Learning Cross-Atlas Consistent Brain Disorder Representations via Disentangled Multi-Atlas Functional Connectivity Learning

Researchers propose MADCLE, a machine learning framework that learns consistent brain disorder representations across multiple brain atlases by disentangling disease-related features from atlas-dependent and covariate factors. The approach demonstrates competitive performance on neurological disorder datasets (ADNI and ADHD-200) while addressing the fundamental problem that different brain parcellation schemes produce heterogeneous and sometimes contradictory functional connectivity representations.

This research addresses a persistent challenge in neuroimaging-based disorder identification: functional connectivity patterns derived from resting-state fMRI critically depend on which brain atlas researchers choose, yet no consensus exists on optimal parcellation schemes. Different atlases emphasize distinct organizational features, making it difficult to develop robust, generalizable diagnostic models. MADCLE tackles this by learning disease-specific representations that remain consistent across multiple atlas schemes simultaneously, rather than forcing artificial fusion of atlas-derived features at shallow levels.

The framework's core innovation lies in structured disentanglement—systematically separating disease-related information from confounding factors like patient covariates and atlas-dependent artifacts. By encouraging cross-atlas alignment of disease representations while allowing atlas-specific modeling of residual factors, MADCLE achieves better interpretability and generalization. This approach reflects broader trends in machine learning toward more principled handling of domain heterogeneity and nuisance variables.

For clinical and research applications, this work has meaningful implications. It suggests that multi-atlas consistency frameworks can improve diagnostic accuracy and robustness compared to single-atlas approaches, potentially increasing the clinical utility of fMRI-based biomarkers for conditions like Alzheimer's disease and ADHD. Developers of neuroimaging analysis pipelines could incorporate similar disentanglement strategies to handle the inherent variability in atlas selection. The competitive performance on established datasets indicates the method is practically viable for real-world deployment in research and clinical settings.

• →MADCLE learns disease representations that remain consistent across different brain atlases through distributional alignment rather than explicit feature fusion.
• →Structured disentanglement separates disease-related information from covariate and atlas-dependent factors, reducing confounding signal leakage.
• →The framework demonstrates competitive or improved performance on ADNI and ADHD-200 datasets compared to single-atlas and recent multi-atlas baselines.
• →This approach addresses a fundamental challenge in neuroimaging: functional connectivity heterogeneity caused by different brain parcellation schemes.
• →The work supports adoption of multi-atlas consistency methods in clinical neuroimaging pipelines for more robust disorder identification.