Brain Tumor Detection

Multi-Class Brain Tumor Detection Using Deep Learning

by

Brain tumors are abnormal growths of cells in the brain that can be life-threatening. Early and accurate detection is crucial for effective treatment. Deep learning, specifically convolutional neural networks (CNNs), has revolutionized medical imaging by providing automated and accurate diagnoses. This project focuses on detecting different types of brain tumors using a deep learning model trained on MRI images.

Dataset

The dataset used in this project consists of MRI scans categorized into three tumor types and one non-tumor class:

  • Glioma Tumor: A type of tumor that arises from glial cells in the brain.
  • Meningioma Tumor: A tumor that forms in the meninges, the protective layers of the brain.
  • Pituitary Tumor: A growth in the pituitary gland, which affects hormone production.
  • No Tumor: MRI images without any tumor.

The dataset is divided into:

  • Training Set: 80% of the total images used for training the model.
  • Validation Set: 10% of images used for tuning hyperparameters and preventing overfitting.
  • Testing Set: 10% of images reserved for evaluating the model’s final performance.

Data Preprocessing

  1. Data Augmentation: Applied to increase dataset variability and improve generalization. Techniques used:
    • Rotation: Randomly rotating images up to 20 degrees.
    • Flipping: Horizontally flipping images.
    • Brightness Adjustment: Enhancing or reducing brightness within a small range.
    • Zooming: Random zoom applied to simulate different perspectives.
  2. Normalization: Pixel values are scaled to a range of [0,1] by dividing by 255.
  3. Shuffling: Ensures random order of images to prevent learning sequence bias.
  4. Resizing: All images are resized to 224×224 pixels to match the input requirement of VGG16.
  5. One-Hot Encoding: Since this is a multi-class problem, labels are converted into one-hot encoded vectors.

Model Architecture

The deep learning model is built using VGG16, a pre-trained CNN model, with modifications for multi-class classification. The architecture includes:

  • Input Layer: Takes in 224×224 MRI images with three color channels.
  • Feature Extractor (VGG16):
    • Uses convolutional layers pre-trained on ImageNet to extract features.
    • The convolutional layers remain frozen initially to leverage pre-learned features.
  • Flatten Layer: Converts the extracted feature maps into a one-dimensional vector.
  • Fully Connected Layers:
    • Dense Layer (512 neurons, ReLU activation)
    • Dropout Layer (rate = 0.5) to prevent overfitting
    • Dense Layer (256 neurons, ReLU activation)
    • Dropout Layer (rate = 0.5)
  • Output Layer: Softmax activation with 4 units for classifying into Glioma, Meningioma, Pituitary, or No Tumor.

Training Process

  • Optimizer: Adam optimizer with a learning rate of 0.0001.
  • Loss Function: Categorical cross-entropy for multi-class classification.
  • Batch Size: 32 images per batch.
  • Epochs: 25 epochs used for training with early stopping to avoid overfitting.
  • Callbacks:
    • Early Stopping: Stops training if validation loss does not improve for 5 consecutive epochs.
    • Model Checkpoint: Saves the best model weights based on validation accuracy.

Results and Analysis

  • The model achieved an overall accuracy of ~95% on the test dataset.
  • Class-wise performance:
    • Glioma Tumor: Accuracy of 96%, with a few misclassifications as Meningioma.
    • Meningioma Tumor: Accuracy of 94%, some false positives classified as No Tumor.
    • Pituitary Tumor: Accuracy of 97%, with high precision and recall.
    • No Tumor: Accuracy of 93%, minor false positives in tumor categories.
  • Confusion Matrix Insights:
    • Some glioma cases were misclassified as meningioma due to overlapping features.
    • Few non-tumor cases were incorrectly classified as tumors (false positives).
    • Pituitary tumors had the highest classification accuracy due to distinct characteristics.
  • Precision, Recall, and F1 Score:
    • Glioma Tumor: Precision 95%, Recall 94%, F1 Score 94.5%.
    • Meningioma Tumor: Precision 92%, Recall 93%, F1 Score 92.5%.
    • Pituitary Tumor: Precision 98%, Recall 97%, F1 Score 97.5%.
    • No Tumor: Precision 91%, Recall 92%, F1 Score 91.5%.

Conclusion

This project successfully demonstrates the potential of deep learning for medical diagnosis, particularly in multi-class brain tumor detection. The VGG16-based model effectively classifies MRI images into four categories with high accuracy.

Future Enhancements:

  • Experiment with deeper architectures like ResNet50 or EfficientNet for improved accuracy.
  • Fine-tune VGG16 layers instead of freezing them to capture domain-specific features.
  • Use more diverse MRI datasets to enhance model generalization.
  • Implement Explainable AI (XAI) techniques to provide visual explanations for model predictions.

References:

  • Medical imaging datasets.
  • Deep learning frameworks like TensorFlow and Keras.
  • Pre-trained models such as VGG16.
  • Research papers on CNN-based brain tumor classification.

Leave a Comment