NiChart_DLMUSE

Overview

NiChart_DLMUSE is a package that allows the users to process their brain imaging (sMRI) data easily and efficiently.

NiChart_DLMUSE offers easy ICV (Intra-Cranial Volume) mask extraction, and brain segmentation into ROIs. This is achieved through the DLICV and DLMUSE methods. Intermediate step results are saved for easy access to the user.

Given an input (sMRI) scan, NiChart_DLMUSE extracts the following:

1. ICV mask
2. Brain MUSE ROI segmentation
3. ROI volumes in a .csv format
4. Individual ROI mask (optionally).

This package uses nnU-Net v2 as a basis model architecture for the deep learning parts, and various other libraries.

Installation

As a locally installed package

You can install NiChart DLMUSE from source:

$ git clone https://github.com/CBICA/NiChart_DLMUSE.git
  cd NiChart_DLMUSE
  python3 -m pip install -e .

Or from out latest stable PyPI wheel:

$ pip install NiChart_DLMUSE

(If needed for your system) Install PyTorch with compatible CUDA. You only need to run this step if you experience errors with CUDA while running NiChart_DLMUSE. Run "pip uninstall torch torchaudio torchvision". Then follow the PyTorch installation instructions for your CUDA version. Specific versions are needed for full compatibility. On Linux, download Torch 2.3.1, on Windows install 2.5.1. For example, after installing NiChart_DLMUSE:

$ pip uninstall torch torchvision torchaudio
  pip install torch==2.3.1 --index-url https://download.pytorch.org/whl/cu121

Run NiChart_DLMUSE

Example usage below:

$ NiChart_DLMUSE   -i                    /path/to/input     \
                   -o                    /path/to/output    \
                   -d                    cpu/cuda/mps

Docker/Singularity/Apptainer-based build and installation

Docker build

The package comes already pre-built as a docker container, for convenience. Please see Usage for more information on how to use it. Alternatively, you can build the docker image locally from the source repo, like so:

$ docker build -t cbica/nichart_dlmuse .

(OUTDATED) Singularity/Apptainer build

Singularity and Apptainer images can be built for NiChart_DLMUSE, allowing for frozen versions of the pipeline and easier installation for end-users. Note that the Singularity project recently underwent a rename to "Apptainer", with a commercial fork still existing under the name "Singularity" (confusing!). Please note that while for now these two versions are largely identical, future versions may diverge. It is recommended to use the AppTainer distribution. For now, these instructions apply to either.

First install the container engine. Then, from the cloned project repository, run:

$ singularity build nichart_dlmuse.sif singularity.def

This will take some time, but will build a containerized version of your current repo. Be aware that this includes any local changes! The nichart_dlmuse.sif file can be distributed via direct download, or pushed to a container registry that accepts SIF images.

Usage

Pre-trained nnUNet models for the skull-stripping can be found in HuggingFace nichart/DLICV and segmentation tasks can be found in HuggingFace nichart/DLMUSE. Feel free to use it under the package's license.

As a locally installed package

A complete command would be (run from the directory of the package):

$ NiChart_DLMUSE -i                    /path/to/input     \
                 -o                    /path/to/output    \
                 -d                    cpu/cuda/mps

For further explanation please refer to the complete documentation:

$ NiChart_DLMUSE -h

Troubleshooting model download failures

Our model download process creates several deep directory structures. If you are running on Windows and your model download process fails, it may be due to Windows file path limitations.

To enable long path support in Windows 10, version 1607, and later, the registry key HKEY_LOCAL_MACHINE\SYSTEM\CurrentControlSet\Control\FileSystem LongPathsEnabled (Type: REG_DWORD) must exist and be set to 1.

If this affects you, we recommend re-running NiChart_DLMUSE with the --clear_cache flag set on the first run.

Using the docker container

Using the file structure explained above, an example command using the docker container is the following:

# Pull the image for your CUDA version (as needed)
$ CUDA_VERSION=11.8 docker pull cbica/nichart_dlmuse:1.0.1-cuda${CUDA_VERSION}
# or, for CPU:
$ docker pull cbica/nichart_dlmuse:1.0.1

# Run the container with proper mounts, GPU enabled
# Place input in /path/to/input/on/host.
# Replace "-d cuda" with "-d mps" or "-d cpu" as needed...
# or don't pass at all to automatically use CPU.
# Each "/path/to/.../on/host" is a placeholder, use your actual paths!
$ docker run -it --name DLMUSE_inference --rm
    --mount type=bind,source=/path/to/input/on/host,target=/input,readonly
    --mount type=bind,source=/path/to/output/on/host,target=/output
    --gpus all cbica/nichart_dlmuse -d cuda

(OUTDATED) singularity container

To use the singularity container, you can just do:

$ singularity run --nv --containall --bind /path/to/.\:/workspace/ nichart_dlmuse.simg NiChart_DLMUSE -i /workspace/temp/nnUNet_raw_data_base/nnUNet_raw_data/ -o /workspace/temp/nnUNet_out -p structural --derived_ROI_mappings_file /NiChart_DLMUSE/shared/dicts/MUSE_mapping_derived_rois.csv --MUSE_ROI_mappings_file /NiChart_DLMUSE/shared/dicts/MUSE_mapping_consecutive_indices.csv --nnUNet_raw_data_base /workspace/temp/nnUNet_raw_data_base/ --nnUNet_preprocessed /workspace/temp/nnUNet_preprocessed/ --model_folder /workspace/temp/nnUNet_model/ --all_in_gpu True --mode fastest --disable_tta