Brain Tumor Classification using CNN & Transfer Learning

Overview

This project implements a brain tumor detection and classification system using Convolutional Neural Networks (CNN) and transfer learning techniques (VGG16, InceptionV3) on MRI scan images. The model classifies images into four categories: No Tumor, Meningioma, Glioma, and Pituitary Tumor.

Features

• Custom CNN and fine-tuned pretrained models (VGG16, InceptionV3)
• Data augmentation, BatchNormalization, and Dropout for robust training
• Optimized with Adam optimizer
• Evaluation using ROC-AUC, precision, recall, and F1-score
• Achieved up to 90% accuracy with InceptionV3

Dataset

• Source: Kaggle - Brain Tumor MRI Dataset
• Four classes: No Tumor, Meningioma, Glioma, Pituitary Tumor

Quick Start

1. Clone the repository

   git clone https://github.com/dhruvkumar24-ai/Brain-Tumor-Classification.git

2. Install dependencies

   • Python 3.10+
   • Required packages: tensorflow, keras, scikit-learn, matplotlib, seaborn, Pillow, opencv-python, kagglehub

3. Download the dataset

   • The notebook automatically downloads the dataset using kagglehub.

4. Run the notebook

   • Open notebooks/Brain_tumor_detection_using_CNN.ipynb in VS Code or Jupyter.
   • Follow the instructions in the notebook cells.

Training Speed Optimizations

• Adjustable image size and batch size for faster training
• Mixed precision training (if supported)
• Optimized data pipeline using tf.data
• Quick testing mode for rapid experimentation

Results

• Custom CNN, VGG16, and InceptionV3 models compared
• InceptionV3 achieved highest accuracy (~90%)
• Detailed evaluation metrics provided in the notebook

Project Structure

BRAIN-TUMOUR-CLASSIFICATION/
│
├── Brain_tumor_detection_using_CNN.ipynb
├── README.md
└── (Dataset downloaded automatically)

Credits

• Developed by Dhruv Kumar Sahu
• Dataset by Masoud Nickparvar (Kaggle)

License

This project is for educational purposes.

---

📊 Results & Evaluation

The models were evaluation based on accuracy, precision, and recall. InceptionV3 emerged as the top performer due to its ability to capture multi-scale features in medical imagery.

Model	Accuracy	Precision	Recall
InceptionV3	90.2%	0.89	0.91
VGG16	84.5%	0.83	0.85
Custom CNN	76.0%	0.74	0.72

Key Insight: The use of Inception modules allowed the network to choose the optimal kernel size for feature extraction, which was critical for identifying the subtle textures of Glioma vs. Meningioma tumors. ![ROC-AUC Curve](images/"C:\Users\dhruv\OneDrive - IIT Kanpur\Pictures\Screenshots\Screenshot 2026-04-20 221014.png")