2 shows the prediction performance of macaque myelin maps based on the GCN model. The best prediction results were obtained using the 20 strongest nodes associated with each cortical vertex as edges (Fig. 2A). Fig. 2B shows the training, validation, and testing processes for both the left and right hemispheres. As the number of training epochs increased, the loss gradually decreased, and the correlation between the true and predicted maps increased. Fig. 2C presents a comparison with existing methods, where 'BPregister' represents the projection based on the KL divergence of blueprint features[5], and 'Joint-embedding' refers to the gradient method based on functional connectivity[6]. It can be observed that the GCN-based method outperforms the other two methods under different numbers of subjects. Specifically, when we use 600 human subjects, the accuracies are: BPregister, L=0.592, R=0.615, mean=0.604; Joint-embedding, L=0.552, R =0.648, mean=0.600; GCN, L=0.631, R=0.652, mean=0.641.
Conclusions:
In the present study, we employed a GCN-based approach to integrate both structural and functional connectivity, constructing a novel data-driven common space for cross-species comparisons between humans and macaques. By using this common space, the projection of human myelin maps onto macaques can be improved. Notably, when training this model for projecting different task-based activations simultaneously, a more comprehensive common space can be constructed. Furthermore, since DTI and resting-state fMRI can be more easily acquired compared to task-based fMRI, this method can also be applied to newborns and other species.
Modeling and Analysis Methods:
fMRI Connectivity and Network Modeling 1
Methods Development 2
Multivariate Approaches
Neuroanatomy, Physiology, Metabolism and Neurotransmission:
Cortical Anatomy and Brain Mapping
Keywords:
Cross-Species Homologues
MRI
Other - Graph Convolutional Neural Network; Blueprint; Functional Connectivity; Multi-modal; Human; Macaque; Common Space
1|2Indicates the priority used for review
Provide references using author date format
[2] Glasser, M. F. (2013), 'The minimal preprocessing pipelines for the Human Connectome Project', Neuroimage, vol. 80, pp. 105-124.
[3] Donahue, C. J. (2016), 'Using diffusion tractography to predict cortical connection strength and distance: a quantitative comparison with tracers in the monkey', Journal of Neuroscience, vol. 36, pp. 6758-6770.
[4] Warrington, S. (2020), 'XTRACT-Standardised protocols for automated tractography in the human and macaque brain', Neuroimage, vol. 217.
[5] Mars, R. B. (2018), 'Whole brain comparative anatomy using connectivity blueprints', eLife, vol. 7.
[6] Xu, T. (2020), 'Cross-species functional alignment reveals evolutionary hierarchy within the connectome', Neuroimage, vol. 223.