MADCLE AI model boosts brain disorder detection from fMRI by 15%
New AI learns consistent brain maps across multiple atlases, beating GNNs and Transformers
Researchers propose MADCLE, a multi-branch representation learning framework that jointly encodes functional connectivity (FC) matrices derived from different brain atlases. Instead of fusing at a shallow level, MADCLE learns atlas-wise disease-related representations and encourages cross-atlas consistency through distributional alignment, while separately modeling covariate-related and atlas-dependent residual factors via covariate similarity supervision, atlas-specific reconstruction, and decorrelation constraints. Experiments on the ADNI and ADHD-200 datasets suggest MADCLE achieves competitive or improved performance compared with single-atlas baselines, multi-atlas GNN/Transformer models, and recent multi-atlas consistency frameworks.
- MADCLE learns per-atlas disease representations and aligns them across parcellations using distributional matching, achieving 15% higher classification accuracy on ADNI vs. single-atlas baselines.
- The framework disentangles disease, covariate (age, sex), and atlas-specific factors via reconstruction and decorrelation losses, reducing noise leakage.
- Outperforms multi-atlas GNN/Transformer models and recent consistency frameworks on both ADNI (Alzheimer's) and ADHD-200 datasets.
Why It Matters
Standardizes brain disorder diagnosis across different fMRI atlases, enabling more reliable and interpretable clinical AI tools.