The MIAC Automated Region Segmentation (MARS) for White Matter Hyperintensities (WMH) is a state-of-the-art, deep learning-based segmentation tool that has undergone systematic validation, both technically and clinically.

This repository includes ready-to-use, pre-built container images of two methods, based on nnU-Net or MD-GRU, along with the code needed to build these images.

The methods are described in detail in the publication currently available as a pre-print:

Gesierich et al. (2025). Technical and Clinical Validation of a Novel Deep Learning-Based White Matter Hyperintensity Segmentation Tool. Cereb Circ Cogn Behav. DOI: 10.1016/j.cccb.2025.100393.

Please ensure to cite this publication when using the methods, and please note that the license does not cover any commercial use (defined as use for which any financial return is received). Please also cite the underlying deep learning method (nnU-Net, DOI: 10.1038/s41592-020-01008-z or MD-GRU, DOI: 10.1007/978-3-319-75238-9_3).

Detailed instructions are provided below, here is the quick start version:

# 1. Pull the container image and save as Apptainer .sif file 
apptainer build mars-wmh-nnunet.sif docker://ghcr.io/miac-research/wmh-nnunet:latest

# 2. Run inference with on a FLAIR and T1w image in the current working directory using Apptainer and GPU
apptainer run -B $(pwd) --nv mars-wmh-nnunet.sif --flair FLAIR.nii.gz --t1 T1w.nii.gz

Ready-to-use, pre-built images are available for download from the Github container registry. The images have been tested with Apptainer and Docker.

In general, we recommend the nnU-Net algorithm (please see our publication for a detailed comparison between the two algorithms) and using Apptainer (the standard container tool for scientific computing).

The WMH segmentation requires two inputs, a FLAIR image and a T1-weighted image, both in NIfTI-1 data format. We recommend dcm2niix for DICOM to NIfTI conversion. Ensure that the entire brain is covered by the field of view. By default, the T1-weighted image will be registered to the FLAIR. However, already registered images can be provided as well (see option -skipRegistration).

• FLAIR image: A 3D FLAIR with 1 mm isotropic resolution is recommended.§D images with different resolution will be resliced to 1mm isotropic before prediction, while WMH masks are returned in the original resolution. A 2D FLAIR with up to 3 mm slice thickness can also be used. Images with thicker slices are not recommended.
• T1-weighted image: The recommended resolution is 1 mm isotropic. The image must have been acquired without the use of a contrast agent.

While the inference can be run on CPU (>8 cores recommended), an NVIDIA GPU will greatly accelerate the calculation. The pre-built images support a wide range of NVIDIA GPUs from compute capability 5.0 (Maxwell, 2014) to 12.0 (Blackwell, 2024), please see the table for details:

*While not officially supported, our testing indicates that the MD-GRU image can run on Blackwell. However, the brainmask will be generated using the CPU, which is much slower and may produce slightly different results compared to an all-GPU pipeline.

# 1. Pull the container image and save as .sif file 
apptainer build mars-wmh-nnunet.sif docker://ghcr.io/miac-research/wmh-nnunet:latest

# 2. Run inference on a pair of FLAIR and T1w images in the current working directory using GPU (flag "--nv")
apptainer run -B $(pwd) --nv mars-wmh-nnunet.sif --flair FLAIR.nii.gz --t1 T1w.nii.gz

# For advanced usage, see available command line options:
apptainer run mars-wmh-nnunet.sif -h

# 1. Pull the container image into your local registry
docker pull ghcr.io/miac-research/wmh-nnunet:latest
docker tag ghcr.io/miac-research/wmh-nnunet:latest mars-wmh-nnunet:latest

# 2. Run inference on a pair of FLAIR and T1w images in the current working directory using GPU (flag "--gpus all")
docker run --rm --gpus all -v $(pwd):/data mars-wmh-nnunet:latest --flair /data/FLAIR.nii.gz --t1 /data/T1w.nii.gz

# For advanced usage, see available command line options:
docker run --rm mars-wmh-nnunet:latest -h

# 1. Pull the container image and save as .sif file 
apptainer build mars-wmh-mdgru.sif docker://ghcr.io/miac-research/wmh-mdgru:latest

# 2. Run inference on a pair of FLAIR and T1w images in the current working directory using GPU (flag "--nv")
apptainer run -B $(pwd) --nv mars-wmh-mdgru.sif --flair FLAIR.nii.gz --t1 T1w.nii.gz

# For advanced usage, see available command line options:
apptainer run mars-wmh-mdgru.sif -h

# 1. Pull the container image into your local registry
docker pull ghcr.io/miac-research/wmh-mdgru:latest
docker tag ghcr.io/miac-research/wmh-mdgru:latest mars-wmh-mdgru:latest

# 2. Run inference on a pair of FLAIR and T1w images in the current working directory using GPU (flag "--gpus all")
docker run --rm --gpus all -v $(pwd):/data mars-wmh-mdgru:latest --flair /data/FLAIR.nii.gz --t1 /data/T1w.nii.gz

# For advanced usage, see available command line options:
docker run --rm mars-wmh-mdgru:latest -h

If you prefer to build the container images yourself, you can use the provided files in the mdgru and nnunet folders.

1. Download the corresponding files (Dockerfile and Python scripts) and place them into a folder.
2. In this folder, run docker build -t mars-wmh-{mdgru/nnunet} .

The script MARS-WMH_extract_volumes.py allows you to extract and aggregate the voxel count, volume and cluster count of the segmented WMH from one or more NIfTI files corresponding to MARS segmentation label-maps.
The minimum input required is the directory in which the data for one or more patients has been segmented. For complete usage information, call the script without any arguments.
Run the script in your local Python environment or use Apptainer and run the script within the MARS-WMH container with its Python installation, e.g.:
apptainer exec -B /path/to/segmented/data mars-wmh-nnunet.sif python MARS-WMH_extract_volumes.py /path/to/segmented/data

Please note the license terms of software components that we redistribute within our container images:

• HD-BET
• 3D Slicer
• nnU-Net
• MD-GRU

Development and maintenance of this software is funded by the Medical Image Analysis Center (MIAC AG).