SynthSegContainer

Containerized SynthSeg.

Building with docker

Before building, download the models as instructed in the SynthSeg installation instructions and place them in synthseg_models/.

Then, build with one of the available Dockerfiles.

Dockerfile.conda

Installs tensorflow-gpu from conda, which includes GPU support. This works for Singularity containers with singularity run --nv.

This also runs without the GPU. MKL support is disabled, which extends execution time (to about 8-10 min) but massively reduces the memory requirements.

Dockerfile.conda-cuda

As above, but use a cuda runtime layer. Possibly useful for use with the GPU in Docker, rather than Singularity.

Dockerfile.conda.mkl

Installs tensorflow-mkl, which uses MKL optimizations. This is 2-3x faster than the non-MKL tensorflow, but it requires MUCH more memory. The memory use can be reduced with the environment variable MKL_DISABLE_FAST_MM but even still, with the default crops, memory use peaks at around 40 Gb.

Dockerfile.tfsrc

This builds without CUDA, but enhances CPU optimization by building tensorflow from source. This takes longer and by default optimizes the tensorflow binary for "native" architecture. This is customizable by changing the configuration options in tf_configure.bazelrc.in.

I'm not sure how much this helps, it doesn't seem much faster than the conda tensorflow-mkl version.

This dockerfile has not been updated to reflect the simplified requirements provided by SynthSeg developers in March 2023.

Usage

The entrypoint to the container is the SynthSeg prediction script SynthSeg_predict.py. Run without args to see usage.

Docker containers

Available on DockerHub.

Singularity

For GPU support, build from synthseg:conda-latest and run with singularity --nv.

Licensing and citations

SynthSeg license.

The authors ask that if it is used in published research, to cite:

SynthSeg: Domain Randomisation for Segmentation of Brain MRI Scans of any Contrast and Resolution B. Billot, D.N. Greve, O. Puonti, A. Thielscher, K. Van Leemput, B. Fischl, A.V. Dalca, J.E. Iglesias (https://pubmed.ncbi.nlm.nih.gov/36857946/)

For cortical parcellation, automated QC, or robust fitting, please also cite the following paper:

Robust Segmentation of Brain MRI in the Wild with Hierarchical CNNs and no Retraining B. Billot, M. Colin, S.E. Arnold, S. Das, J.E. Iglesias MICCAI 2022

cookpa/SynthSegContainer