Brain Tumor Segmentation using 3D UNet

This project involves training a 3D UNet model for the task of brain tumor segmentation using volumetric MRI data. The goal is to accurately segment tumor regions from 3D MRI scans.

Table of Contents

• Introduction
• Dataset
• Model Architecture
• Training
• Evaluation
• Installation
• Usage
• Results
• Acknowledgements

Introduction

Brain tumor segmentation is a critical task in medical imaging that aids in the diagnosis and treatment planning for patients. This project implements a 3D UNet model, which is commonly used for medical image segmentation tasks due to its ability to capture both local and global contextual information.

Dataset

The dataset used in this project is collected from Synapse, a collaborative platform for sharing data, tools, and ideas. The dataset consists of volumetric MRI scans stored in .h5 files. Each file contains a 3D image of the brain and the corresponding segmentation mask, where different labels represent different tumor regions.

Data Preprocessing

The images and masks are normalized, resized, and augmented during training to improve the model's robustness. The preprocessing steps include:

• Normalization: Scaling the pixel values to a standard range.
• Resizing: Adjusting the images and masks to a uniform size.
• Augmentation: Applying random transformations like flips and rotations to the data.

Model Architecture

The model used is a 3D UNet, which consists of the following components:

• Encoder: A series of convolutional layers that downsample the input and extract features.
• Bottleneck: The central part of the network that captures the most abstract features.
• Decoder: A series of upsampling layers that reconstruct the segmented image from the bottleneck features.
• Skip Connections: Connections between the encoder and decoder layers to retain spatial information.

Training

The model is trained using a combined loss function that includes Dice Loss and Binary Cross-Entropy (BCE) Loss. The training process includes:

• Optimizer: Adam optimizer with an initial learning rate of 0.001.
• Learning Rate Scheduler: Reduces the learning rate when the validation loss plateaus.
• Early Stopping: Stops training when the validation loss does not improve for a specified number of epochs.

Commands

To train the model, run:

python main.py