ARTFEED — Contemporary Art Intelligence

Multi-Atlas Disentangled Connectivity Learning for Brain Disorder Representations

publication · 2026-05-11

A team of researchers has introduced Multi-Atlas Disentangled Connectivity LEarning (MADCLE), a framework for multi-branch representation learning that encodes functional connectivity (FC) matrices from various brain atlases simultaneously. While resting-state fMRI-derived FC is frequently utilized to examine changes in brain networks associated with disorders, outcomes can differ based on the selected atlas. MADCLE tackles the issue of cross-atlas variability by developing atlas-specific representations related to diseases, avoiding reliance on a unified latent variable. Further information about this method can be found in the arXiv preprint 2605.07026.

Key facts

  • MADCLE is a multi-branch representation learning framework
  • It jointly encodes FC matrices from different brain atlases
  • FC construction depends on brain atlas choice
  • Different parcellations emphasize distinct organizational features
  • Existing multi-atlas approaches fuse features at shallow level
  • Single-atlas disentanglement methods do not address cross-atlas heterogeneity
  • MADCLE learns atlas-wise disease-related representations
  • The preprint is on arXiv with ID 2605.07026

Entities

Institutions

  • arXiv

Sources