README

Overview

This repository accompanies the paper entitled Multi-disease segmentation of glioblastoma and white matter hyperintensities in BraTS 2018, submitted to Frontiers in Computational Neuroscience.

It is primarily intended for code peer review rather than general consumption.

U-Net model for brain tumor segmentation

The module unet.py implements a U-Net style encoder-decoder model, in the mxnet deep learning framework, for the voxelwise segmentation of BraTS 2018 data.

The main training notebook is training_all.ipynb, which will train on the entire training set, with hyperparameters reasonably optimized via internal cross validation on the training set. This was run multiple times to obtain a bagged ensemble set of models.

The main prediction notebook is predict_val.ipynb, which will predict BraTS validation cases based on an ensemble of models trained above.

Chronic small vessel ischemic disease (SVID)

Modifications were made to investigate the joint prediction of non-tumor white matter injury related to chronic small vessel ischemic disease (SVID). 5 channel segmentation included background (label = 0), necrotic/nonenhancing tumor (label = 1), edema (label = 2), SVID (label = 3), and enhancing tumor (label = 4).

The *_svid5.ipynb notebooks contain the modifications needed to retrain the model to additionally segment SVID. (The training notebook will obviously not run unless the additional manual segmentations are available.)

Brats18_TCIA10_261_1_flair.nii.gz	5 channel segmentation (with SVID)

4D inputs

For each subject, input FLAIR, T1, T1CE, T2, and (where applicable) segmentation .nii.gz files were combined into a single 4D NIfTI file using the convert_to_4D.R script. These notebooks presume such files are available.

Data normalization

Subject-wise and input-channel-wise data normalization was performed using whole-brain mean and standard deviation as calculated in data_normalization.ipynb.

Notebooks

At the very least, paths will need to be updated for your local environment.

Filename	Mode	Dataset	Comments
training_all.ipynb	Training	Training	Main parent notebook
training_all_svid5.ipynb	Training	Training	Modified for 5 class prediction including SVID
training_cv.ipynb	Training	Training CV
training_cv_svid5.ipynb	Training	Training CV	Modified for 5 class prediction including SVID
predict_val.ipynb	Inference	Validation
predict_val_svid5.ipynb	Inference	Validation	Modified for 5 class prediction including SVID
predict_training_cv.ipynb	Inference	Training CV
predict_training_cv_svid5.ipynb	Inference	Training CV	Modified for 5 class prediction including SVID