HD-GLIO

Introduction

This repository provides easy to use access to our HD-GLIO brain tumor segmentation tool. HD-GLIO is the result of a joint project between the Department of Neuroradiology at the Heidelberg University Hospital, Germany and the Division of Medical Image Computing at the German Cancer Research Center (DKFZ) Heidelberg, Germany. If you are using HD-GLIO, please cite the following two publications:

• Kickingereder P, Isensee F, Tursunova I, Petersen J, Neuberger U, Bonekamp D, Brugnara G, Schell M, Kessler T, Foltyn M, Harting I, Sahm F, Prager M, Nowosielski M, Wick A, Nolden M, Radbruch A, Debus J, Schlemmer HP, Heiland S, Platten M, von Deimling A, van den Bent MJ, Gorlia T, Wick W, Bendszus M, Maier-Hein KH. Automated quantitative tumour response assessment of MRI in neuro-oncology with artificial neural networks: a multicentre, retrospective study. Lancet Oncol. 2019 May;20(5):728-740. (https://doi.org/10.1016/S1470-2045(19)30098-1)
• Isensee F, Petersen J, Kohl SAA, Jaeger PF, Maier-Hein KH. nnU-Net: Breaking the Spell on Successful Medical Image Segmentation. arXiv preprint 2019 arXiv:1904.08128. (https://arxiv.org/abs/1904.08128)

HD-GLIO was developed with 3220 MRI examinations from 1450 brain tumor patients (80% for training and 20% for testing). Specifically the data included:

1. a single-institutional retrospective dataset with 694 MRI examinations from 495 patients acquired at the Department of Neuroradiology, Heidelberg University Hospital, Germany (corresponding to the “Heidelberg training dataset and Heidelberg test dataset” described in Kickingereder et al. Lancet Oncol. 2019)
2. a multicentric clinical trial dataset with 2034 MRI examinations from 532 patients acquired across 34 institutions in Europe (corresponding to the “EORTC-26101 test dataset” described in Kickingereder et al. Lancet Oncol. 2019)
3. a single-institutional retrospective dataset with 492 MRI examinations from 423 patients (80% glial brain tumors, 20% other histological entities) undergoing routine MRI at different stages of the disease (including 79 early postoperative MRI scans acquired <72h after surgery) at the Department of Neuroradiology, Heidelberg University Hospital, Germany

Specifically, each MRI examination included precontrast T1-weighted, postcontrast T1-weighted, T2-weighted and FLAIR sequences (all sequences brain extracted and co-registered) as well as corresponding ground-truth tumor segmentation masks. HD-GLIO performs brain tumor segmentation for contrast-enhancing tumor and non-enhancing T2/FLAIR signal abnormality. We applied a variant of the nnU-Net ('no-new-Net') framework (as described in Isensee et al. arXiv preprint 2019) for training the HD-GLIO algorithm.

HD-GLIO is very fast on GPU with <10s run time per MRI examination.

Installation Instructions

Installation Requirements

Unlike HD-BET, HD-GLIO requires a GPU to perform brain tumor segmentation. Any GPU with 4 GB of VRAM and cuda/pytorch support will do. Running the prediction on CPU is not supported.

Installation with Pip

Installation with pip is quick an easy, just run the following command and everything will be done for you:

pip install hd_glio

Manual installation

If you intend to modify HD-GLIO, you can also install is manually:

1. Clone this repository:

   git clone https://github.com/MIC-DKFZ/HD-GLIO

2. Go into the repository (the folder with the setup.py file) and install:

   cd HD-GLIO

   pip install -e .

Per default, model parameters will be downloaded to ~/.hd_glio_params. If you wish to use a different folder, open hd_glio/paths.py in a text editor and modify folder_with_parameter_files.

Both manual and pip installation will install two commands with which you can use HD-GLIO from anywhere in your system: hd_glio_predict and hd_glio_predict_folder.

How to use it

Using HD_GLIO is straightforward. You can use it in any terminal on your linux system. The hd_glio_predict and hd_glio_predict_folder commands were installed automatically. HD-GLIO requires a GPU with at least 4 GB of VRAM to run.

Prerequisites

HD-GLIO was trained with four MRI modalities: T1, constrast-enhanced T1, T2 and FLAIR. All these modalities must be present in order to run HD-GLIO.

