Reproducible and Individualized Striatal Parcellation Based on Multi-Level Contrastive Learning

Functional magnetic resonance imaging (fMRI)-based parcellation provides a non-invasive approach to study brain integration and segregation across multiple scales, facilitating understanding of the brain organization. However, prevalent clustering-based methods are typically applied independently to each scan or individual, failing to capture both shared and individual-specific features. They often lead to poorly reproducible parcellations. Additionally, low signal-to-noise ratios make the produced parcellation susceptible to spatially discrete parcels. To address these issues, we propose a Multi-Level Contrastive Learning-Based Graph Convolutional Network (MCL-GCN) for reproducible and individualized striatal parcellation. This method employs a spatial graph convolutional network to integrate spatial information while learning functional connectivity features. By incorporating contrastive learning at both voxel and individual levels, the proposed model simultaneously extracts common features within individuals and specific features across individuals, enhancing the reproducibility of the parcellation. Moreover, a homogeneity loss is designed to further improve the functional homogeneity of parcels. Compared with other methods, MCL-GCN demonstrates superior reproducibility, spatial continuity, and comparable homogeneity on human connectome project (HCP) dataset. Furthermore, significant correlations between cognitive behaviors and identified topological features demonstrate the model's ability to capture individual-specific information which is potential to extend our understanding of striatal function.