Perelman School of Medicine at the University of Pennsylvania

Section for Biomedical Image Analysis (SBIA)

participating with CBICA

Multimodal Brain Tumor Segmentation Challenge 2018

ScopeRelevanceTasksDataEvaluationParticipation Summary Data RequestPrevious BraTSPeople

BraTS Algorithmic Repository

In a nutshell: We would like to have your algorithms in a Docker container, as well as in their original source code. We intend to run your dockerized algorithm on and additional dataset to blindly compare segmentation results, and to make all contributed Docker containers available through the upcoming BraTS algorithmic repository. Your source code will not be distributed and will only be used internally by the BraTS organizers, as a proof of code ownership (contact us if you cannot share your source code).

Your algorithm should be able to generate a tumor segmentation on any multimodal brain scan that is preprocessed like a BraTS test subject. To allow for fair comparisons and assessing performance differences across algorithms, you are expected to indicate what training set you have used, during training and/or design of your algorithm, and in the case of private datasets then a description of those datasets.

What we need from you

- Your version of the Docker run command below with the name of your container and the script call
- Your requirements as far as resource usage is concerned
- Additional notes regarding functionality
- Links to Docker Hub or another platform where the image is available
- The source code (irrespective of the language it is written, e.g. MatLab, Python, C++, etc), with some comments included stating the dependencies. Note that this is only for internal evaluation, in order to confirm that the code is something you have written.


In most cases, containerization of your code is a simple and straightforward procedure. We will provide an example container with functional code soon. Furthermore, many containers are available on the internet to be used as a basis, probably including your favorite programming environment and neuroimaging tools.

The concept of containerization is to simplify the deployment of applications. To ensure maximum compatibility and (potentially) distributability, we want to collect algorithms using a container approach with Docker. Docker is a technique to do so and will be used for the BraTS Algorithmic Repository.

Docker can be used to “wrap” your entire segmentation method (including all dependencies) into a single container. This container can be run as if it would be a single standalone application, anywhere, on any platform. Because your method and all dependencies are included in the container, the method is guaranteed to run exactly the same at all times.

Dockerization is a very popular concept and has been used successfully in previous MICCAI challenges, e.g., in the MSSEG challenge and the WMH Segmentation Challenge (The current text is also based on info from the latter). Docker Hub provides a large overview of existing Docker containers (base images), that can be used to build your own container. Furthermore, many popular programming environments and image analysis methods have Dockerfiles available.


Data access

All test sets will be identical to the 2017 or 2018 test data that you have already processed. They are co-registered, skull-stripped, resampled to 1mm^3 isotropic resolution, and aligned to the SRI space. Data will be in NIfTI GZIP Compressed Tar Archive (.nii.gz) format, with all header information except the spatial resolution removed, and the individual volumes will be named ‘fla.nii.gz’, ‘t1c.nii.gz’, ‘t1.nii.gz’, ‘t2.nii.gz’. You can use any of the BraTS training or testing image volumes to check whether your Docker image runs as expected.

Because your container runs in an isolated environment, the data needs to be mapped into the container. The input data files, i.e., the ‘fla.nii.gz’, ‘t1c.nii.gz’, ‘t1.nii.gz’, ‘t2.nii.gz’ volumes, will be linked to /data and your segmentations must be placed in /data/results. Results should be a NIfTI file with the same resolution as the input data. Please call the resulting file "tumor_’your_image’_class.nii.gz", where ‘your_image’ is an eight digit identifier of your algorithm. If your algorithm returns probabilities as well, you can return them accordingly, and name them, e.g., "tumor_’your_image’_prob_4.nii.gz" for results of class 4. If your algorithm returns tissue classes, please use “tissue_’your_name’_wm.nii.gz” for white matter (*‘_gm.nii.gz’ and ‘*_csf.nii.gz’ for the other two).

Please also see this for additional information.

There should be no interaction with the container required other than running the Docker command below, i.e., we can only support fully automatic algorithms.


Computing environment & resources, GPUs

We will run your container on a selection of cases from an additional dataset, based on the quality of their skull-stripping. Docker can set resource limits on containers. Please give us an indication how many CPUs and how much RAM is needed for you method, and what the resulting computation time will be.

In our first instalment, we would like to run all code CPU-only to retain compatibility. Docker does not yet support GPU mapping on all platforms, so please provide a CPU-only version of you code. If you really want/need to use a GPU, please contact us.


Running your Docker / Docker commands

Your container will be run with the following commands:

docker run −v <directory>:/data −it <your image> <your script call>

“directory" will be our test directory containing the four modalities and the empty folder for your results.

"your image" is the name of your Docker image.

"your script call" is the script that should be called when running the container.


Examples and Assistance

To help you containerize your segmentation method with Docker, we have provided some examples. Some exemplary Dockerfiles can be found on Github.

If you are unsure whether your method can be containerized or how to proceed, please contact us in advance. We will try to help you with Docker.

We have created a PDF-Document containing the exact specifications which you can download from here.


Feel free to send any communication related to the BraTS challenge to