All input files must be provided as nifti (.nii.gz) files containing 2D or 3D MRI image data. Sequences with multiple temporal volumes (i.e. 4D sequences) are not supported (however can be splitted upfront into the individual temporal volumes using fslsplit1).

• INPUT_T1 must be a T1-weighted sequence before contrast-agent administration (T1-w) acquired as 2D with axial orientation (e.g. TSE) or as 3D (e.g. MPRAGE)
• INPUT_CT1 must be a T1-weighted sequence after contrast-agent administration (cT1-w) acquired as 2D with axial orientation (e.g. TSE) or as 3D (e.g. MPRAGE)
• INPUT_T2 must be a T2-weighted sequence (T2-w) acquired as 2D
• INPUT_FLAIR must be a fluid attenuated inversion recovery (FLAIR) sequence acquired as 2D with axial orientation (e.g. TSE). A 3D acquisition (e.g. 3D TSE/FSE) may work as well.

(These specifications are in line with the consensus recommendations for a standardized brain tumor imaging protocol in clinical trials - see Ellingson et al. Neuro Oncol. 2015 Sep;17(9):1188-98 - www.ncbi.nlm.nih.gov/pubmed/26250565)

Input files must contain 3D images; Sequences with multiple temporal volumes (i.e. 4D sequences) are not supported (however can be splitted upfront into the individual temporal volumes using fslsplit¹).

All input files must match the orientation of standard MNI152 template and must be brain extracted and co-registered. All non-brain voxels must be 0. To ensure that these pre-processing steps are performed correctly you may adhere to the following example:

Reorient MRI sequences to standard space

fslreorient2std T1.nii.gz T1_reorient.nii.gz
fslreorient2std CT1.nii.gz CT1_reorient.nii.gz
fslreorient2std T2.nii.gz T2_reorient.nii.gz
fslreorient2std FLAIR.nii.gz FLAIR_reorient.nii.gz

The following is the recommended workflow for FSL5. There is a better way to do this but this requires FSL6 (see below)

# perform brain extraction using HD-BET (https://github.com/MIC-DKFZ/HD-BET)
hd-bet -i T1_reorient.nii.gz
hd-bet -i CT1_reorient.nii.gz
hd-bet -i T2_reorient.nii.gz
hd-bet -i FLAIR_reorient.nii.gz 

# register all sequences to T1
fsl5.0-flirt -in CT1_reorient_bet.nii.gz -ref T1_reorient_bet.nii.gz -out CT1_reorient_bet_reg.nii.gz -dof 6 -interp spline
fsl5.0-flirt -in T2_reorient_bet.nii.gz -ref T1_reorient.nii.gz -out T2_reorient_bet_reg.nii.gz -dof 6 -interp spline
fsl5.0-flirt -in FLAIR_reorient_bet.nii.gz -ref T1_reorient.nii.gz -out FLAIR_reorient_bet_reg.nii.gz -dof 6 -interp spline

# reapply T1 brain mask (this is important because HD-GLIO expects non-brain voxels to be 0 and the registration 
process can introduce nonzero values
# T1_BRAIN_MASK.nii.gz is the mask (not the brain extracted image!) as obtained from HD-Bet
fsl5.0-fslmaths CT1_reorient_bet_reg.nii.gz -mas T1_BRAIN_MASK.nii.gz CT1_reorient_bet_reg.nii.gz
fsl5.0-fslmaths T2_reorient_bet_reg.nii.gz -mas T1_BRAIN_MASK.nii.gz T2_reorient_bet_reg.nii.gz
fsl5.0-fslmaths FLAIR_reorient_bet_reg.nii.gz -mas T1_BRAIN_MASK.nii.gz FLAIR_reorient_bet_reg.nii.gz

This is how to do it with FSL6:

# run hd bet
hd-bet -i T1_reorient.nii.gz -o t1_bet.nii.gz -s 1
hd-bet -i CT1_reorient.nii.gz -o ct1_bet.nii.gz
hd-bet -i T2_reorient.nii.gz -o t2_bet.nii.gz
hd-bet -i FLAIR_reorient.nii.gz -o flair_bet.nii.gz

