Hyper Morphological Graph for Brain Structural Connection Analysis

Cortical morphological features are considered to be effective biomarkers, yet the high-order interregional cortical connections remain underexplored. Previous advances have predominantly focused on establishing structural correlations between pairs of regions of interest (ROIs), often overlooking the inherent modularity characteristic of human brains. In this work, we propose a cortical hyper-connection representation learning (CHRL) framework to systematically characterize brain modularity. To begin with, we develop a hyper morphological graph where multiple ROIs are densely connected to simulate the modular organization of the human brain. Then, we employ a hypergraph neural network to facilitate information aggregation between ROIs within the same module. Next, considering the prevalent heterogeneity in the structure of human brains, individuals assigned to the same group may exhibit different traits. Therefore, we design a pretraining scheme based on contrastive learning to prebuild a separable feature space with a weak reliance on ground truths. Experimental results on two public datasets validate the effectiveness of our model in both autism spectrum disorder diagnosis and gender classification tasks, outperforming existing competing methods by a large margin.