Brain Tumor MRI Classification

A comprehensive Jupyter notebook project for classifying brain tumor MRI scans into four categories: Glioma, Meningioma, Pituitary, or No Tumor.

Overview

This project demonstrates end-to-end deep learning for medical image classification, featuring:

• Data exploration and preprocessing
• Two neural network architectures (transfer learning and custom CNN)
• Model training, evaluation, and visualization
• Interactive web application with AI-powered explanations
• Deployment capabilities

Features

• Data Analysis: Exploratory data analysis with class distribution and sample visualization
• Model Training: Train and compare Xception (transfer learning) and custom CNN models
• Evaluation: Comprehensive metrics, confusion matrices, and classification reports
• Web App: Streamlit interface for real-time classification with saliency maps
• AI Explanations: Google Gemini integration for model decision explanations
• Deployment: Ngrok tunneling for public access

Dataset

The project uses the Brain Tumor MRI Dataset from Kaggle, containing:

• Glioma tumors
• Meningioma tumors
• Pituitary tumors
• No tumor (healthy scans)

Requirements

• Python 3.7+
• Jupyter Notebook or Google Colab
• Required packages:
  • tensorflow
  • keras
  • pandas
  • numpy
  • matplotlib
  • seaborn
  • scikit-learn
  • opencv-python
  • pillow
  • streamlit
  • google-generativeai
  • python-dotenv
  • pyngrok
  • plotly

Installation

1. Clone the repository:

   git clone https://github.com/anbguye/Brain-Tumor-MRI-Classification.git
   cd Brain-Tumor-MRI-Classification

2. Install dependencies:

   pip install tensorflow pandas numpy matplotlib seaborn scikit-learn opencv-python pillow streamlit google-generativeai python-dotenv pyngrok plotly

3. Set up API keys:

   • Create a .env file in the project root
   • Add your Google Gemini API key: GOOGLE_API_KEY=your_api_key_here
   • Add your Ngrok token: NGROK_AUTH_TOKEN=your_token_here

How to Run

Option 1: Google Colab (Recommended)

1. Upload Brain_Tumor_Classification.ipynb to Google Colab
2. The notebook includes automatic dataset download via Kaggle API
3. Run all cells sequentially
4. The Streamlit app will launch with a public URL via Ngrok

Option 2: Local Environment

1. Download the dataset from Kaggle and extract to Training/ and Testing/ folders
2. Open Brain_Tumor_Classification.ipynb in Jupyter Notebook
3. Update file paths in the notebook if necessary
4. Run the cells to train models and launch the web app

Project Structure

The notebook is divided into two main parts:

Part 1: Model Development

• Dataset loading and preprocessing
• Data augmentation and generators
• Xception model training and evaluation
• Custom CNN model training and evaluation
• Model comparison and visualization

Part 2: Web Application

• Streamlit app code for interactive classification
• Saliency map generation
• Google Gemini integration for explanations
• Ngrok deployment

Model Architectures

Xception (Transfer Learning)

• Input: 299x299 RGB images
• Base Model: Xception pretrained on ImageNet
• Top Layers: Flatten → Dropout(0.3) → Dense(128, ReLU) → Dropout(0.25) → Dense(4, Softmax)
• Training: Adamax optimizer, 0.001 learning rate, 5 epochs

Custom CNN

• Input: 224x224 RGB images
• Architecture:
  • Conv2D(512) → MaxPool → Conv2D(256) → MaxPool → Dropout(0.25)
  • Conv2D(128) → MaxPool → Dropout(0.25) → Conv2D(64) → MaxPool
  • Flatten → Dense(256, L2 reg) → Dropout(0.35) → Dense(256, L2 reg) → Dropout(0.35) → Dense(4, Softmax)
• Training: Adamax optimizer, 0.001 learning rate, 5 epochs

Results

Both models achieve high accuracy on the test set with comprehensive evaluation metrics including precision, recall, and confusion matrices.

Web Application Features

• Upload MRI images for classification
• Real-time prediction with confidence scores
• Saliency maps highlighting model focus areas
• AI-generated explanations using Google Gemini
• Interactive visualizations with Plotly

License

This project uses the CC0-1.0 license.