# register brain extracted images to t1, save matrix
flirt -in ct1_bet.nii.gz -out ct1_bet_reg.nii.gz -ref t1_bet.nii.gz -omat ct1_to_t1.mat -interp spline -dof 6 &
flirt -in t2_bet.nii.gz -out t2_bet_reg.nii.gz -ref t1_bet.nii.gz -omat t2_to_t1.mat -interp spline -dof 6 &
flirt -in flair_bet.nii.gz -out flair_bet_reg.nii.gz -ref t1_bet.nii.gz -omat flair_to_t1.mat -interp spline -dof 6 &
wait

# we are only interested in the matrices, delete the other output images
rm ct1_bet.nii.gz t2_bet.nii.gz flair_bet.nii.gz
rm ct1_bet_reg.nii.gz t2_bet_reg.nii.gz flair_bet_reg.nii.gz

# now apply the transformation matrices to the original images (pre hd-bet)
flirt -in CT1_reorient.nii.gz -out ct1_reg.nii.gz -ref t1_bet.nii.gz -applyxfm -init ct1_to_t1.mat -interp spline &
flirt -in T2_reorient.nii.gz -out t2_reg.nii.gz -ref t1_bet.nii.gz -applyxfm -init t2_to_t1.mat -interp spline &
flirt -in FLAIR_reorient.nii.gz -out flair_reg.nii.gz -ref t1_bet.nii.gz -applyxfm -init flair_to_t1.mat -interp spline &
wait

# now apply t1 brain mask to all registered images
fslmaths ct1_reg.nii.gz -mas t1_bet_mask.nii.gz CT1_reorient_reg_bet.nii.gz & # t1_bet_mask.nii.gz was generated by hd-bet (see above)
fslmaths t2_reg.nii.gz -mas t1_bet_mask.nii.gz T2_reorient_reg_bet.nii.gz & # t1_bet_mask.nii.gz was generated by hd-bet (see above)
fslmaths flair_reg.nii.gz -mas t1_bet_mask.nii.gz FLAIR_reorient_reg_bet.nii.gz & # t1_bet_mask.nii.gz was generated by hd-bet (see above)
wait

# done

After applying this example you would use T1_reorient.nii.gz, CT1_reorient_reg_bet.nii.gz, T2_reorient_reg_bet.nii.gz and FLAIR_reorient_reg_bet.nii.gz to proceed.

Run HD-GLIO

HD-GLIO provides two main scripts: hd_glio_predict and hd_glio_predict_folder.

Predicting a single case

hd_glio_predict can be used to predict a single case. It is useful for exploration or if the number of cases to be procesed is low. Here is how to use it:

hd_glio_predict -t1 T1_FILE -t1c CT1_FILE -t2 T2_FILE -flair FLAIR_FILE -o OUTPUT_FILE

T1_FILE, CT1_FILE, T2_FILE, FLAIR_FILE and OUTPUT_FILE must all be niftis (end with .nii.gz). The four input files must be preprocesed as specified in How to use it - Prerequisites (ses above).

Predicting multiple cases

hd_glio_predict_folder is useful for batch processing, especially if the number of cases to be processed is large. By interleaving preprocessing, inference and segmentation export we can speed up the prediction significantly. Furthermore, the pipeline is initialized only once for all cases, again saving a lot of computation time. Here is how to use it:

hd_glio_predict_folder -i INPUT_FOLDER -o OUTPUT_FOLDER

INPUT_FOLDER hereby contains the T1, T1c, T2 and FLAIR images. In order to ensure that HD-GLIO correctly assigns filenames to modalities, you must apply the following naming convention to your data

• INPUT_T1: PATIENT_IDENTIFIER_0000.nii.gz
• INPUT_CT1: PATIENT_IDENTIFIER_0001.nii.gz
• INPUT_T2: PATIENT_IDENTIFIER_0002.nii.gz
• INPUT_FLAIR: PATIENT_IDENTIFIER_0003.nii.gz

Hereby, PATIENT_IDENTIFIER can be anything. You can use an arbitrary number of patients (by using a different PATIENT_IDENTIFIER for each patient). Predicted segmentations will be saved as PATIENT_IDENTIFIER.nii.gz in the OUTPUT_FOLDER

¹https://fsl.fmrib.ox.ac.uk/fsl/fslwiki/Fslutils