🧠 Brain Tumor Detection using Deep Learning

A deep learning project for binary classification of brain MRI images to detect the presence of tumors. Three different architectures (Custom CNN, VGG-inspired, and Lightweight ResNet) are implemented and compared.

📁 Dataset Structure

./Tumor/
├── Brain Tumor/     # MRI images with tumors (class 1)
└── Healthy/         # Healthy MRI images (class 0)

Data Processing:

• Original image size: 512×512 pixels
• Preprocessing: Converted to grayscale and resized to 128×128
• Dataset loaded using os.walk() and processed with OpenCV
• Labels: 1 for tumor, 0 for healthy (tumor images loaded first)

🏗️ Model Architectures

1️⃣ Custom CNN (Baseline Model)

A lightweight convolutional network designed for small datasets:

Input (128,128,1)
↓
Conv2D(8) → BatchNorm → MaxPooling
↓
Conv2D(16) → BatchNorm → MaxPooling
↓
Conv2D(24) → BatchNorm → MaxPooling
↓
Flatten → Dense(64) → Dropout
↓
Output (1, Sigmoid)

Configuration:

• Loss: Binary Crossentropy
• Optimizer: Adam
• Metric: Accuracy
• Epochs: 15
• Batch Size: 32

Performance: ✅ 98.97% Accuracy

---

2️⃣ VGG-inspired Architecture

Deeper architecture with multiple consecutive convolutions per block:

Input (128,128,1)
↓
[Conv2D(32) → Conv2D(32)] → BatchNorm → MaxPooling
↓
[Conv2D(64) → Conv2D(64)] → BatchNorm → MaxPooling
↓
[Conv2D(128) → Conv2D(128)] → BatchNorm → MaxPooling
↓
Flatten → Dense(64) → Dropout
↓
Output (1, Sigmoid)

Features:

• 3×3 kernels with 'same' padding
• Batch Normalization after each convolution
• Progressive filter increase
• 15 epochs, batch size 32

Performance: ✅ 97.37% Accuracy

---

3️⃣ Lightweight ResNet

Residual architecture with skip connections to prevent gradient vanishing:

Input (128,128,1)
↓
Conv2D(64) → BatchNorm → MaxPooling
    ↓
    ┌── Residual Block (64 filters) ──┐
    ↓                                 ↓
Conv2D(32) → BatchNorm → MaxPooling ← Skip Connection
    ↓
    ┌── Residual Block (32 filters) ──┐
    ↓                                 ↓
Conv2D(16) → BatchNorm → MaxPooling ← Skip Connection
    ↓
    ┌── Residual Block (16 filters) ──┐
    ↓                                 ↓
Global Average Pooling → Dense(128) → Dropout
↓
Output (1, Sigmoid)

Features:

• Skip connections for better gradient flow
• Global Average Pooling instead of Flatten
• Progressive filter reduction (64 → 32 → 16)
• 15 epochs, batch size 32

Performance: 🏆 99.29% Accuracy (Best Model)

📊 Results Summary

Architecture	Accuracy	Key Feature
Custom CNN	98.97%	Simple, efficient, low overfitting
VGG-inspired	97.37%	Deeper, multiple conv layers
Lightweight ResNet	99.29%	Skip connections, best performance

🚀 Quick Start

# Clone repository
git clone https://github.com/yourusername/brain-tumor-detection.git
cd brain-tumor-detection

# Install dependencies
pip install -r requirements.txt

# Run training
python train.py --model resnet  # or cnn/vgg

📦 Requirements

• TensorFlow 2.x
• OpenCV (cv2)
• NumPy
• scikit-learn
• Matplotlib

📝 License

This project is licensed under the MIT License.

---

⭐ If you find this project useful, please consider giving it a star!