Researchers Introduce M2LG-DG Framework for Depressive Disorder Classification

Researchers have introduced M2LG-DG, a multimodal domain generalization framework designed to classify major depressive disorder (MDD) across disparate clinical imaging sites using resting-state functional magnetic resonance imaging (rs-fMRI). The work addresses a common challenge in neuroimaging AI, where scanner hardware and protocol variations degrade model performance on unseen data.

The proposed architecture leverages a dual-stream rs-fMRI encoder combining a global self-attention mechanism with local graph-constrained aggregation. It fuses imaging and non-imaging representations through bidirectional cross-attention with modality gating, alongside a cross-site supervised contrastive objective that aligns diagnostic features across acquisition domains.

Evaluated on four held-out REST-meta-MDD sites, M2LG-DG attained an area under the curve (AUC) of 69.48%, outperforming baseline methods by 2.18 percentage points. Additional experiments on the Autism Brain Imaging Data Exchange (ABIDE) dataset confirmed its broader applicability to psychiatric neuroimaging tasks.