tsantini/wmh7t repository overview

⁠wmh7T

wmh7T is a nnUnet-based FLAIR-only Docker application for automated white matter hyperintensity segmentation tailored for 7T MRI (it may work well for 3T and 1.5T as well!). It supports NVIDIA GPU and CPU inference for individual NIfTI images or folders.

NVIDIA GPU inference is substantially faster than CPU inference and is recommended when a compatible GPU is available. CPU execution is provided for compatibility.

⁠Usage

⁠NVIDIA GPU — single FLAIR image

Run this command from the directory containing the input NIfTI image:

docker run --gpus all --rm \
  --shm-size=8g \
  --user "$(id -u):$(id -g)" \
  -v "$PWD:/data" \
  tsantini/wmh7t:latest \
  /data/input_flair.nii.gz \
  /data/input_flair_wmh7T.nii.gz \
  --device cuda

⁠NVIDIA GPU — folder of FLAIR images

Run this command from the directory containing the NIfTI images:

docker run --gpus all --rm \
  --shm-size=8g \
  --user "$(id -u):$(id -g)" \
  -v "$PWD:/data" \
  tsantini/wmh7t:latest \
  /data \
  /data \
  --device cuda \
  --batch-size 32

For each input FLAIR image, wmh7T produces a binary NIfTI segmentation using the suffix _wmh7T.

⁠NVIDIA GPU + Apptainer — folder of FLAIR images

apptainer pull wmh7t_latest.sif docker://tsantini/wmh7t:latest

apptainer run --nv \
  --bind "$PWD:/data" \
  wmh7t_latest.sif \
  /data/input_flair.nii.gz \
  /data/input_flair_wmh7T.nii.gz \
  --device cuda
  --batch-size 32

⁠CPU — single FLAIR image

docker run --rm \
  --shm-size=8g \
  --user "$(id -u):$(id -g)" \
  -v "$PWD:/data" \
  tsantini/wmh7t:latest \
  /data/input_flair.nii.gz \
  /data/input_flair_wmh7T.nii.gz \
  --device cpu \
  --threads 8

CPU inference is supported for systems without an NVIDIA GPU, but it is considerably slower.

⁠Project status and support

wmh7T is a work in progress. Results should be reviewed before being used in research analyses.

If you encounter a problem or unexpected segmentation, please submit a ticket through the wmh7T GitHub issue tracker⁠.