Brain Tumor Segmentation Project

Dataset

• Source: Brain Tumor MRI Dataset
• Classes:
  • glioma: 0
  • meningioma: 1
  • notumor: 2
  • pituitary: 3

Models and Approaches

1. Normal CNN Model

• Architecture: Custom Convolutional Neural Network
  • Convolutional layers with ReLU activation
  • Max pooling layers
  • Dropout for regularization
  • Fully connected layers
• Key Features:
  • 3 convolutional blocks
  • 2 fully connected layers
  • Softmax output layer

2. VGG16 Model

• Pre-trained VGG16 architecture
• Modified final classification layer
• 4-class classification

3. Faster R-CNN

• Pretrained Faster R-CNN with ResNet50 backbone
• Object detection and localization
• Used for precise tumor region identification

Data Augmentation Techniques

• Random horizontal flips
• Random rotations (±10 degrees)
• Random resized crop
• Color jittering

Training Details

Loss Function

• Cross-Entropy Loss for all models
• Optimized using Adam optimizer

Performance Metrics

• Accuracy
• Precision
• Recall
• F1 Score
• Confusion Matrix Analysis

Results

Normal CNN

• Accuracy: ~70%
• Precision/Recall: Varies by class
• Best performance on 'notumor' class

VGG16 (4 Classes)

• Accuracy: ~60%
• Precision range: 0.45 - 0.72
• Strong performance on 'notumor' class

VGG16 (2 Classes)

• Accuracy: ~94%
• High precision and recall
• Binary classification (tumor vs. no tumor)

Faster R-CNN Segmentation Technique

• Localization and classification of tumor regions
• In the 4-class classification, implemented Chan-Vese segmentation algorithm
• Enhances tumor region localization and segmentation
• Intersection over Union (IoU) based accuracy
• Provides precise region of interest extraction after initial tumor localization