Brain Tumor Classification with Grad-CAM Visualization

This Streamlit application enables users to upload brain MRI images and perform tumor classification using a pre-trained ResNet18 model, with visualizations provided by various Grad-CAM techniques. The app supports four tumor classes: glioma, meningioma, no tumor, and pituitary.

Features

• Image Upload: Upload brain MRI images in JPG or PNG format.
• Model Prediction: Uses a pre-trained ResNet18 model to classify the uploaded image into one of four tumor classes.
• Grad-CAM Visualization: Displays Grad-CAM heatmap for the predicted class, highlighting important regions for the model's decision.
• Multi-Target Grad-CAM: Visualizes Grad-CAM heatmaps for all four classes alongside their predicted probabilities.
• CAM Methods Comparison: Compares visualizations from different CAM methods (GradCAM, GradCAM++, XGradCAM, EigenCAM) for the predicted class.
• Responsive Design: Adapts visualization sizes based on the browser window width.

Requirements

• Python 3.8+
• Streamlit
• PyTorch
• OpenCV
• NumPy
• Matplotlib
• PIL (Pillow)
• torchvision
• pytorch-grad-cam
• streamlit-js-eval

Install dependencies using:

pip install streamlit torch opencv-python numpy matplotlib pillow torchvision pytorch-grad-cam streamlit-js-eval

Setup

1. Model File:

   • Place your pre-trained ResNet18 model file (brain_tumor_resnet18.pth) in the project directory.
   • Update the model_path variable in app.py if the model file is located elsewhere.

2. Directory Structure:

   project_directory/
   ├── app.py
   ├── brain_tumor_resnet18.pth
   └── README.md
   

3. Run the Application:

   streamlit run app.py

Usage

1. Open the app in your browser (typically at http://localhost:8501).
2. Upload a brain MRI image (JPG or PNG).
3. View the results:
   • Uploaded Image: Displays the original MRI image.
   • Grad-CAM for Predicted Class: Shows the Grad-CAM heatmap for the model's predicted tumor class.
   • Multi-Target Grad-CAM: Displays heatmaps for all four classes with their respective probabilities.
   • Comparison of Different CAM Methods: Compares visualizations from GradCAM, GradCAM++, XGradCAM, and EigenCAM.

Code Overview

• Model Loading: Loads a pre-trained ResNet18 model with a modified fully connected layer for 4-class classification.
• Image Preprocessing: Resizes images to 224x224, normalizes them, and converts them to tensors for model input.
• Grad-CAM Implementation: Uses the pytorch_grad_cam library to generate heatmaps for model interpretability.
• Visualization:
  • Single-class Grad-CAM for the predicted class.
  • Multi-class Grad-CAM for all classes with probability scores.
  • Comparison of different CAM methods for the predicted class.
• Streamlit Interface: Provides an interactive UI for image upload and result visualization.

Notes

• Ensure the model file path is correct; otherwise, the app will stop with an error.
• The app uses the last convolutional layer (layer4[-1].conv2) of ResNet18 for Grad-CAM visualizations.
• Visualizations are normalized for display and use RGB format.
• The app dynamically adjusts figure sizes based on browser window width for better viewing.
• CUDA is used if available; otherwise, it falls back to CPU.

Limitations

• The model must be trained and saved as brain_tumor_resnet18.pth in the specified format.
• Only JPG and PNG images are supported.
• The app assumes the input images are brain MRIs; other image types may lead to unreliable predictions.
• Visualization quality depends on the input image resolution and model performance.

Troubleshooting

• Model Not Found: Verify the brain_tumor_resnet18.pth file exists in the specified path.
• Dependency Issues: Ensure all required packages are installed with compatible versions.
• CUDA Errors: If GPU is unavailable or misconfigured, the app automatically uses CPU.
• Visualization Issues: Check browser compatibility or adjust the page_width calculation if figures appear misaligned.