Explainable Alzheimer's screening from a single MRI scan
A research system that takes a routine T1 MPRAGE brain scan and produces a diagnosis estimate, a conversion-risk forecast, and a visual explanation of which brain regions drove the prediction — built as a research project, not a medical product.
What Alzheimer's actually does to the brain
Alzheimer's disease progressively kills neurons, starting in specific structures years before memory symptoms become obvious. Abnormal protein buildup (amyloid plaques outside neurons, tau tangles inside them) disrupts cell function and eventually causes tissue loss you can see directly on a structural MRI — the hippocampus and entorhinal cortex shrink first, then surrounding cortex thins and the fluid-filled ventricles expand to fill the space. By the time someone is diagnosed clinically, these changes have often been building for years.
That's the opening this project is built around: if the structural signature is already visible on a routine scan, a model trained to recognize it can flag risk earlier than symptoms alone would — and, just as important, show which regions drove that conclusion, instead of returning an unexplained number.
This is a research prototype. It is not FDA/CE cleared, has not been validated prospectively, and should never be the sole basis for a clinical decision — see the disclaimer above on every page.
How it works, end to end
An uploaded scan goes through N4 bias correction, deep-learning skull stripping (HD-BET), and registration to a standard template, then into a 3D convolutional network fused with structural biomarkers (hippocampal volume, cortical thickness, ventricular volume, and similar FreeSurfer measures). A separate discrete-time survival model estimates conversion risk over time and acts as a second opinion on the binary conversion classifier. Both outputs come with a confidence estimate from repeated stochastic inference, and a Grad-CAM-style heatmap shows which regions the classification model weighted most heavily for that specific scan.
Where it stands today
Measured on a held-out ADNI test set; see live model validation for the full confusion matrices and current numbers.
balanced accuracy
balanced accuracy
concordance index
These are the numbers this deployed tool actually produces on a single uploaded scan, from the 3-seed biomarker-fusion ensemble — the exact checkpoints the server loads, the same figures shown on the live validation page. A separate, clinician-only tool estimates MCI→AD conversion risk from cognitive test scores alone (no scan) and scores higher in testing (0.788 balanced accuracy) — see the paper for why cognition, not imaging, carries most of the signal for that specific task.
The conversion model is trained on a small cohort (170 subjects) — treat that number as exploratory, not diagnosis-grade. A plasma-biomarker extension (p-tau181, NfL) was tested and reverted after it measurably hurt both tasks; that negative result is documented in the project history rather than hidden.
What the model actually highlights
A real result from a real scan run on this exact deployment — not a mockup, not training data, not cherry-picked for a clean-looking output.
Self-consistency check: repeating the computation with different random dropout patterns moved this prediction by about ±6.0 percentage points. Of the regions the model focused on most for this one scan, 4.7% fell inside the hippocampus — versus about 0.5% expected by pure chance. That's one example, not the typical case; see model validation for the aggregate figure across the full test set.
Where it was looking
Brighter (yellow/red) areas mattered more to the model's decision, faint areas mattered little. Try your own scan to see this generated on demand.
What's next
- Prospective validation on scans and sites not seen during training — the current numbers are all retrospective ADNI test-set performance.
- A written paper covering the full methodology, ablations, and statistical significance testing (DeLong's test / bootstrap confidence intervals), not just point estimates.
- Broader explainability comparison — the current heatmap is Grad-CAM-style; SHAP and integrated gradients are being evaluated against it for which gives clinicians the most useful signal.
- Additional cohorts beyond ADNI, to check whether performance holds outside the population the model was trained on.
About
Built by Bahrom Ashurov as an independent research project combining 3D medical imaging, structural biomarker fusion, and explainability methods for early Alzheimer's detection. The code is MIT-licensed and open; the ADNI imaging and clinical data it's trained on is not redistributed here and is governed by ADNI's own Data Use Agreement — anyone extending this work needs their own DUA-approved access.