BrainAware Early Alzheimer's Detection

Hippocampus-focused VGG16 for CN / MCI / AD classification from T1 MRI

BrainAware takes a T1-weighted brain MRI, extracts the axial slices that contain the hippocampus, runs each slice through a fine-tuned VGG16 and averages the class probabilities. This page shows the viewer and the preprocessing on a few open-data scans.

Demo. Preprocessing approximated without FSL; open-data scans (healthy subjects), not ADNI. Research only, not for clinical use.

Prediction

Live predictions coming soon.

Axial
Coronal
Sagittal
3D drag to rotate
Loading mesh…
Click a view to move the other planes. Radiological convention (patient right on image left).

Model input

150×150, as fed to the model (same orientation as training)

What the model sees

Each axial slice in the hippocampus range is skull stripped, cropped to the brain and resized to 150×150. The model classifies each slice on its own and averages the probabilities. Use the slider or the axial view to step through the slices.

All model input slices for this case

About this demo

How it works

  1. Data. ADNI T1-weighted MRI, 1,598 subjects (339 AD, 452 CN, 807 MCI).
  2. Preprocessing. Register to MNI152, segment the hippocampus, skull strip, remove background, keep the slices that contain the hippocampus, crop to the brain, resize to 150×150.
  3. Model. VGG16 pretrained on ImageNet. First 3 conv layers frozen, the rest fine-tuned, new head with Dense 1024 → 512 → 3 (softmax).
  4. Training. Adam, exponential LR decay (0.96), class weights for the imbalance, progressive augmentation (strength increased in steps during training).
  5. Inference. Predict each slice, average the probabilities over slices, report CN / MCI / AD percentages.
Modified VGG16 architecture: input 150x150x3, VGG16 conv blocks, new classifier with 1024, 512 and 3 units
Modified VGG16. Figure from the paper.
The BrainAware web app: prediction donut chart, patient progression chart, three MRI planes and a 3D brain
The full app (React + Cornerstone + vtk.js, Django backend), with patient history across visits. Figure from the paper.

Paper results (ADNI)

SplitAccuracyLoss
Train99.62%0.019
Validation97.52%0.086
Test (10-fold CV)98.17%0.077
ClassPrecisionRecallF1
AD0.990.960.98
CN0.980.990.99
MCI0.970.990.98

These numbers are slice-level results on ADNI. Reported in the paper on the ADNI test set; they are not results on the scans shown on this page. External validation is still open work.

Team

Built by Mohamed Hafez, Noha Tarek, Caroline Magdy, Mohammed Kamel and Alaa Melek. Supervised by Ibrahim Sadek. Co-authors: Lamees N. Mahmoud and Shafiq Ul Rehman.

Mohammed Abdelhalim on the paper is Mohamed Hafez.

Citation

Rehman S.U., Tarek N., Magdy C., Kamel M., Abdelhalim M., Melek A., Mahmoud L.N., Sadek I. AI-based tool for early detection of Alzheimer's disease. Heliyon 10(8), e29375, 2024. doi:10.1016/j.heliyon.2024.e29375

@article{rehman2024alzheimer,
  title   = {AI-based tool for early detection of Alzheimer's disease},
  author  = {Rehman, Shafiq Ul and Tarek, Noha and Magdy, Caroline and Kamel, Mohammed and
             Abdelhalim, Mohammed and Melek, Alaa and Mahmoud, Lamees N. and Sadek, Ibrahim},
  journal = {Heliyon},
  volume  = {10},
  number  = {8},
  pages   = {e29375},
  year    = {2024},
  doi     = {10.1016/j.heliyon.2024.e29375}
}

Data and credits