Validated screening workflow

Early Alzheimer prediction from brain MRI

We compared deterministic and CNN-based paths, then selected a two-stage MLP because it produced the lowest missed-dementia rate while keeping alpha in the single digits at the recommended threshold.

MRI image pixels + regions feature vectors PCA two-stage MLP

Very mild demented MRI example — Very Mild

Healthy brain and typical Alzheimer's disease brain MRI comparison — **Medial temporal context:** hippocampal atrophy is assessed on lower axial or coronal views, as in this reference, not on the high axial slice above.

Domain knowledge guides features

Match each feature to anatomy the slice actually shows

Hover or select an indicator to locate its image region. These are numerical proxies, not clinical segmentations.

Important: the hippocampi and medial temporal lobes sit lower and more medial. They cannot be identified reliably in this high axial slice.

Recommended model

Two-stage MLP: minimize beta while keeping alpha low

The model first detects dementia, then assigns severity. Threshold 0.50 gives the lowest beta among the operating points with single-digit test alpha.

Selected threshold0.50

MRI1,110 deterministic features120 PCA values128 -> 64 dementia MLP96 -> 48 severity MLP

Stage 1 estimates dementia probability from pixel, texture, asymmetry, and coarse region summaries. Stage 2 separates very mild from mild disease only after dementia is predicted.

Test sensitivity96.82%

Beta3.18%

Alpha7.71%

Validation MRI features projected onto two linear discriminant analysis directions — **LDA view.** Class overlap limits a straight three-class boundary to 56.1% balanced accuracy.

Validation and test ROC curves for the two-stage MLP dementia detector — **ROC/AUC.** Validation AUC is 0.993; the test audit is 0.989. Dots mark threshold 0.50.

Exploratory model interpretation

The MLP relies most on broad cortical information

One coarse area at a time was replaced by its training-mean image patch. A larger ROC-AUC drop means the fitted MLP lost more useful signal.

Validation ROC-AUC decrease after occluding five coarse brain MRI regions — Paired bootstrap intervals use the same 950 validation images. All five areas produced a positive performance drop.

Strongest model reliance

Upper and lower cortical proxies

9.3-9.7 AUC points lost

Broad outer-brain texture and atrophy patterns produced the largest measured performance change.

Moderate model reliance

Central / ventricle zone

5.7 AUC points lost

Central fluid-space size and surrounding intensity added useful, but smaller, discrimination.

Smaller model reliance

Lower-lateral proxies

2.6-3.1 AUC points lost

Both sides helped, but less than the broad cortical and central regions in this slice-level model.

MLP threshold comparison

Zero misses is not useful when alpha reaches 68%

Beta is the primary objective, but alpha still prevents the model from flagging nearly everyone. The middle point preserves most sensitivity without that failure.

Three two-stage MLP operating points comparing alpha and beta on test — Lower-left is better. Thresholds were chosen on validation; the chart shows their observed test behavior.

Rejected extreme

Maximum sensitivity

Beta 0.00%Alpha 68.13%

Zero misses is achieved by flagging 327 of 480 non-demented images.

Recommended

Threshold 0.50

Beta 3.18%Alpha 7.71%

Only 15 of 471 dementia images are missed while alpha remains single-digit.

Lower-alpha option

Threshold 0.748

Beta 4.25%Alpha 5.83%

Nine fewer false alarms, but five additional missed dementia images.

Recommended operating point

Use the two-stage MLP at threshold 0.50

The primary goal is beta reduction. Alpha is monitored as a secondary cost rather than enforced as a hard target.

Recommended

Threshold 0.50

Sensitivity96.82%

Beta3.18%

Alpha7.71%

Pred. nonPred. dementia Actual non44337 Actual dementia15456

Lower-alpha option

Threshold 0.748

Sensitivity95.75%

Beta4.25%

Alpha5.83%

Pred. nonPred. dementia Actual non45228 Actual dementia20451

Threshold diagnostics

The selected threshold keeps beta low without extreme alpha

Three MLP operating points comparing alpha and beta — **Operating-point tradeoff.** Zero beta requires unusably high alpha. Threshold 0.50 retains most sensitivity with single-digit alpha.

Test false alarms and missed dementia counts for two MLP thresholds — **Observed test errors.** The lower-alpha threshold saves 9 false alarms but creates 5 additional misses.