I’m reaching out to see if MRIQC users have any tips for excluding task-based BOLD runs and/or participants due to their BOLD time series. I know this question has come up a couple times (Excluding subjects based on mriqc, Excluding outliers with fmriprep output or mriqc), but I figured I’d pose the question again in case anyone has new insight.
I’ll probably delve into the annotations at the PCP-QAP site, but since these are pilot subjects, I’m also trying to get an idea if any part of the protocol or scan parameters need to be changed in order to optimize signal.
Unfortunately, excluding datasets because of quality remains a very artesanal task, especially for BOLD fMRI data. There are some good practices that may help you navigate through the task:
Establish some exclusion criteria before you start (will come back to your exclusion criteria later)
Screen all your subjects/tasks/runs with the same settings along the project (e.g., if you chose MRIQC, then keep the same version of MRIQC for the whole sample). I highly recommend using the latest MRIQC (v0.15.0).
Randomize, if possible, the order of screening.
Average assessments, if you can do several of them.
Look at the group distribution of IQMs (group report of MRIQC) to spot outliers.
Visualize BOLD series in movie mode before any exclusion decision.
Obvious artifacts (spikes, unrecoverable dropout in temporal and vmPFC regions, interference artifacts, reconstruction issues such as eye leakage through slices in multiband sequences, missing data, etc)
Subjects above a preset threshold of recurrent motion. E.g. more than 20% of the timepoints with FD > 0.5mm.
People are sharing their ratings with our MRIQC-WebAPI (https://doi.org/10.1038/s41597-019-0035-4). That could be a way of learning what others are doing (cross-checking your ratings with public datasets already on the repository).
The MRIQC WebAPI notebook on github was also very useful as I delved into the WebAPI. (Posting for my own reference and anyone else who may be getting started.)
