Deployment#
Arcana provides tools for deploying pipelines in Docker containers that can be run in XNAT’s container service. Pipelines can be built done on an individual basis or as part of a wider a suite (e.g. Australian Imaging Service Pipelines). As well as building Docker images, the deployment workflow includes procedures to test and generate documentation.
Command definitions#
The XNAT container service uses command configuration files
saved in the org.nrg.commands image label to resolve metadata for the pipelines
that available on a given Docker image. The XnatViaCS.generate_xnat_command()
method is used to generate the JSON metadata to be saved in this field.
There are four key fields that will determine the functionality of the command (the rest are metadata fields that are exposed to the XNAT UI):
pydra_task
inputs
outputs
parameters
The pydra_task
keyword argument should be the path to an installed
Python module containing a Pydra task followed by a colon and the name of
the task, e.g. pydra.tasks.fsl.preprocess.fast:Fast
. Note that Arcana
will attempt to resolve the package that contains the Pydra task and install the
same version (including local development versions) within the Anaconda environment
in the image. inputs
and parameters
expose text boxes in the XNAT dialog when
the pipelines are run. outputs
determine where the outputs will
be stored in the XNAT data tree.
Inputs prompt the user to enter selection criteria for input data and are used by the entrypoint of the Docker containers to add source columns to the dataset (see Frames: Rows and Columns). They are specified by 4-tuple consisting of
name of field in the pydra task input interface
format required by pydra task
description of input that will be exposed to the XNAT UI
the row row_frequency of the column (see Spaces and Frames: Rows and Columns)
Parameters are passed directly through the pipeline add method (see Pydra workflows) that is run in the container, and consist of a 2-tuple with
name of field in the pydra task input interface
description of parameters that will be exposed to the XNAT UI
Outputs do not show up in the XNAT dialog and are specified by a 3-tuple:
name of field in the pydra task output interface
format produced by pydra task
destination path (slashes are permitted interpreted as a relative path from the derivatives root)
from arcana.data.stores.medimage import XnatViaCS
from arcana.data.spaces.medimage import Clinical
from arcana.data.formats.medimage import NiftiGz
xnat_command = XnatViaCS.generate_xnat_command(
pipeline_name='example_pipeline',
pydra_task='pydra.tasks.fsl.preprocess.fast:FAST',
image_tag='example/0.1',
description=(
"FAST (FMRIB's Automated Segmentation Tool) segments a 3D image of "
"the brain into different tissue types (Grey Matter, White Matter, "
"CSF, etc.), whilst also correcting for spatial intensity variations "
"(also known as bias field or RF inhomogeneities)."),
version='6.0-1',
info_url='https://fsl.fmrib.ox.ac.uk/fsl/fslwiki/FAST',
inputs=[
('in_files', NiftiGz, 'File to segment', 'session'),
('number_of_classes', int, 'Number of classes', 'session')],
outputs=[
('tissue_class_files', NiftiGz, 'fast/tissue-classes'),
('partial_volume_map', NiftiGz, 'fast/partial-volumes'),
('partial_volume_files', NiftiGz, 'fast/partial-volume-files'),
('bias_field', NiftiGz, 'fast/bias-field'),
('probability_maps', NiftiGz, 'fast/probability-map')],
parameters=[
('use_priors', 'Use priors'),
('bias_lowpass', 'Low-pass filter bias field')],
configuration=[ # If different from the Pydra task
('output_biasfield', True),
('output_biascorrected', True),
('bias_lowpass', 5.0)],
row_frequency='session')
When working with the CLI, command configurations are stored in YAML format,
with keys matching the arguments of XnatViaCS.generate_xnat_command()
.
Note
image_tag
and registry
are omitted from the YAML representation
of the commands as they are provided by the image configuration
(see Building)
Building#
Dockerfiles for pipeline images are created using Neurodocker
and can therefore work with any Debian/Ubuntu or Red-Hat based images
(using a value for package_manager
keyword argument of "apt"
for
Debian based or "yum"
for Red-Hat based). Arcana installs itself into the Docker image
within an Anaconda environment named “arcana”. Therefore, it won’t typically
conflict with packages on existing Docker images for third-party pipelines
unless they are also installed using Anaconda.
Extending the YAML format used to define the command configurations, the full configuration required to build an XNAT docker image looks like
pkg_name: FSL
pkg_version: &pkg_version '6.0.1'
wrapper_version: '1'
authors:
- name: Thomas G. Close
email: thomas.close@sydney.edu.au
base_image: !join [ 'brainlife/fsl:', *pkg_version ]
info_url: https://fsl.fmrib.ox.ac.uk/fsl/fslwiki
package_manager: apt
system_packages:
package_templates:
- name: dcm2niix
version: v1.0.20201102
python_packages:
- name: pydra-dcm2niix
commands:
pipeline_name: fast
pydra_task: pydra.tasks.fsl.preprocess.fast:FAST
description:
FAST (FMRIBs Automated Segmentation Tool) segments a 3D image of
the brain into different tissue types (Grey Matter, White Matter,
CSF, etc.), whilst also correcting for spatial intensity variations
(also known as bias field or RF inhomogeneities).
version: 1
info_url: https://fsl.fmrib.ox.ac.uk/fsl/fslwiki/FAST
inputs:
- name: in_files
format: medimage:NiftiGzX
stored_format: medimage:Dicom
description: Anatomical image to segment into different tissues
outputs:
- name: tissue_classes
format: medimage:NiftiGz
path: fast/tissue-classes
- name: probability_maps
format: medimage:NiftiGz
path: fast/probability-map
parameters:
- name: use_priors
description: Use priors in tissue estimation
- name: bias_lowpass
description: Low-pass filter bias field
configuration:
- output_biasfield: true
- bias_lowpass: 5.0
row_frequency: session
where fields in the top-level YAML are provided as arguments to
XnatViaCS.generate_dockerfile()
, i.e.
