Brain Tumor Segmentation with Deep Learning and Deep Ensembles

This project focuses on brain tumor segmentation using a deep learning model implemented in TensorFlow and Keras. The goal is to identify and segment brain tumors from MRI scans, a crucial step in medical image analysis. I used Deep Ensembles to create a uncertainty map for the prediction of the U-Net.

Project Structure

• brain-tumor-segmentation.ipynb - The main Jupyter notebook containing all the steps from preprocessing to model evaluation.
• README.md - Overview and instructions for the project.
• data/ - Folder where the dataset should be placed (not included in the repository).
• models/ - Saved models and checkpoints (optional).

Features

• Preprocessing of MRI images (e.g. resizing, normalization)
• U-Net-based architecture for semantic segmentation
• Dice coefficient and binary cross-entropy loss for training
• Visualizations of predictions vs. ground truth

Requirements

• Python 3.7+
• Jupyter Notebook
• TensorFlow >= 2.x
• NumPy
• Matplotlib
• OpenCV
• scikit-learn

Dataset

The notebook expects brain MRI images with corresponding segmentation masks. You can use publicly available datasets like:

• BraTS Challenge Dataset
• Kaggle datasets for brain MRI images

Ensure your dataset is structured like this:

data/
├── images/
│   ├── patient1.png
│   └── ...
└── masks/
    ├── patient1_mask.png
    └── ...

Usage

1. Clone the repo and navigate to the folder:

git clone https://github.com/yourusername/brain-tumor-segmentation.git
cd brain-tumor-segmentation

2. Launch the notebook:

jupyter notebook brain-tumor-segmentation.ipynb

3. Follow along in the notebook to:

• Load the dataset
• Train the model
• Visualize predictions

Results

Sample output:

• Dice Score: ~0.85 (varies based on dataset and training)
• Segmentation outputs displayed as side-by-side comparisons

Future Improvements

• Add data augmentation
• Hyperparameter tuning