Inspiration

Alzheimer's disease affects over 55 million people worldwide, with numbers projected to triple by 2050. What struck us most is that early detection can significantly improve patient outcomes — yet current diagnostic methods rely heavily on subjective clinical assessments and expert radiologists, who are scarce in many regions.

We were inspired by the potential of AI to democratize healthcare access. Imagine a screening tool that could help identify early signs of cognitive decline in underserved communities, giving patients and families precious time to plan and seek treatment.

This hackathon gave us the opportunity to build something that could genuinely save lives.

What It Does

Our AI model classifies brain MRI scans into four stages of Alzheimer's disease progression:

Stage	Label	Description
🟢	NonDemented	Healthy brain with no cognitive impairment
🟡	VeryMildDemented	Very early stage — subtle changes detectable
🟠	MildDemented	Mild cognitive impairment (MCI)
🔴	ModerateDemented	Moderate-stage Alzheimer's disease

Key Capabilities

Automated Classification
Upload an MRI scan and receive instant predictions with confidence scores.
Explainable AI
Grad-CAM visualizations show exactly which brain regions influenced the prediction.
Class Imbalance Handling
Robust performance even on rare disease stages.
Reproducible Pipeline
Fully documented Jupyter notebook runnable on Google Colab.

How We Built It

Architecture

We used EfficientNet-B0 with transfer learning, a state-of-the-art CNN architecture known for its optimal accuracy-to-parameter efficiency.
The model was pre-trained on ImageNet and fine-tuned on our medical imaging dataset.

Tech Stack

Framework: PyTorch 2.0+
Model: EfficientNet-B0 with a custom classification head
Explainability: Grad-CAM
Data Processing: torchvision transforms with augmentation
Evaluation: scikit-learn metrics (Accuracy, Precision, Recall, F1-score, AUC-ROC)

Training Strategy

Optimizer: AdamW with weight decay
$$ \lambda = 10^{-4} $$
Learning Rate:
$$ 10^{-4} $$
with ReduceLROnPlateau scheduler
Class Imbalance:
- Weighted CrossEntropy loss
- Weighted random sampling
Regularization:
- Dropout
- Early stopping (patience = 7)
- Gradient clipping

Challenges We Ran Into

1. Severe Class Imbalance

The dataset was extremely skewed:

ModerateDemented: 64 samples (~1%)
NonDemented: 3,200 samples (~50%)

Solutions:

Inverse-frequency class weights in the loss function
Weighted random sampling during training
Strategic data augmentation

2. Medical Imaging Constraints

MRI scans differ significantly from natural images:

Grayscale images with subtle anatomical variations
Over-aggressive augmentation risks anatomical distortion

Adjustments:

Carefully tuned augmentation parameters
Normalization using ImageNet statistics for transfer learning compatibility

3. Interpretability Requirements

Medical AI must be explainable.

Implemented Grad-CAM from scratch
Visualized attention over clinically relevant regions:
- Hippocampus
- Temporal lobe
- Ventricles

This ensured predictions were clinically meaningful, not artifact-driven.

4. Reproducibility

Ensuring the notebook runs seamlessly on both local machines and Google Colab required:

Careful dependency management
Robust path handling
Fixed random seeds

Accomplishments We’re Proud Of

✅ High Performance
Strong classification accuracy with robust minority-class handling
✅ Explainable AI
Grad-CAM confirms focus on clinically relevant brain regions
✅ Production-Ready Code
Modular architecture separating data loading, modeling, training, and evaluation
✅ Comprehensive Documentation
- Model Card with bias and fairness considerations
- Technical report covering methodology and limitations
- Fully documented Jupyter notebook
✅ Ethical Considerations
Clearly defined as a screening aid, not a diagnostic tool, with responsible-use guidelines

What We Learned

Technical

Transfer Learning Works
Pre-trained models dramatically reduce data requirements for medical imaging tasks
Class Imbalance Strategies
Combining weighted loss and weighted sampling outperforms either alone
Grad-CAM Internals
Building explainability from scratch deepened our understanding of CNNs

Domain

Medical AI Responsibility
False negatives can delay treatment, while false positives can cause unnecessary anxiety
Alzheimer’s Pathology
Understanding hippocampal atrophy and ventricular enlargement helped validate model attention

Process

Documentation Matters
A well-documented notebook is more valuable than marginal accuracy gains
Reproducibility First
Fixed seeds and multi-environment testing saved significant debugging time