1. Quality Assessment

Of the 2341 raw images, 1581 had previously been viewed to assess quality as part of the PNC or GRMPY. To determine which of the remaining 760 images required viewing, quality metrics based on the surface reconstruction from Freesurfer's cross-sectional pipeline were calculated. FreeQC (v 0.0.9) was used to calculate CNR, Euler's number, and the number of holes in the cortical surface. The image was pushed to DockerHub (pennbbl/freeqc:0.0.9), then pulled via Singularity to PMACS for running. It was confirmed that Euler's number and the number of holes in the cortical surface are correlated -1. Histograms for each quality metric were generated.

More plots were generated to determine which of the remaining 760 images needed viewing. The first plot was a scatterplot of CNR versus the total number of holes, colored by manual exclusion status in the 1581 sample. The second plot was of ROC curves for CNR in the left hemisphere (gray versus CSF) and the total number of holes (see below). The AUC for the number of holes was very high (0.9158), while the AUC for CNR was low (0.6057), so only the number of holes was used to determine which images should be viewed. The cut-off was chosen such that sensitivity would be very high: we prioritized finding bad images over saving time (sensitivity = 0.97, specificity = 0.60). After applying the cut-off (120 or more holes), 292 images required manual review.

Of these images, 30 were very poor quality (rating = 0), 82 were okay (rating = 1), and 180 were good (rating = 2). Very poor image quality was characterized by ghosting, ringing throughout the cortex, obvious motion, or field of view problems. The code can be found here.

Noticing that some of the original manual ratings failed to exclude very bad images (e.g., sub-100278 ses-PNC3, which was originally rated 1.667), images with more than 250 holes that had previously been viewed were viewed again (N = 49). Of these images, 23 were very poor quality (rating = 0), 14 were okay (rating = 1), and 12 were good (rating = 2). The code can be found here.

Scans that should be included in their respective single subject templates (antssst) were then identified. Any scan with a manual rating of 0 was excluded (N = 66), as well as any scan that was the only image remaining for a subject after manual rating exclusions were applied (N = 19). The code can be found here.

Notes

• (October 23, 2020): Numbers will need to be updated slightly when the three images that did not make it through fMRIPrep are processed.
• (November 9, 2020): sub-100278 ses-PNC3 and sub-117595 ses-PNC1 have made it through fMRIPrep now, but sub-98585 ses-PNC1 keeps hanging. Since sub-98585 ses-PNC1's t1w image is clearly of poor quality, I will proceed without it.