“Bare-metal” installation (Python 2/3)¶
mriqc follows the [BIDSApps] specification, the execution is
split in two consecutive steps: a first level (or
by a second level (or
participant level, all individual images to be processed are run
through the pipeline, and the MRIQC measures are extracted and
the individual reports (see The MRIQC Reports) generated.
group level, the IQMs extracted in
first place are combined in a table and the group reports are generated.
The first (
participant) level is executed as follows:
mriqc bids-dataset/ out/ participant
Please note the keyword
participant as fourth positional argument.
It is possible to run
mriqc on specific subjects using
mriqc bids-dataset/ out/ participant --participant_label S001 S002
S002 are subject identifiers, corresponding to the folders
sub-S002 in the BIDS tree.
Here, it is also accepted to use the
mriqc bids-dataset/ out/ participant --participant_label sub-S001 sub-S002
If the argument
--participant_label is not provided, then all
subjects will be processed and the group level analysis will
automatically be executed without need of running the command in item 3.
After running the
participant level with the
group level will not be automatically triggered.
To run the group level analysis:
mriqc bids-dataset/ out/ group
Examples of the generated visual reports are found in mriqc.org.
Depending on the input images, the resulting outputs will vary as described next.
If you have Docker installed, the quickest way to get
mriqc to work
is following the running with docker guide.
Running MRIQC on HPC clusters¶
We have profiled cores and memory usages with the resource profiler tool of nipype.
An MRIQC run of one subject (from the ABIDE) dataset, containing only one run, one BOLD task (resting-state) yielded the following report:
MultiProcplugin of nipype with
nprocs=10, the workflow nodes run across the available processors for 41.68 minutes. A memory peak of 8GB is reached by the end of the runtime, when the plotting nodes are fired up.
We also profiled MRIQC on a dataset with 8 tasks (one run per task), on ds030 of OpenfMRI:
Again, we used
n_procs=10. The software run for roughly about the same time (47.11 min). Most of the run time, memory usage keeps around a maximum of 10GB. Since we saw a memory consumption of 1-2GB during the the 1-task example, a rule of thumb may be that each task takes around 1GB of memory.