search

Thunderhead-exe / Multiclass-Brain-Tumor-Segmentation-3D-MRI

Multiclass brain tumor segmentation for T2-weighted 3D MRI using 3D-UNet architecture. Plus, multiple visualizations techniques for 3D MRI Data.
MIT License
4 stars 0 forks source link

readme

Multiclass Brain Tumor Segmentation from 3D MRI

Brain MRI Segmentation

Table of Contents

Project Overview

This project focuses on multiclass brain tumor segmentation using 3D MRI T2-weighted images. We employ a 3D-UNet architecture to segment four distinct classes within brain MRI 3D scans, utilizing the BraTS2020 Dataset from Kaggle. \ \ Date: July 2023

Data

The dataset consists of 3D MRI T2-weighted scans from the BraTS2020 Dataset, available on Kaggle. These images are used for multiclass (4 classes) brain tumor segmentation.

Sample MRI Slices

Figure 1: Sample MRI slices showing different tumor classes

Model Architecture

We utilize a 3D-UNet model architecture for multiclass segmentation. This architecture is particularly well-suited for 3D medical imaging data as it can effectively capture and utilize spatial information in all three dimensions.

Illustration of the 3D U-Net architecture

Figure 2: 3D-UNet model architecture

Why 3D-UNet?

  1. Spatial Information: 3D-UNet is designed to process volumetric data, making it ideal for 3D MRI scans. It preserves and utilizes spatial information in all three dimensions, which is crucial for accurate tumor segmentation.

  2. Skip Connections: The architecture includes skip connections that combine low-level and high-level features, allowing for more precise localization of tumors.

  3. Depth: The model's depth allows it to learn hierarchical features, from simple edges to complex tumor patterns.

  4. Efficiency: Despite working with 3D data, the model is relatively efficient due to its U-shaped structure and the use of max-pooling operations.

Parameters

Performance

Our model achieved the following performance metrics:

MRI Data Augmentation

To improve model generalization and performance, we applied the following data augmentation techniques:

  1. Horizontal Flip
  2. Elastic Deformation
  3. Contrast and Brightness Adjustment

MRI Data Augmentation 1 MRI Data Augmentation 2

Figure 5: MRI Data Augmentation

Visualizations

To better understand the 3D nature of our data and results, we provide various visualization techniques:

1. 3D Visualization

3D Tumor Visualization 3D Tumor Visualization 3D Tumor Visualization

Figure 6: 3D visualizations of segmented tumor

2. GIF Animation

MRI Slice Animation MRI Slice Animation

Figure 7: Animated GIFs showing MRI slices

3. 2D Montage

2D Montage of MRI Slices

Figure 8: 2D montage of key MRI slices

Getting Started

To use this project, follow these steps:

  1. Clone the repository:

    git clone https://github.com/Thunderhead-exe/Multiclass-Brain-Tumor-Segmentation-3D-MRI.git

  2. Follow the provided notebook

Note: The trained model is provided too, feel free to experiment!

Contributing

We welcome contributions to improve this project. Please feel free to submit issues, feature requests, or pull requests.

License

This project is licensed under the MIT License - see the LICENSE file for details.