fbk-nilab / bl_app_dbb_disseg

Deep Learning based segmentation for distorted brains

bl_app_dbb_DisSeg

This application implements a 3D U-Net for brain tissue segmentation. This implementation of 3D U-Net is used as the baseline method in the Distorted Brain Benchmark (DBB) for automatic tissue segmentation in paediatric patients.

Author

Gabriele Amorosino ([email protected])

Citation

If you use this code for your research please cite:

Gabriele Amorosino, Denis Peruzzo, Daniela Redaelli, Emanuele Olivetti, Filippo Arrigoni, Paolo Avesani,
DBB - A Distorted Brain Benchmark for Automatic Tissue Segmentation in Paediatric Patients,
NeuroImage, 2022, 119486, ISSN 1053-8119, https://doi.org/10.1016/j.neuroimage.2022.119486.

Running the Brainlife App

You can run the BrainLife App DBB_DisSeg on the brainlife.io platform via the web user interface (UI) or using the brainlife CLI. With both of these two solutions, the inputs and outputs are stored on the brainlife.io platform, under the specified project, and the computations are performed using the brainlife.io cloud computing resources.

On Brainlife.io via UI

You can see DBB_DisSeg currently registered on Brainlife. Find the App on brainlife.io and click "Execute" tab and specify dataset e.g. "DBB Distorted Brain Benchmark".

On Brainlife.io using CLI

Brainlife CLI could be installed on UNIX/Linux-based system following the instruction reported in https://brainlife.io/docs/cli/install/.

The first time you use the BrainLife CLI on a machine, you need to log in with the brainlife.io credentials

bl login

You can run the App with CLI as follow:

bl app run --id 60f83394b99111089ab55f80  --project <project_id> --input t1:<t1_object_id> --input mask:<mask_object_id> 

the output is stored in the reference project specified with the id <project_id>. You can retrieve the object_id using the command bl data query, e.g to get the id of the mask file for the subject 0001 :

bl data query --subject 0001 --datatype neuro/mask --project <projectid>

If not present yet, you can upload a new file in a project using bl data upload. For example, in the case of T1-w file, for the subject 0001 you can run:

bl data upload --project <project_id> --subject 0001 --datatype "neuro/anat/t1w" --t1 <full_path>

Running the code locally

You can run the code on your local machine by git cloning this repository. You can choose to run it with dockers, avoiding to install any software except for singularity. Furthermore, you can run the original script using local software installed.

Run the script using the dockers (recommended)

It is possible to run the app locally, using the dockers that embedded all needed software. This is exactly the same way that apps run code on brainlife.io

Inside the cloned directory, create config.json with something like the following content with the fullpaths to your local input files:

{   
    "t1": "./t1.nii.gz",
    "mask": "./mask.nii.gz"
}

Launch the app by executing main.

./main

To avoid using the config file, you can input directly the fullpath of the filess using the script main.sh:

main.sh <t1.ext> <mask.ext> [<outputdir>] [--no-histmatch]

where the --no-histmatch option can be used if you want to avoid performing histogram matching on the reference image.

Script Dependecies

The App needs singularity to run.

Output

The output of bl_app_dbb_DisSeg are the predicted segmentation volume of the 3D U-Net and a json file describing the labels of the segmented volume.

The files are stored in the working directory, under the folder ./segmentation with the name segmentation.nii.gz , for the semgnetaion volume and label.json, for the json file.

Run the script (local installed softwares)

Clone this repository using git on your local machine to run this script.

Usage

main_local.sh <t1.ext> <mask.ext> [<outputdir>] [--no-histmatch]

Output

The outputs of bl_app_dbb_DisSeg are the predicted segmentation volume of the 3D U-Net and a json file describing the labels of the segmented volume.

The files are stored in the working directory, under the folder ./segmentation with the name segmentation.nii.gz , for the semgnetaion volume and label.json, for the json file.

Script Dependecies

In order to use the script, the following software must be installed:

• ANTs, Advanced Normalization Tools (version >= 2.1.0)

It is also necessary that Python 2.7.x is installed, with the following modules:

• scikit-image=0.14.2
• requests=2.22.0
• nibabel=2.5.1
• pydicom=1.3.0
• tqdm=4.38.0
• pyvista
• vtk=8.1.2
• libnetcdf=4.6.2
• tensorflow-gpu=1.10.0
• Cuda toolkit 9.0 or 9.1

It is suggested to install python modules using conda.

conda install -c anaconda python=2.7 scikit-image=0.14.2 \
      && conda install -c anaconda python=2.7 requests=2.22.0 \
      && conda install -c conda-forge nibabel=2.5.1 \
      && conda install -c conda-forge pydicom=1.3.0 \
      && conda install -c conda-forge tqdm=4.38.0 \
      && conda install -c anaconda tensorflow-gpu=1.10.0 cudatoolkit=9.0  \
      && conda install -c conda-forge pyvista "vtk=8.1.2" "libnetcdf=4.6.2"

Run test on DBB Distorted Brain Benchmark test set

You can run the tool to reproduce the results on the test set of DBB Distorted Brain Benchmark using the script with dockers:

run_test.sh <download_dir> <output_dir>

or with local softwares installed:

run_test_local.sh <download_dir> <output_dir>

The script performs the automatic download of the published testset (using brainlife CLI) of the DBB benchmark (https://doi.org/10.25663/brainlife.pub.24) in the folder <download_dir>. Then, the script predicts the segmentation volume for each subject and stores the results in <output_dir>. Finally, the script computes the dice score between the predicted segmentation and the test set ground-truth, and create the final csv file, named average_dice_score.csv (stored in <output_dir>), reporting the average dice score across the subjects for each label of the segmented volumes.