MPRAGE Alzheimer’s Detection
Research use only. This tool is not a certified medical device and has not been reviewed or cleared by any regulatory authority (FDA, CE, or otherwise). Outputs are model predictions on a research dataset (ADNI) and must not be used as the sole basis for any clinical diagnosis or treatment decision. No account is required and this is a demonstration surface only — uploaded files are deleted automatically after a result is shown.

Model validation

Aggregate numbers from the held-out test set and a same-session repeat-scan reliability check, so you can judge overall trustworthiness before trusting any one prediction. See the project README and paper for full methodology.

Diagnosis (CN / MCI / AD), test set

Total accuracy: 83.1%

Balanced accuracy: 0.839

Predicted CNPredicted MCIPredicted AD
True CN178179
True MCI3015535
True AD38168

MCI → AD conversion risk, test set

Trained on a small cohort (170 subjects) — treat as exploratory, not diagnosis-grade.

Balanced accuracy: 0.752

Predicted stablePredicted converter
True stable10514
True converter2236

Survival model (conversion timing), separate ensemble

A separate discrete-time survival model, evaluated on its own time-to-event task rather than the binary stable/converter split above.

Concordance index: 0.775 (3-seed ensemble)

Test-retest reliability

Same person, same session, two MRI acquisitions ~8 minutes apart — does the model agree with itself?

Agreement rate: 0.870 across 376 repeat-scan pairs.

Interpretability

Grad-CAM attention checked against a hippocampus mask — the region AD is known to affect first. Higher overlap than the mask's share of total brain volume suggests the model is using anatomically relevant signal, not just any texture correlated with the label.

Sample scan: mean 1.63% across 59 test-set scans, vs. 0.54% chance level (the mask's share of the cropped volume) -- roughly 3x chance, though modest in absolute terms