ALZ detector

Inspiration

Alzheimer's disease affects millions worldwide, yet early and accurate diagnosis remains a challenge. We were inspired to leverage the power of AI to support clinicians in detecting Alzheimer’s from medical data—both structured clinical features and MRI images—to enable earlier intervention and improve patient outcomes.

What it does

Our project is a dual-model AI system for Alzheimer’s disease classification.

The first model uses structured clinical and genomic data (130 features) to classify disease stages into 9 categories with a custom PyTorch neural network.
The second model analyzes MRI scans with a Convolutional Neural Network (CNN) to detect visual biomarkers of Alzheimer’s.
Together, they provide a multi-modal diagnostic aid that can assist medical professionals in making more informed decisions.

How we built it

We built two pipelines:

Structured Data Pipeline:
- Loaded preprocessed .npz files containing clinical and genomic features.
- Engineered a 3-layer neural network using PyTorch, with normalization and robust handling of missing values.
- Trained and evaluated the model, achieving 99.84% test accuracy.
MRI Image Pipeline:
- Loaded brain MRI data from Parquet files, extracted byte images, and preprocessed them (resize, normalize, reshape).
- Built a CNN using TensorFlow/Keras with convolutional, pooling, and dense layers.
- Trained on grayscale MRI images and achieved 91.64% test accuracy.

Challenges we ran into

Data Quality Issues: Encountered NaN and infinite values in the structured dataset, causing training instability.
Image Preprocessing Complexity: MRI byte data required careful extraction, format handling, and normalization to avoid PIL deprecation warnings and array shape errors.
Model Convergence: Initial training runs resulted in nan loss due to unscaled data and high learning rates.
Memory & Speed: Processing large MRI images required efficient resizing and batch operations to fit within Colab’s memory limits.

Accomplishments that we're proud of

Successfully built and trained two distinct AI models for Alzheimer’s classification using different data modalities.
Achieved strong performance: 99.84% accuracy on structured data and 91.64% accuracy on MRI images.
Implemented robust data preprocessing pipelines that handle real-world data issues gracefully.
Visualized results with confusion matrices and training history plots to interpret model behavior clearly.

What we learned

The importance of thorough data cleaning and normalization for model stability.
How to handle medical image data stored in Parquet format and convert byte streams into usable arrays.
Techniques for diagnosing and fixing training issues like nan loss through learning rate adjustment and data scaling.
The value of multi-modal approaches in medical AI—combining different data types can lead to more reliable systems.

What's next for Untitled

Model Fusion: Combine the structured and image models into an ensemble for improved accuracy and robustness.
Explainability: Integrate tools like SHAP or Grad-CAM to provide interpretable insights for clinicians.
Deployment: Package the system into a lightweight web or mobile interface for easy use in clinical settings.
Dataset Expansion: Train on larger, more diverse datasets to improve generalization across populations.
Real-Time Processing: Optimize the pipeline for faster inference to support real-time diagnostic assistance.