shayandaneshvar/brats-2020

Deep Learning with CNNs course project

Brain Tumor Segmentation using 3D UNet and Variants

Guidelines

Code

All code will be written in the src folder.

Commits

Always create a new branch to add new work, and then merge it using merge/pull request option to merge with master. This is to encounter less merge conflicts.

Useful Links

• Markdown Cheatsheet
• ImageDataGenerator tutorial
• ImageDataGenerator tutorial - TF website
• ImageDataGenerator tutorial flow function - SO

Related papers' repositories

Important Papers

• 3D U-Net Based Brain Tumor Segmentation and Survival Days Prediction, 2019, Wang et al. -> crap
• Brain Tumor Segmentation Using an Ensemble of 3D U-Nets and Overall Survival Prediction Using Radiomic Features
  • They ensembled 6 3DUNets with different number of layers, they split the data 6:4, hence I do the same splits

Related Papers

Residual UNet paper https://arxiv.org/pdf/1909.12901v2.pdf https://www.frontiersin.org/articles/10.3389/fncom.2020.00025/full Attention U-Net: Learning Where to Look for the Pancreas (2D Unet with sophisticated attention) Brain Tumor Segmentation and Survival Prediction using 3D Attention UNet (Trivial Channel Attention on 3D UNet)

Datasets

• BraTS 2020 (Test + Validation sets)
  • Multi-modal scans available as NIfTI images .nii
  • Four channels of information - four different volumes of the same image
    • T1/Native
    • T1CE/ post-contrast T1-weighted (same as first one but contrasted)
    • T2 Weighted
    • T2 Fluid attenuated inversion recovery volumes/ FLAIR
  • Labels/Annotations
    • 0: unlabeled volume: the background and parts of the brain which is normal
    • 1: Necrotic and Non-enhancing tumor core (NCR/NET)
    • 2: Peritumoral Edema (ED)
    • 4: GD-enhancing tumor (ET)

Report

Dataset stuff:

• Download dataset and unzip + install nibabel (Shayan)
• (FIX) Rename W39_1998.09.19_Segm -> BraTS20_Training_355_seg (Shayan)
• MinMax Scaler + Combine all volumes except for T1 native as T1 Native is the same as T1CE with worse contrast (Shayan)
• label 4 -> 3 (Shayan)
• Crop images and remove most of the black section (Shayan)
• (Extra) Drop volumes where there's not much annotation?? (Did not do this as there's not many images, to just lose one!)

Code stuff:

• wrote data loader and dataset
• wrote and debug 3DUNet
• TODO: add dice or focal loss or both, and train (shayan)
• Going with BCEwithLogit to see how it works

Metrics

• Dice Coefficient
• Accuracy
• ...

Models

For segmentation, variations of 3D Unet is being used, namely 3DUNet (Concatenative skips), Residual 3DUNet (Additive skips), Attention 3DUNet