Clustering-based Multi-View Network Fusion for Estimating Brain Network Atlases of Healthy and Disordered Populations

Background: While several research methods were developed to estimate individual-based representations of brain connectional wiring (i.e., a connectome), traditionally captured using multimodal MRI data (e.g., functional and diffusion MRI), very limited works aimed to estimate brain network atlas for a population of connectomes. Estimating well-representative brain templates is a key step for group comparison studies. However, estimating a network atlas for a population of multi-source brain connectomes lying on different manifolds is absent.

New method: To fill this gap, we propose a cluster-based multi-view brain connectivity fusion framework to estimate a brain network atlas for a population of multi-view brain networks, where each view captures a specific facet of the brain construct. Specifically, given a population of subjects, each with multi-view networks, we first non-linearly fuse multi-view networks into a single fused network for each subject. Then, we cluster the fused networks to identify individuals sharing similar connectional traits in an unsupervised way, which are next averaged within each cluster to generate a representative network atlas. Finally, we construct the final multi-view network atlas by averaging the obtained templates of all clusters.

Results: We evaluated our method on both healthy and disordered populations (with autism and dementia) and spotted differences between network atlases for healthy and autistic groups.

Comparison with existing methods and conclusions: Compared to other baseline methods, our fusion strategy achieved the best results in terms of template centeredness and population representativeness.