BraTs

[TOC]

1. Overview

Fig 1: Brain Complete Tumor Segmention

Fig 2: Brain Core Tumor Segmention

Ground Truth
Prediction

BRaTS stands for Brain Tumor Segmentation. The BRaTS challenge has always been focusing on the evaluation of the state-of-the-art methods for the segmentation of brain tumors in multi-modal magnetic resonance imaging (MRI) scans. This is a coordinated effort for Tumor Segmentation from the University of Pennsylvania, Perelman School of Medicine.

MRI Scans of Glioblastomas/High Grade Glioma (GBM/HGG) and low grade glioma (LGG) with pathologically confirmed diagnosis are labelled and are available for download.

2. Models

• U-Net

pytorch/models/unet.py

Fig 3: U-Net Diagram

• DeepLab V3 +

tensorflow/models/research/deeplab

Fig 4: DeepLab V3+ Diagram

3. Dataset

3.1 Overview

File: FLAIR MRI Sequence Data of One Person

File Type: png files

Image-Shape: 240(Slide Width) × 240(Slide Height) × 31(Number of Slide) × 1(Multi-mode)

Image Subjects: 31 persons

3.2 Labels

• GAD-Enhancing Tumor - WHITE
• Tumor Core - BLACK
• Whole Tumor - GREY
• Background - GREYISH BLACK

3.3 Data Preprocessing

• Co-registering
• Interpolation to the same resolution (1 mm^3)
• Skull Stripped

4. Train

3.1 Loss Function

Dice Coefficient Loss

4.2 Optimizer

Adam Optimizer

Stochastic Gradient Descent

4.3 Hyperparameters

learning rate = 1e-4

maximum number of epochs = 1000

Weights Init: Normal Distribution (mean:0, std:0.01)

Bias Init: Initialized as 0

5. Steps to Run

Step 1: Download complete model and unzip from here.

Step 2:

git clone https://github.com/geekswaroop/BRaTS.git

cd into the downloaded folder and then run

Step 3:

python train.py

6. Results

7. References

1. The Multimodal Brain Tumor Image Segmentation Benchmark (BRATS)

2. Investigator's Implementation

3. U-Net: Convolutional Networks for Biomedical Image Segmentation