DSB2017 challenge: grt123 processor

This repository uses a modified grt123 solution to implement a DIAG-processor docker image. The processor can be applied to a list of 3D lung images and the algorithm will:

• Compute cancer scores

  Cancer scores indicate the likelihood of the image containing a cancer.

• Designate regions of interest

  Regions of interest are locations in the input image that the algorithm analyzed to compute its "cancer score".

Original grt123 solution can be found here.

Building

The processor uses docker to build a containerized runtime environment for the grt123 algorithm to run in.

If you are in a linux environment with docker installed, issue the command:

./build_processor_docker.sh processor [--version-tag VERSION_TAG] [--git-commit GIT_COMMIT_ID] [--push] [-h|--help]

to build the DIAG docker image.

• --version-tag VERSION_TAG labels the docker image with a version tag, if not specified the tag is left empty creating latest.
• --git-commit GIT_COMMIT_ID overrides the internally baked gid commit id used by the processor, if not specified will attempt to look for it in the .git folder.
• --push will attempt to push the images to the private DIAG-docker-registry after building.
• -h, --help will display some a short help message.

Running

The algorithm requires the following hardware configuration to run:

• 8 CPUs
• 30GB of memory
• 1 CUDA capable GPU with at least 4 GB of memory

The measured runtime for the algorithm on a system matching the specifications is:

• 40 seconds startup time plus ...
  • ... 40 seconds per scan when processing many scans at once
  • ... 60 seconds when processing single scans

Input

The input directory is a directory lung image volumes. Images are:

• MHA or MHD+ZRAW MetaIO files containing image volumes

Output

• Per processed image an XML file with candidates and cancer scores
• A JSON file summarizing all findings