Neuro-Anatomy–Informed Self-Supervised Learning for Structural Brain MRI

Self-supervised learning (SSL) has become a foundational paradigm for representation learning from large-scale unlabeled data and underpins many modern foundation models across biological and biomedical domains, including biomedical imaging. However, most existing SSL approaches for biomedical imaging rely on objectives originally designed for natural images, such as contrastive alignment or reconstruction, which primarily enforce instance-level consistency based on generic features. As a result, these objectives are insufficient to capture biologically meaningful structural relationships that emerge across samples and populations, limiting their ability to learn representations aligned with underlying biological organization. In this work, we introduce neuro-anatomy-informed self-supervised learning (NAI-SSL), a framework that moves beyond conventional instance-level objectives by incorporating biologically grounded inductive bias. Focusing on structural brain MRI, we leverage inter-hemispheric structural consistency, a fundamental principle of brain organization, to formulate a structure-aware self-supervised task based on inter-regional covariance. Unlike conventional SSL objectives that operate on individual samples, the proposed task leverages cross-sample covariance statistics to model coordinated variation across anatomically related brain regions, enabling the learning of biologically meaningful cross-sample structural relationships. Empirical results demonstrate that the proposed approach consistently improves representation quality across downstream neuroimaging tasks, while producing representations that better reflect underlying brain organization. Additional analysis and experiments are included in the Supp.