Brain Tumor Segmentation (3D) Using FPN-3D U-Net

This repository contains the implementation of a deep learning-based brain tumor segmentation model using a hybrid FPN-3D U-Net architecture. This project addresses the critical need for automated, accurate, and efficient segmentation of brain tumors in MRI scans, aiding medical professionals in diagnosis and treatment planning.

---

Table of Contents

1. Project Overview
2. Dataset
3. Preprocessing Pipeline
4. Model Architecture
5. Training and Evaluation
6. Results and Analysis
7. How to Use
8. References

---

Project Overview

Objective

This project aims to develop a 3D segmentation model that can accurately segment brain tumors into four classes:

• Necrotic core
• Peritumoral edema
• Enhancing tumor
• Background

Why Automated Brain Tumor Segmentation?

Manual segmentation of brain tumors is time-consuming, prone to variability, and dependent on expert radiologists. This project leverages deep learning to:

• Automate segmentation tasks.
• Provide consistent and reliable results.
• Improve diagnostic workflows, especially in resource-limited settings.

Key Features of the Project

• Utilizes BraTS21 Dataset, which includes four MRI modalities: FLAIR, T1, T1Gd, and T2.
• Integrates a Feature Pyramid Network (FPN) with a 3D U-Net for enhanced multi-scale feature extraction.
• Employs a custom Dice-Cross Entropy Loss function to handle class imbalance and improve segmentation accuracy.
• Comprehensive preprocessing pipeline to standardize and augment data.

---

Dataset

The BraTS21 Dataset is a benchmark dataset for brain tumor segmentation. It contains 3D MRI scans with expert-annotated segmentation labels.

MRI Modalities

1. FLAIR: Highlights abnormalities such as edema.
2. T1-weighted: Provides anatomical details of brain structures.
3. Post-contrast T1-weighted (T1Gd): Enhances active tumor regions.
4. T2-weighted: Emphasizes fluid-rich areas and complements FLAIR.

Dataset Characteristics

• 1,251 MRI scans.
• Four modalities per scan.
• Each scan is resampled to a uniform voxel size of 1 mm isotropic.
• Segmentation annotations include four classes.

Visualization of MRI Modalities:

---

Preprocessing Pipeline

A robust preprocessing pipeline was developed to standardize and prepare the dataset for training:

1. Skull-Stripping:

   • Removes non-brain tissues, focusing solely on the brain.

2. Resampling:

   • Resamples scans to a uniform voxel resolution of 1 mm isotropic.

3. Cropping to Foreground:

   • Reduces unnecessary background regions by cropping around the brain.

4. Intensity Normalization:

   • Standardizes voxel intensity to have zero mean and unit variance.

5. Augmentation:

   • Includes random flips, rotations, and gamma adjustments to enhance diversity.

6. Resizing:

   • Rescales scans to a fixed size of 128 x 128 x 128 voxels.

Example of a Preprocessed Image:

---

Model Architecture

The proposed model combines the 3D U-Net with a Feature Pyramid Network (FPN) to leverage multi-scale feature extraction.

Key Components

1. Encoder:

   • Extracts hierarchical features using convolutional blocks and downsampling.

2. FPN Layers:

   • Integrates multi-scale features for better context and detail.

3. Decoder:

   • Reconstructs high-resolution segmentation maps using skip connections and upsampling layers.

4. Custom Loss Function:

   • Combines Dice Loss and Cross-Entropy Loss to address imbalanced data.

FPN-3D U-Net Architecture

---

Training and Evaluation

Training Details

• Batch Size: 1
• Epochs: 10
• Learning Rate: 0.0003
• Optimizer: Adam

Loss Function

The combined loss function:

• Dice Loss: Measures overlap between predicted and true segmentation.
• Cross-Entropy Loss: Penalizes incorrect class predictions.

Evaluation Metrics

• Dice Similarity Coefficient (DSC): Evaluates segmentation accuracy.

---

How to Use

Clone the Repository

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

Install Dependencies

pip install -r requirements.txt

Preprocess the Dataset

python preprocess.py

Train the Model

python main.py --mode train

Evaluate the Model

python main.py --mode test --weights final_model.pth

---

References

1. Havaei, M., et al., "Brain Tumor Segmentation with Deep Neural Networks," Medical Image Analysis, 2017.
2. Bakas, S., et al., "Segmentation Labels and Radiomic Features for TCGA-GBM Collection," The Cancer Imaging Archive, 2017.
3. Çiçek, Ö., et al., "3D U-Net: Learning Dense Volumetric Segmentation from Sparse Annotation," MICCAI 2016.
4. Brain tumor image segmentation based on improved FPN,Haitao Sun, Shuai Yang, Lijuan Chen, Pingyan Liao, Xiangping Liu, Ying Liu & Ning Wang BMC Medical Imaging

---