CNNs for Alzheimer’s Detection

Inspiration

Alzheimer’s disease is a progressive neurological disorder that profoundly impacts patients, families, and healthcare systems. One of the major challenges lies in early detection and reliable assessment of disease severity, especially when visual differences between MRI scans can be subtle and difficult to interpret, even for trained clinicians.

We were inspired by a simple but powerful question :

Can deep learning models not only classify Alzheimer’s stages, but also help us understand how brain structures progressively diverge from a healthy pattern?

Rather than relying on a single modeling approach, we wanted to explore multiple perspectives :

Classical supervised learning, transfer learning with state-of-the-art architectures, and finally a self-supervised anomaly detection approach that does not require labels at all.

What it does

Our project analyzes brain MRI images to study Alzheimer’s disease through three complementary approaches :

1. Supervised CNN models trained from scratch to classify MRI scans into four stages :

Non-Demented
Very Mild Dementia
Mild Dementia
Moderate Dementia

2. Transfer Learning models (EfficientNet and ResNet) leveraging ImageNet pretraining, fine-tuned for medical imaging.

3. Self-supervised anomaly detection using an autoencoder trained only on healthy (Non-Demented) brains, enabling :

Detection of structural deviations
Visualization of anomaly maps highlighting regions that diverge from the learned healthy pattern

Together, these approaches allow both quantitative evaluation (classification performance) and qualitative interpretation (visual anomaly maps) of disease progression.

How we built it

1. Supervised CNN Baselines

We started with a simple convolutional neural network trained on grayscale MRI images (128×128).

From this baseline, we progressively improved performance by :

Adding data augmentation
Handling class imbalance with class weights
Applying oversampling
Performing hyperparameter tuning

Each model variant was evaluated independently to clearly measure incremental improvements.

2. Transfer Learning (EfficientNet & ResNet)

To push performance further, we implemented transfer learning using :

EfficientNetB0
ResNet50

Key design choices :

Conversion of grayscale MRI scans to RGB
Freezing pretrained layers initially
Gradual fine-tuning of deeper layers
Use of adaptive learning rate scheduling (ReduceLROnPlateau) and early stopping

This allowed us to benefit from rich pretrained features while adapting the models to medical imaging data.

3. Production-like Cross-Dataset Evaluation

To simulate real-world deployment, we tested all trained models on previously unseen MRI datasets.

This step revealed a critical insight :

Models with excellent validation accuracy can collapse when faced with distribution shifts.

This motivated the exploration of a fundamentally different paradigm.

4. Self-Supervised Anomaly Detection with Autoencoder

Instead of predicting labels, we trained a convolutional autoencoder only on healthy (Non-Demented) MRI scans.

The idea is simple :

Learn a compact representation of a healthy brain
Reconstruct healthy images accurately
Observe reconstruction errors when the brain structure deviates from normality

We used :

Keras Tuner to automatically search for the optimal encoder–decoder architecture
Mean Squared Error (MSE) as reconstruction loss
A statistically calibrated pixel-wise anomaly threshold based on healthy images

This enabled :

Pixel-level anomaly maps
Quantitative severity indicators (mean reconstruction error, anomaly ratio)
Clear visualization of progressive structural divergence across Alzheimer’s stages

Challenges we ran into

Dataset heterogeneity : MRI scans from different sources vary in contrast, resolution, and acquisition protocols.
Class imbalance : Severe stages are underrepresented, complicating supervised learning.
Generalization : High validation accuracy does not guarantee robustness across datasets.
Interpretability : Classification alone does not explain why a prediction is made.
Compute constraints : Efficient memory handling was required to avoid GPU out-of-memory errors.

Accomplishments that we are proud of

Built a complete experimental pipeline from baseline CNNs to advanced self-supervised learning.
Demonstrated the limits of pure supervised classification under dataset shift.
Achieved clear visual separation between Alzheimer’s stages using anomaly maps.
Designed an approach that is label-efficient, interpretable, and clinically intuitive.
Combined quantitative metrics and qualitative visual explanations in a single project.

What we learned

High accuracy does not necessarily imply robustness or clinical usefulness.
Transfer learning can significantly improve performance but still suffers from domain shift.
Self-supervised learning is a powerful alternative when labels are scarce or unreliable.
Autoencoders can reveal progressive structural changes without explicit supervision.
Visualization is crucial for trust and interpretability in medical AI.