A Large-Scale Deep Normative Modeling of Primary Sulcal Patterns Reveals Deviations in a Spectrum of Disorders

Abstract

Primary sulcal patterns, the spatial arrangement of the earliest cortical folds, emerge prenatally and remain essentially stable after birth. Alterations are associated with cognitive outcomes and hold promise as clinical biomarkers. However, detecting abnormalities is challenging due to the high topological variability and complexity of normal individual sulcal patterns. Here, we introduce an unsupervised generative model that quantifies individual sulcal pattern deviations against a normative baseline learned from 10,349 typically developing subjects spanning early childhood through adulthood. We combine a generative reconstruction error with a discriminator trained on diffusion-generated pseudo-atypical graphs to leverage their complementary strengths in sensitivity and robustness. The discriminator captures subtle anomalies and is robust to demographic and data quality variations, whereas the generative model enables severity stratification and demonstrates cross-site generalizability. Across two clinical cohorts spanning subtle to overt sulcal abnormalities (congenital heart disease and polymicrogyria), our model captured global and localized condition-related deviations and linked them to neurodevelopmental outcomes. By enabling population-scale, individualized quantification of sulcal deviations, this framework provides a foundation for future clinical risk stratification.