AI Now Spots Depression in Brain Scans, Even at a New Hospital
Brain scans vary by machine — this AI ignores that and still finds depression
Researchers developed M2LG-DG, a framework that classifies major depressive disorder across imaging sites using resting-state fMRI. The problem it targets: these models often show lower performance at sites not included during development, which can limit their use in clinical settings. M2LG-DG combines a dual-stream encoder with imaging and non-imaging data and a cross-site contrastive objective. On four held-out REST-meta-MDD sites, it reached an AUC of 69.48%, beating the closest comparison method by 2.18 percentage points. Experiments on the ABIDE dataset further support its use for other psychiatric neuroimaging tasks. This is research, not a clinical diagnosis tool.
- The AI combines brain scans with basic patient info like age and sex to spot signs of depression
- It scored about 70 out of 100 on a standard accuracy measure, beating the next-best method by roughly 2 points
- It still worked on hospitals it had never seen before — the exact problem that has stalled this kind of research
Why It Matters
Could one day give doctors a fast, objective depression check that works no matter which hospital or scanner you visit.