SSL for Alzheimers

Website Dashboard
Project Architecture
Cross-Domain Generalization Test Accuracy comparison Graph
Sample images
Tumor classification with Grad-CAM analysis
Alzheimer classification with Grad-CAM analysis
MRI upload page
Results Webpage
Confusion Matrix for Tumor
Confusion Matrix for Alzheimers
Radar chart for Metrics across different Alzheimer classes

Inspiration

Neurodegenerative diseases like Alzheimer’s affect millions of people worldwide, yet early diagnosis remains difficult due to the lack of large, well-annotated medical imaging datasets. MRI scans contain rich structural information, but labeling them requires expert clinicians, making supervised learning approaches expensive, slow, and difficult to scale.

We were inspired by a simple question: how can we leverage the vast amount of unlabeled brain MRI data to build accurate, trustworthy AI systems for brain disease analysis?
This led us to explore self-supervised learning (SSL) as a way to reduce dependency on labels while still learning meaningful, clinically relevant representations.

What it does

SSL for Alzheimer’s is a self-supervised learning framework that learns powerful representations from unlabeled brain MRI scans and uses them to classify neurological diseases with very limited labeled data.

The model is pretrained using SimCLR on large unlabeled MRI datasets and then fine-tuned to:

Classify Alzheimer’s disease into four stages
Detect brain tumors
Identify Parkinson’s disease
Classify multiple sclerosis

To improve trust and transparency, the system also generates Grad-CAM visualizations that highlight which regions of the brain influenced each prediction.

How we built it

Collected and preprocessed large-scale unlabeled brain MRI slices
Designed medically safe data augmentations to preserve anatomical features
Implemented SimCLR with a from-scratch developed ResNet-50 backbone for self-supervised pretraining
Fine-tuned the pretrained encoder using very small labeled datasets (as few as 450 images per class)
Integrated Grad-CAM for visual explanations and added quantitative explainability metrics
Evaluated performance across multiple brain disease classification tasks

All experiments were implemented in PyTorch and executed through a single reproducible notebook.

Challenges we ran into

Adapting self-supervised learning techniques to medical imaging without destroying clinically important information
Designing augmentations that improve representation learning while preserving diagnostic signals
Handling class imbalance and limited labeled samples
Ensuring model interpretability aligned with known neurological patterns
Balancing model accuracy with computational efficiency for real-world feasibility

Accomplishments that we're proud of

Achieved strong performance using minimal labeled data across multiple brain diseases
Demonstrated that SSL representations generalize across neurological conditions
Outperformed a fully supervised baseline in low-label settings
Integrated explainability tools that provide clinically meaningful insights
Built a scalable, data-efficient pipeline suitable for real-world medical imaging scenarios

What we learned

Self-supervised learning is highly effective for medical imaging tasks where labeled data is scarce
Representation quality matters more than model complexity in low-label regimes
Explainability is critical for trust in healthcare AI systems
Medical AI requires careful engineering choices beyond standard computer vision pipelines
Research-driven methods can be successfully adapted into practical, hackathon-ready solutions

What's next for SSL for Alzheimers

Extend the framework to additional neurological and neurodegenerative diseases
Incorporate longitudinal MRI data to model disease progression
Explore other SSL techniques such as masked autoencoders
Improve clinical validation with multi-institutional datasets
Integrate the model into clinician-facing tools for real-time decision support using RAG