BrainGaurd

Early Alzheimer’s detection from MRI, with clear visual explanations.

Comment

BrainGuard — MRI-Based Alzheimer’s Detection

About the Project

BrainGuard is a Flask + PyTorch web app that classifies brain MRI scans into four dementia severity levels (Non-Demented, Very Mild, Mild, Moderate) and provides Grad-CAM visual explanations to highlight the regions influencing the prediction. Built for the Hack4Health AI for Alzheimer’s challenge, the system emphasizes interpretability, reliability, and ease of use.

What It Does

  • Upload MRI images (PNG/JPG or NIfTI); automatically handles middle-slice extraction for 3D volumes.
  • ResNet50 transfer-learning model delivers high accuracy across four dementia classes.
  • Grad-CAM overlays show the model’s attention for transparent, explainable AI.
  • Web UI for drag-and-drop uploads, probability breakdowns, and confidence display.

Tech Stack

  • Backend: Flask, Python, PyTorch (ResNet50), Torchvision
  • Explainability: Grad-CAM
  • Data Handling: Pillow, NumPy, (optional) NiBabel for NIfTI
  • Frontend: HTML/CSS/JS (vanilla), Font Awesome

How We Built It

  1. Data Prep: Grayscale conversion, resize to 224×224, normalization (ImageNet stats). For NIfTI, extract the middle axial slice.
  2. Model: ResNet50 with single-channel first conv and a 4-class classifier head.
  3. Training: CrossEntropyLoss + Adam (lr=1e-3), ReduceLROnPlateau scheduler, batch size 32, early-stopping mindset.
  4. Explainability: Grad-CAM on the final conv block to generate saliency overlays.
  5. App: Flask API for inference; web UI for uploads, predictions, probabilities, and attention maps.

Results

  • Strong validation performance with clear saliency maps aligning to clinically relevant regions (e.g., hippocampal areas).
  • Fast inference on CPU; optional GPU support if available.

How to Run (Local)

python -m venv venv
# On Windows: venv\Scripts\activate
pip install -r requirements.txt
python app/app.py
# Open http://localhost:8000

Challenges

  • Handling diverse MRI formats (2D and 3D) with consistent preprocessing.
  • Ensuring interpretability via Grad-CAM while keeping inference fast.

What’s Next

  • Full 3D model support (volumetric CNNs / transformers).
  • Domain adaptation for multi-scanner generalization.
  • Calibration and threshold tuning for deployment contexts.
  • Additional explainability methods (e.g., Integrated Gradients).

Team

BrainGuard — Hack4Health AI for Alzheimer’s Challenge

Built With

Share this project:

Updates