Brain Tumor MRI Classification
Transfer learning with ResNet18 to classify brain MRI scans into three tumor types, with Grad-CAM to show where the model looks.
- Built with
- Python, PyTorch
- Source
- github ↗
This project uses transfer learning with ResNet18 to classify MRI brain scans into three types of brain cancer: Glioma, Meningioma, and General Tumor. Built with PyTorch and trained on a curated dataset of 6,000+ labeled medical images, the model achieves strong performance and interpretability using Grad-CAM.
Highlights
- Real-world medical dataset from hospitals (6,056 MRI scans)
- Transfer learning with ResNet18 pretrained on ImageNet
- Achieved ~98% accuracy on unseen test MRI images
- Grad-CAM visualizations to highlight what the model focuses on
- Built and trained entirely in Google Colab with GPU acceleration
- Clean training pipeline with train/val/test splits and augmentation
Dataset
-
Classes:
brain_glioma/brain_menin/brain_tumor/
-
All images resized to 224x224, normalized using ImageNet statistics
Technologies Used
- Python, PyTorch, torchvision
- Scikit-learn, Seaborn, Matplotlib
- Google Colab (GPU)
Results
| Metric | Score |
|---|---|
| Test Accuracy | ~98% |
| Precision | ~98% |
| Recall | ~98% |
| F1-Score | ~98% |
Grad-CAM Example
Model focuses attention on tumor regions, validating spatial learning.
Future Improvements
- Try ResNet50 or EfficientNet for higher accuracy
- Try applying Learning Rate Scheduling
Citation
Rahman, Md Mizanur (2024), "Brain Cancer - MRI dataset", Mendeley Data, V1, doi: 10.17632/mk56jw9rns.1
https://medium.com/@stepanulyanin/implementing-grad-cam-in-pytorch-ea0937c31e82