USDPnet: An Unsupervised Symmetric Deep Framework for Robust Parcellation of Infant Subcortical Nuclei

ICLR 2026 Conference Submission11022 Authors

Keywords: Keywords: unsupervised learning, brain parcellation, subcortical nuclei, infant brain development, symmetric neural networks, Cauchy-Schwarz divergence, neuroimaging, deep clustering, brain atlas, developmental neuroscience

Abstract: Accurate infant subcortical parcellation is vital for understanding early brain development and neurodevelopmental pathology. However, existing methods suffer from initialization sensitivity, poor bilateral consistency, and limited applicability to early postnatal data. We propose USDPnet, an Unsupervised Symmetric Deep Parcellation Network, which integrates deep autoencoder-based feature embedding with divergence-driven clustering. By introducing the generalized Cauchy-Schwarz divergence (GCSD) as the clustering objective, we enhance inter-cluster separability across complex developmental features. A symmetry constraint further enforces bilateral consistency, leading to anatomically plausible and robust delineations. USDPnet operates on surface-based features extracted from infant subcortical nuclei. Experiments show it outperforms traditional and deep clustering baselines. Visualizations are largely consistent with the parcellation results based on anatomy and function connectivity. The resulting parcellations are developmentally grounded, anatomically symmetric, and functionally relevant, offering fine-grained and biologically coherent maps of early subcortical organization. Code is available at https://anonymous.4open.science/r/USDPnet-X12D.

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Submission Number: 11022