Simple Skull Stripping (S3)

Skull stripping tool with optional approximate segmentation of brain tissue based on atlas registration
Input: patient head scan
Output:
- skull stripped scan,
- binary brain mask,
- probabilistic segmentation of grey matter (GM), white matter (WM) and cerbospinal fluid (CSF).
The solver works with any input head scan, however T1 non-contrast scan is preferred. The reason is that brain pathologies like lesions are less pronounce on T1 non-contrast scan, compare to other modalities like T1-contrast or FLAIR scan.
The tissue segmentations provide good approximation for healthy brains. However, they might not correctly estimate tissue in the regions with significant pathologies, like tumor.

Method

The method is based on atlas registrations. Here we use atlas of normal adult brain anatomy extracted from sri24. The method consist of the following steps:

Input: Head MRI scan (shown in the light blue frame in Figurew)
Compute a rigid registration f that maps the brain atlas to the subject (Fig. blue box)
Apply f to map the atlas brain mask and tissue segmentations to the subject scan (purple box). The registered mask provides a coarse approximation of the subject brain mask.
Use the coarse brain mask to skull-strip the subject and the atlas scan (yellow box).
Compute a non-rigid registration g that maps the masked atlas to the masked subject, accounting for the subject specific brain morphology (green box).
Apply g to map the atlas tissue segmentations to the subject anatomies (orange box).
Fuse the subject tissue segmentations to create a refined brain mask (orangebox). Compute a 95% percentile of the sum of all tissue segmentations and exclude values smaller than the lower bound of the 95% percentiles.
Apply the refined brain mask to the input MRI scan (orange box).
Output: Coarse and refined brain mask,skull-stripped input scan and tissue segmentations.

Distributions:

The software comes with three branches

master - python3 implementation
s3_p27 - same as the master brunch but python2.7 implementation
s3_ants - experimental branch, which use only ANTS registration (not recommended)

Installation:

First install packages from requirements.txt

pip install -r requirements.txt

Install ANTs with instructions provided at https://github.com/ANTsX/ANTs/wiki/Compiling-ANTs-on-Linux-and-Mac-OS
Install NiftyReg with instructions provided at http://cmictig.cs.ucl.ac.uk/wiki/index.php/NiftyReg_install. The software come with experimental branch called s3_ants which does not require NiftyReg installation, however this branch is still under development and might not be stable. However, if you can't install NiftyReg you can use this branch for skull stripping.
Set the path to the installed libraries by updating your ~/.profile or ~/.basrc file with the following lines, where the first line contains path to YOUR library:

export ANTSPATH=/home/jana/Work/Ants/stnava-ANTs-a430c38/antsbin/bin/
export PATH=$PATH:$ANTSPATH

export NIFTYREG_INSTALL=/home/jana/Work/libs/nifty_reg-1.3.9/niftyreg_install
export PATH=${PATH}:${NIFTYREG_INSTALL}/bin

Run

Parameters:

-i : Path to the input modality, nifti file (.nii or nii.gz) (Mandatory)
-o : Path to the output folder (Optional)
-t : Use this when the tissue registrations for white matter, gray matter and cerebral spinal fluid are required (Optional)

Example

Folder s3/example/ contains test scan called T1.nii To apply the s3 method to the example scan:

cd s3
python s3.py -i example/T1.nii

This command would apply the skull stripping procedure on the input file "t1.nii" file and stores the output, the brain mask and stripped brain scan, in the same folder as the input file.

To save results to differen than the input folder, use flag -o

python s3.py -i example/T1.nii -o output

Where output is the name of the output folder. If the specified output folder does not exist, it will be created.

To enable computation of tissue segmentation use flag -t:

python s3.py -i example/T1.nii -o output/  -t

This command performs skulls stripping of input image, and outputs the brain mask, skull-stripped scan, soft segmentations of white, grey matter and csf.

References

Please cite: Lipkova et al., Personalized Radiotherapy Design for Glioblastoma: Integrating Mathematical Tumor Models, Multimodal Scans and Bayesian Inference., IEEE Transaction on Medical Imaging, (2019), (also available at https://arxiv.org/pdf/1807.00499.pdf)

Acknowledgement

Esther Alberts, Enes Senel, Florian Kofler

🐼

MingR-Ma / s3