Hugging Face's logo

Spaces:
saifulislamoni
/
SSL-Former
Sleeping

App Files Community
SSL-Former / README.md
saifulislamoni's picture
saifulislamoni
Deploy: Brain Tumor Classifier with GradCAM and PDF reports - MoCo SSL + Swin-Transformer
c703689
preview code
|
raw
history blame contribute delete
3.7 kB
metadata 
title: Brain Tumor Classifier
emoji: 🧠
colorFrom: blue
colorTo: green
sdk: docker
pinned: false

🧠 Brain Tumor Classification AI

AI-powered MRI analysis for brain tumor detection and classification

Using MoCo Self-Supervised Learning with Swin-Transformer backbone for high-accuracy tumor classification.

✨ Features

  • 🎯 High Accuracy: ~95% on test dataset
  • 🧠 4-Class Classification: Glioma, Meningioma, No Tumor, Pituitary
  • 📊 GradCAM Visualization: See which regions influenced the prediction
  • 📄 PDF Reports: Download professional analysis reports
  • 📱 Responsive Design: Works on desktop, tablet, and mobile

🚀 Quick Start

  1. Upload MRI Image

    • Click upload area or drag & drop
    • Formats: JPEG, PNG, BMP, TIFF
    • Max size: 16 MB

  2. View Results

    • Classification with confidence score
    • Probability distribution for all classes
    • GradCAM visualization showing influential regions

  3. Download Report

    • Enter your name
    • Get professional PDF report
    • Includes all visualizations

🔬 Model Details

Architecture:

  • Backbone: Swin-Transformer (ImageNet pre-trained)
  • Pre-training: MoCo (Momentum Contrast) - 30 epochs
  • Fine-tuning: 2-stage process (30 epochs total)
  • Head: Linear classifier (192 → 4 classes)

Training:

  • Dataset: 17,784 MRI images
  • Classes: 4 (balanced)
  • Validation: 10% held-out
  • Test: 10% locked evaluation

Performance:

  • Accuracy: ~95%
  • Precision: ~95% (weighted)
  • Recall: ~95% (weighted)
  • F1-Score: ~95% (weighted)

📊 Class Information

Class Description
Glioma Aggressive tumor from glial cells
Meningioma Tumor from brain membranes
No Tumor Healthy brain with no abnormalities
Pituitary Tumor in pituitary gland

🎓 About GradCAM

The GradCAM (Gradient-weighted Class Activation Map) visualization shows:

  • Which regions of the MRI influenced the model's decision
  • Warmer colors (red/yellow) = higher importance
  • Cooler colors (blue) = lower importance

This helps build trust in AI predictions by showing what the model "looked at" when making its decision.

⚙️ Technical Stack

  • Framework: Flask
  • Deep Learning: PyTorch
  • Model: Swin-Transformer
  • Visualization: GradCAM
  • Reports: ReportLab
  • Frontend: HTML5, CSS3, JavaScript

⏱️ Performance

Operation Time
Model Load ~10-15s (first time)
Classification ~2-3s
GradCAM Included in classification
PDF Report ~2-3s
Total ~15-20s per request

🛡️ Data Privacy

  • ✅ Images are NOT stored
  • ✅ Processed in memory only
  • ✅ Temporary files cleaned up
  • ✅ No data collection
  • ✅ Local processing (no external API calls)

⚠️ Disclaimer

This tool is for research and educational purposes only.

This AI model is NOT a substitute for professional medical diagnosis. All results should be verified by qualified healthcare professionals. Always consult a radiologist or medical doctor for clinical interpretation of MRI scans.

📚 Learn More

  • Paper: MoCo: Momentum Contrast for Unsupervised Visual Representation Learning
  • Swin-Transformer: Shifted Windows Transformer for CV

🤝 Contributing

Feedback and suggestions welcome!

For issues or improvements:

  1. Note the problem
  2. Share details about the image/results
  3. Report via GitHub Issues

📄 License

Academic use - Research and educational purposes

Created with ❤️ for AI in Medical Imaging

Last Updated: 2026-04-20