Overview

This repository provides source code and pre-trained models for brain tumor segmentation with BraTS dataset. The method is detailed in [1], and it won the 2nd place of MICCAI 2017 BraTS Challenge. In addition, it is adapted to deal with BraTS 2015 dataset.

This implementation is based on NiftyNet and Tensorflow. While NiftyNet provides more automatic pipelines for dataloading, training, testing and evaluation, this naive implementation only makes use of NiftyNet for network definition, so that it is lightweight and extensible. A demo that makes more use of NiftyNet for brain tumor segmentation is proivde at https://github.com/NifTK/NiftyNet/tree/dev/demos/BRATS17

If you use any resources in this repository, please cite the following papers:

[1] Guotai Wang, Wenqi Li, Sebastien Ourselin, Tom Vercauteren. "Automatic Brain Tumor Segmentation using Cascaded Anisotropic Convolutional Neural Networks." In Brainlesion: Glioma, Multiple Sclerosis, Stroke and Traumatic Brain Injuries. Pages 179-190. Springer, 2018. https://arxiv.org/abs/1709.00382
[2] Eli Gibson, Wenqi Li, Carole Sudre, Lucas Fidon, Dzhoshkun I. Shakir, Guotai Wang, Zach Eaton-Rosen, Robert Gray, Tom Doel, Yipeng Hu, Tom Whyntie, Parashkev Nachev, Marc Modat, Dean C. Barratt, Sébastien Ourselin, M. Jorge Cardoso^, Tom Vercauteren^. "NiftyNet: a deep-learning platform for medical imaging." Computer Methods and Programs in Biomedicine, 158 (2018): 113-122. https://arxiv.org/pdf/1709.03485

An example of brain tumor segmentation result.

Requirements

A CUDA compatable GPU with memoery not less than 6GB is recommended for training. For testing only, a CUDA compatable GPU may not be required.
Tensorflow (v1.4.0). Install tensorflow following instructions from https://www.tensorflow.org/install/
NiftyNet (v0.2.0). Install it by following instructions from http://niftynet.readthedocs.io/en/dev/installation.html
BraTS 2015 or 2017 dataset. Data can be downloaded from http://braintumorsegmentation.org/

How to use

1, Prepare data

Download BraTS dataset, and uncompress the training and tesing zip files. For example, the training set will be in data_root/BRATS2015_Training or data_root/Brats17TrainingData and the validation set will be in data_root/BRATS2015_Validation or data_root/Brats17ValidationData.

2, Use pre-trained models to segment images

To segment BraTS 2015 data, run:

python test.py config15/test_all_class.txt

To segment BraTS 2017 data, run:

python test.py config17/test_all_class.txt

3, How to train

The trainig process needs 9 steps, with axial view, sagittal view, coronal view for whole tumor, tumor core, and enhancing core, respectively.

The following commands are examples for BraTS 2017. However, you can edit the corresponding *.txt files for different configurations.

Train models for whole tumor in axial, sagittal and coronal views respectively. Run:

python train.py config17/train_wt_ax.txt
python train.py config17/train_wt_sg.txt
python train.py config17/train_wt_cr.txt

Train models for tumor core in axial, sagittal and coronal views respectively. Run:

python train.py config17/train_tc_ax.txt
python train.py config17/train_tc_sg.txt
python train.py config17/train_tc_cr.txt

Train models for enhancing core in axial, sagittal and coronal views respectively. Run:

python train.py config17/train_en_ax.txt
python train.py config17/train_en_sg.txt
python train.py config17/train_en_cr.txt

To save the time for training, you may use the modals in axial view as initalizations for sagittal and coronal views. Copy variales in axial view to those in sagittal or coronal view by running:

python util/rename_variables.py

You may need to edit this file to set different parameters. As an example for Brats 2015, after running this command, you will see a model named model15/msnet_tc32sg_init that is copied from model15/msnet_tc32_20000.ckpt. Then just set start_iteration=1 and model_pre_trained=model15/msnet_tc32sg_init in config15/train_tc_sg.txt.

4, How to test

Similar to 'Use pre-trained models', write a configure file that is similar to config15/test_all_class.txt or config17/test_all_class.txt and set the value of model_file to your own model files. Run:

python test.py your_own_config_for_test.txt

5, Evaluation

Calcuate dice scores between segmentation and the ground truth, run:

python util/evaluation.py

You may need to edit this file to specify folders for segmentation and ground truth.

taigw / brats17

readme