Description. We found that we were unable to run RISH, because brain extraction of (some) matched participants was poor (?). Here, we review the results of brain extraction (via FSL’s BET) for matched participants.

Participants. In total, we have 124 participants in the matched sample, including n=44 unique participants from the reference site (CMH) and n=80 participants from the target sites (CMP, MRC, MRP, ZHH, ZHP); i.e., 16 participants per target site.

Summary of BET quality in RISH sample. We QC’d the static BET images, available in archive/data/SPINS/pipelines/bids_apps/dmriprep/QC, for all participants included in the matched RISH sample. The following table indicates how many participants failed, passed, were missing outputs (are being re-run), or didn’t have outputs by design (human phantoms; also being re-run). ZHH was the only site without issue. The PRISMA sites had pronounced problems with BET: no participants from CMP nor ZHP passed, and only 2 MRP participants passed. (Note that all participants included in the RISH sample passed visual QC of raw diffusion images.)

site pass fail missing (error) missing (phantom)
CMH 31 4 0 9
CMP 0 7 6 3
MRC 8 5 0 3
MRP 2 10 1 3
ZHH 13 0 0 3
ZHP 0 6 7 3

Reference images. A good extraction will show the red line tracing the boundary between brain and skull, and encompass only brain (no skull or black/void space). A bad extraction is anything that violates this definition.

Example good

Example bad


Example of BET failures per site. Here, we show two examples of failed extraction per site. There were a diversity of failed extractions; some were clearly off across multiple slices, others might better be described as less than ideal in a circumscribed region.

CMH

CMP

MRC

MRP

ZHP

Review of T1 tissue segmentation. We also checked to see if the n=32 subset of participants with failed/problematic BET extractions had problematic T1 tissue segmentations, derived from fmriprep. Some did, but most did not. The general pattern seemed to be that those with especially poor extractions also had poor tissue segmentation. For example, the segmentation for the ‘bad’ example above (ZHP0108) looks as follows:


Review of derived masks. We also checked the appearance of T1 masks for the subset of participants with failed/problematic BET extractions and problematic T1 tissue segmentations. We found that the masks had ‘holes’ through the volumes. For example, the segmentation for the ‘bad’ example above (ZHP0108) looks as follows:


Ajustment to brain extraction – using BET and AFNI. Michael has recently tested a new method of brain extraction that also uses AFNI. He created an initial mask from BET (more liberal), and then created another mask using AFNI, and finally created an intersection: agreement was saved as a new mask. This appears to result in better extractions (ZHP108 shown):


We also see that tissue segmentation on the T1 is much improved. Note, however, that this data has one other difference: it has also been run with a newer version of fMRIprep, i.e. version 0.4.1. So, the improved extraction may be a result of new fMRIprep or the combined BET and AFNI method.


Current hypothesis. We think that if BET looks really bad (e.g., ZHP0108), something has likely gone wrong with the distortion correction. If the extraction is more subtley poor, perhaps something else has gone wrong (could be any number of things?).

Next steps / discussion. Figure out why extraction is so bad at the PRISMA sites. Is it distortion correction? How can we individually fix? Michael and his group have delinerated a timeline for dmriprep. Michael has also learned about another brain extraction method, QSIPREP, which apparently performs well even on bad b0s, and is working on implementing it into a nipype workflow. In the interim, can we test RISH with ZHH?