Docker container for HD-BET.
The models are downloaded into the container at build time, to allow the container to be run offline. The model cache directory is set outside the user's home directory for compatibility with Singularity.
Run without args to see usage. Note input requirements from HD-BET:
INPUT_FILENAME must be a nifti (.nii.gz) file containing 3D MRI image data. 4D image sequences are not supported (however can be splitted upfront into the individual temporal volumes using fslsplit). INPUT_FILENAME can be either a pre- or postcontrast T1-w, T2-w or FLAIR MRI sequence. Other modalities might work as well. Input images must match the orientation of standard MNI152 template! Use fslreorient2std upfront to ensure that this is the case.
To enable multi-threading, pass the environment variable OMP_NUM_THREADS=N
at run time.
I have not tested the GPU on Docker. On Singularity, the GPU mode works well with
singularity run --nv
, and allows use of -mode accurate -tta 1
for better results in
less time than -mode fast
requires on CPU.
If using HD-BET in published research, please cite
Isensee F, Schell M, Pflueger I, Brugnara G, Bonekamp D, Neuberger U, Wick A, Schlemmer HP, Heiland S, Wick W, Bendszus M, Maier-Hein KH, Kickingereder P. Automated brain extraction of multisequence MRI using artificial neural networks. Hum Brain Mapp. 2019 Dec 1;40(17):4952-4964. doi: 10.1002/hbm.24750. Epub 2019 Aug 12. PMID: 31403237; PMCID: PMC6865732.