Abstract

Brain connectomes—the structural or functional connections between distinct brain regions—are widely used for neuroimaging studies. However, different ways of brain parcellation are proposed and used by different research groups without any consensus of their superiority. The variety of choices in brain parcellation makes data sharing and result comparison between studies difficult. Here, we propose a framework for transforming connectomes from one parcellation to another to address this problem. The optimal transport between nodes of two parcellations is learned in a data-driven way using graph matching methods. Spectral embedding is applied to the source connectomes to obtain node embeddings. These node embeddings are then transformed into the target space using the optimal transport. The target connectomes are estimated using the transformed node embeddings. We test the effectiveness of the proposed framework by learning the optimal transport based on data from the Human Connectome Project Young Adult, and applying it to structural connectomes data from the Lifespan Human Connectome Project Development. The efficacy of our approach is validated by comparing the estimated connectomes against their counterparts (connectomes generated directly from the target parcellation) and testing the pre-trained predictive models on estimated connectomes. We show that the estimated connectomes are highly correlated with the actual data, and predictive models for age achieve high accuracies. Overall, our proposed framework holds great promises in facilitating the generalization of connectome-based models across different parcellations.

Full Text: [PDF]