REPRESENTATION LEARNING ON NATIVE CORTICAL SURFACES: FROM GEOMETRY TO INDIVIDUAL TRAITS

ICLR 2026 Conference Submission16848 Authors

Keywords: cortical surface, self-attention, transformer

Abstract:

Analyzing the intricate geometry of the cerebral cortex is fundamental to understanding the neuroanatomical basis of individual traits. However, the fundamental conflict between powerful, grid-dependent architectures like Transformers and the irregular cortical mesh has forced a compromise: the distortive practice of spherical projection. This act of simplification discards the geometric subtleties we aim to study.
To resolve this foundational data-architecture mismatch, we propose the Native Cortical Surface Representation Learning Model (NCS-RL), an end-to-end framework that reshapes the data to fit the model, not the other way around. Its first component, the Canonical Surface Generator, creates a shared, regular topological grid across all subjects. Onto this grid, it precisely maps each individual's unique geometric details via diffeomorphic deformation. This single process achieves three critical goals simultaneously: it establishes a principled tokenization for Transformers, resolves inter-subject correspondence, and yields a spectrum of anatomically faithful variations for data augmentation.
With the cortical surface now represented as a structured and geometrically rich sequence of tokens, the second component, the Cortical Transformer, is designed to interpret it. Its dual-pathway architecture is built to leverage this new data structure: one pathway uses our novel Adjacency Self-Attention to learn fine-grained local geometric patterns directly from the native surface priors, while the other captures global context. A gated mechanism then fuses these pathways, forging a holistic representation that understands not just what a cortical region is, but precisely how it is shaped.
Moreover, to ensure geometric fidelity, our model was pre-trained on over $5,000$ subjects from the ABCD, HCP, and ABIDE datasets. Our method demonstrates state-of-the-art performance in experiments and ablation studies, including phenotype prediction and functional map regression. Our implementation is available in the supplementary material and will be released.

Primary Area: applications to neuroscience & cognitive science

Submission Number: 16848