Neuroimaging-Example/README.md

BraTS 2018 - Semantic Segmentation on Brain MRI Scans

![](eef7d5d5c8/readme_images/ex1.gif)

Intro

BraTS 2018 utilizes multi-institutional pre-operative MRI scans and focuses on the segmentation of intrinsically heterogeneous (in appearance, shape, and histology) brain tumors, namely gliomas. Furthermore, to pinpoint the clinical relevance of this segmentation task, BraTS’18 also focuses on the prediction of patient overall survival, via integrative analyses of radiomic features and machine learning algorithms

Dataset

• The dataset is divided into 2 types of cancer - HGG (High-grade gliomas) and LGG (Low-grade gliomas). For every type of cancer, we will have 3D MRI scans (voxel) of patients that were acquired with different clinical protocols and various scanners from multiple (n=19) institutions. For every patient with four filters:

  • Native (T1).
  • Post-contrast T1-weighted (T1Gd).
  • T2-weighted (T2).
  • T2 Fluid Attenuated Inversion Recovery (FLAIR).

  Each filter is an MRI scan that acts as a feature. and

• The segmented MRI scenes have been segmented manually, by one to four raters, following the same annotation protocol, and their annotations were approved by experienced neuro-radiologists. The segmented data have been labeled using 3 different pixels:

  • GD-enhancing tumor (ET — label 4)
  • Peritumoral edema (ED — label 2)
  • Necrotic and non-enhancing tumor core (NCR/NET — label 1)

• Image type: NIfTI files (.nii.gz)

• Dataset tree:

      .
      ├── HGG
      │   ├── Brats18_2013_10_1
      │   │   ├── Brats18_2013_10_1_flair.nii.gz
      │   │   ├── Brats18_2013_10_1_seg.nii.gz
      │   │   ├── Brats18_2013_10_1_t1.nii.gz
      │   │   ├── Brats18_2013_10_1_t1ce.nii.gz
      │   │   └── Brats18_2013_10_1_t2.nii.gz
      │   ├── Brats18_2013_11_1
      │   │   ├── Brats18_2013_11_1_flair.nii.gz
      │   │   ├── Brats18_2013_11_1_seg.nii.gz
      │   │   ├── Brats18_2013_11_1_t1.nii.gz
      │   │   ├── Brats18_2013_11_1_t1ce.nii.gz
      │   │   └── Brats18_2013_11_1_t2.nii.gz
  
      │   ├── ...
  
      ├── LGG
      │   ├── Brats18_2013_0_1
      │   │   ├── Brats18_2013_0_1_flair.nii.gz
      │   │   ├── Brats18_2013_0_1_seg.nii.gz
      │   │   ├── Brats18_2013_0_1_t1.nii.gz
      │   │   ├── Brats18_2013_0_1_t1ce.nii.gz
      │   │   └── Brats18_2013_0_1_t2.nii.gz
      │   ├── Brats18_2013_15_1
      │   │   ├── Brats18_2013_15_1_flair.nii.gz
      │   │   ├── Brats18_2013_15_1_seg.nii.gz
      │   │   ├── Brats18_2013_15_1_t1.nii.gz
      │   │   ├── Brats18_2013_15_1_t1ce.nii.gz
      │   │   └── Brats18_2013_15_1_t2.nii.gz
      │   ├── ...
  
      ├── survival_data.csv
  

Models

• We will use a 3D U-net model from the winning paper with the following architecture:

• The model was implemented with Keras and TensorFlow V 1.* by Suyog Jadhav

Links:

• DAGsHub Repository
• Colab Notebook

TODO:

• Add global const file and change the stage const
• Split and save to csv by names of patients

Hypothesis:

• Create two models for the different types of tumors - HGG and LGG.
• Change the classification from multi-class (ncr,ed,et) to mask (one channel).
• Use 2D U-net model.
• Change the image size to the paper's size (4,160,192,128) with a batch size of 1.
• Use different model architecture.
• Change the normalization of the images.

Additional Information ℹ️

• BraTS 2018 Winning Paper
• Group Normalization
• loss functions for semantic segmentation
• The BraTS 2018 challenge