A complete pipeline for BraTS 2020: Multimodal Brain Tumor Segmentation Challenge 2020 based on 3D U-net

The github repo lets you train a 3D U-net model using BraTS 2020 dataset (perhaps it can be used for previous BraTS dataset). While this repo is a ready-to-use pipeline for segmentation task, one may extend this repo for other tasks such as survival task and Uncertainty task. Even the repo may be used for other 3D dataset/task.

If you face any problem, please feel free to open an issue.

Directory structure

.
├─ data
  ├─ brats20	# Data provided by the BraTS 2020 competition host
    ├─ TrainingData
	├─ BraTS20_Training_001
	    ├─ BraTS20_Training_001_flair.nii.gz
	    ├─ BraTS20_Training_001_seg.nii.gz
	    ├─ BraTS20_Training_001_t1.nii.gz
	    ├─ BraTS20_Training_001_t1ce.nii.gz
	    ├─ BraTS20_Training_001_t2.nii.gz
	├─ BraTS20_Training_002
	├─ ...
    ├─ ValidationData
	├─ BraTS20_Validation_001
	    ├─ BraTS20_Validation_001_flair.nii.gz
	    ├─ BraTS20_Validation_001_t1.nii.gz
	    ├─ BraTS20_Validation_001_t1ce.nii.gz
	    ├─ BraTS20_Validation_001_t2.nii.gz
	    ├─ ...
├─ model	# Generated training and validation split (training_ids.pkl, validation_ids_pkl, test_ids.pkl), processed data file (brats20_data.h5, brats20_data_test.h5), and save best training model (isensee_2017_model.h5)
├─ output	# Generated prediction file
├─ src		# Souce code
    ├─ unet3d
    ├─ config.py
    ├─ inference.py
    ├─ train.py
├─ inference.sh
├─ train.sh

Packages

• Python >= 3.5 (my current version is 3.7.7)
• tensorflowgpu==1.15 (other 1.x version should work)
• Other packages: pytables, SimpleITK, nilearn, nibabel
• Optional package: nipype (For n4itk bias correction preprocessing only. However, I didn't achieve that much performance gain using this technique!)

• Note for installing 'pytables': install it using conda ('conda install pytables'). Installing using pip ('pip install tables') raises 'memory dump' issue!

How to run

If you prepare directory structure properly, you are done!

Train:

$ ./train.sh

Validation/Test:

$ ./inference.sh

Results on validation set

Label Dice_ET Dice_WT Dice_TC Sensitivity_ET Sensitivity_WT Sensitivity_TC Specificity_ET Specificity_WT Specificity_TC Hausdorff95_ET Hausdorff95_WT Hausdorff95_TC

Mean 0.60954 0.82701 0.74769 0.66736 0.86277 0.75625 0.99939 0.99834 0.99931 46.96773 10.70035 11.63979

StdDev 0.31934 0.15026 0.21439 0.33152 0.14198 0.22683 0.00098 0.00127 0.00108 112.51574 17.23635 18.01854

Median 0.7783 0.87372 0.84036 0.81398 0.90829 0.85127 0.99967 0.99875 0.99967 3.60555 5.47723 5.74456

25quantile 0.44618 0.814 0.63882 0.5355 0.83669 0.64064 0.99921 0.998 0.99928 2 3.74166 3

75quantile 0.84443 0.90629 0.89807 0.89286 0.94265 0.91698 0.99987 0.99916 0.99987 15.67788 10.44031 12.36932

Acknowledgment

Significant code has been borrowed from ellisdg's repository which is based on Isensee et al.'s paper. I really appreciate David G Ellis's contributions to the community.