from arcana.data.stores.medimage import XnatViaCS
xnat_command = XnatViaCS.generate_dockerfile(
xnat_commands=[xnat_command], # List of commands available on the image generated by XnatViaCS.generate_xnat_command()
python_packages=[
('pydra-fsl', '0.1.0')], # Required python packages (aside from arcana and its dependencies)
maintainer='your-email@your-institute.org', # maintainer of wrapper (i.e. not pipeline unless they are the same)
base_image='brainlife/fsl', # base Docker image
package_manager='apt', # package manager of base image
packages=[], # system packages to install (i.e. with 'apt')
extra_labels={}, # extra labels you might want to put into the image
arcana_extras=[] # install extras for Arcana package (e.g. 'test'
))
The CLI command to build the image from the YAML configuration is
$ arcana deploy build 'your-pipeline-config.yml'
Successfully built "FSL" image with ["fast"] commands
To build a suite of pipelines from a series of YAML files stored in a directory tree simply provide the root directory instead and Arcana will walk the sub-directories and attempt to build any YAML files it finds, e.g.
$ arcana deploy build 'config-root-dir'
./config-root-dir/mri/neuro/fsl.yml: FSL [fast]
./config-root-dir/mri/neuro/mrtrix3.yml: MRtrix3 [dwi2fod, dwi2tensor, tckgen]
./config-root-dir/mri/neuro/freesurfer.yml: Freesurfer [recon-all]
...
Testing#
After an image has been built successfully, it can be tested against previously generated results to check for consistency with previous versions. This can be particularly useful when updating dependency versions. Tests that don’t match previous results within a given tolerance will be flagged for manual review.
To avoid expensive runs when not necessarily (particularly within CI/CD pipelines), in the case that the provenance data saved along the generated reference data will be checked before running the pipelines. If the provenance data would be unchanged (including software dependency versions), then the pipeline test will be skipped.
Test data, both inputs to the pipeline and reference data to check against
pipeline outputs, need to be stored in separate directories for each command.
Under the pipeline data directory, there should be one or more subdirectories
for different tests of the pipeline, and in each of these subdirectories there
should be an inputs
and an outputs
directory, and optionally a YAML
file named parameters.yml
. Inside the inputs
directory there should be
file-groups named after each input of the pipeline, and likewise in the
outputs
directory there should be file-groups named after each output
of the pipeline. Any field inputs or outputs should be placed alongside the
file-groups in a JSON file called __fields__.json
.
Specifying two tests (‘test1’ and ‘test2’) for the FSL FAST example given above (see Building) the directory structure would look like:
FAST
├── test1
│ ├── inputs
│ │ └── in_files.nii.gz
│ ├── outputs
| │ └── fast
| │ ├── tissue_class_files.nii.gz
| │ ├── partial_volumes.nii.gz
| │ ├── partial-volume-files.nii.gz
| │ ├── bias-field.nii.gz
| │ └── probability-map.nii.gz
│ └── parameters.yml
└── test2
├── inputs
│ └── in_files.nii.gz
├── outputs
│ └── fast
│ ├── tissue_class_files.nii.gz
│ ├── partial_volumes.nii.gz
│ ├── partial-volume-files.nii.gz
│ ├── bias-field.nii.gz
│ └── probability-map.nii.gz
└── parameters.yml
To run a test via the CLI point the test command to the YAML configuration file and the data directory containing the test data, e.g.
$ arcana deploy test ./fast.yml ./fast-data
Pipeline test 'test1' ran successfully and outputs matched saved
Pipeline test 'test2' ran successfully and outputs matched saved
To run tests over a suite of image configurations in a directory containing a
number of YAML configuration files (i.e. same as building) simply provide the
directory to arcana deploy test
instead of the path to the YAML config
file and supply a directory tree containing the test data, with matching
sub-directory structure to the configuration dir. For example, given the following
directory structure for the configuration files
mri
└── neuro
├── fsl.yml
├── mrtrix3.yml
...
The test data should be laid out like
mri-data
└── neuro
├── fsl
│ └── fast
| ├── test1
| │ ├── inputs
| │ │ └── in_files.nii.gz
| │ ├── outputs
| | │ └── fast
| | │ ├── tissue_class_files.nii.gz
| | │ ├── partial_volumes.nii.gz
| | │ ├── partial-volume-files.nii.gz
| | │ ├── bias-field.nii.gz
| | │ └── probability-map.nii.gz
| │ └── parameters.yml
| └── test2
| ├── inputs
| │ └── in_files.nii.gz
| ├── outputs
| │ └── fast
| │ ├── tissue_class_files.nii.gz
| │ ├── partial_volumes.nii.gz
| │ ├── partial-volume-files.nii.gz
| │ ├── bias-field.nii.gz
| │ └── probability-map.nii.gz
| └── parameters.yml
└── mrtrix3
├── dwi2fod
| ├── test1
| | ├── inputs
...
Like in the case of a single YAML configuration file, the CLI command to test a suite of image/command configurations is.
$ arcana deploy test ./mri ./mri-data --output test-results.json
...E..F..
While not strictly necessary, it is strongly advised to store test data alongside image/command configurations inside some kind of version control. However, storing large files inside vanilla Git repositories is not recommended, therefore, you will probably want to use one of the extensions designed for dealing with large files:
Autodocs#
Documentation can be automatically generated using from the pipeline configuration YAML files (see Building) using
$ arcana deploy docs <path-to-yaml-or-directory> <docs-output-dir>
Generated HTML documents will be placed in the output dir, with pipelines organised hierarchically to match the structure of the source directory.