Constructing fine‐grained subcortical atlases with connectional consensus graph representation learning

Abstract Objective. The complex internal organization of subcortical structures forms the foundation of critical neural circuits that support sensorimotor processing, emotion regulation, and memory. However, their complex internal organization poses a significant challenge to reliable, fine-scale parcellation. Approach. To overcome the trade-off between anatomical specificity and cross-subject consistency, we propose a novel multiscale subcortical parcellation framework grounded in consensus graph representation learning of diffusion magnetic resonance imaging (dMRI) tractography data. We propose a novel fiber-cluster-based connectivity representation to address the limitations of conventional voxel-level tractography features, thereby enhancing anatomical fidelity and reducing tracking noise. Furthermore, our method preserves local structural coherence while significantly mitigating the curse of dimensionality by leveraging 3D-SLIC supervoxel preparcellation. Finally, we integrate consensus graph representation learning with low-rank tensor modeling, enabling population-level regularization that refines individual embeddings and ensures consistent subcortical parcellations across subjects. By utilizing this framework, we create a new, fine-grained subcortical atlas. Main results. Evaluations using ultra‐high‐field dMRI from Human Connectome Project demonstrate that our method yields subcortical parcels with enhanced reproducibility and microstructural homogeneity. Across diffusion-derived microstructure indices, our atlas consistently achieves the lowest or second-lowest coefficient of variation, with average reductions of 15%–25% compared to existing atlases, thereby supporting robust downstream analyses of structural homology and regional variability. Significance. Our pipeline provides a powerful tool for detailed mapping of subcortical organization, offering promising applications in precision neuroimaging and the discovery of clinical biomarkers for neurological and psychiatric disorders that affect these structures (e.g. Parkinson’s disease, schizophrenia, and major depressive disorder). Our code is available at https://github.com/WanZhonghua/SubcorticalParcellation .