| T O P I C R E V I E W |
| yaxley |
Posted - Jul 08 2008 : 4:59:46 PM
Recent analyses (after mid-May) have shown some strange output at the second level with the inclusion of NaNs.
First, I've encountered NaNs in the second level report_log as follows:
--
Post-stats
/usr/local/fsl/bin/fslmaths stats/zstat1 -mas mask thresh_zstat1
echo 276871 > thresh_zstat1.vol
zstat1: DLH=nan VOLUME=276871 RESELS=nan
--
Second, in the second level Cluster List, many more clusters are listed than usual (eg, ~90 rather than ~6). However, for each of the listed clusters there are NaNs in the P-value and the Z-COG columns.
Has anyone else stumbled across this?
Rich
|
| 11 L A T E S T R E P L I E S (Newest First) |
| dvsmith |
Posted - Jul 18 2008 : 02:10:32 AM
OK, when I run an analysis using my mask fix, I do not encounter the NaNs. If I let FEAT do the masking, then I have the problem. If use data that was pre-processed before the cluster was updated with Steve Smith's fix, I do not encounter the problem.
I'm not sure how this wouldn't be related to the altered code on the cluster. I suspect everyone using FEAT and GRF for thresholding on the cluster is having this problem. You're not going to notice it unless you look at the activation tables that Feat gives. You should also be able to notice very tiny clusters of activation that normally wouldn't pass GRF.
We can wait a couple of weeks for the new release; however, I'm hoping it won't have any of these issues (i.e., masking problems and/or NaNs). In the interim, we can switch back to my mask fix.
David
|
| yaxley |
Posted - Jul 15 2008 : 1:00:10 PM
[quote]Originally posted by petty
>>is your data being written to goldman?
Yes
>>i know of some other people having the same issue, but the only common link i've seen is data written to that particular server.
I know that Bethany and O'Dhaniel have had the same problem, but I'm not sure which server holds their data.
|
| yaxley |
Posted - Jul 15 2008 : 12:56:50 PM
[quote]Originally posted by dvsmith
>>So no NaNs anywhere before the second-level analyses (e.g., check all the image files from the first level)?
Yes, the NaNs are only present in the second-level output.
>>Are the NaNs only showing up around the edge of the brain?
Yes, the NaNs are limited to the edge of the brain.
|
| dvsmith |
Posted - Jul 14 2008 : 5:00:35 PM
It happens to data being written to Goldman or Fatt. My suspicion is something is wrong with the altered Feat code that Steve Smith supplied to fix the masking problem. Specifically, I think something is going awry when the smoothness and resels are being calculated for thresholding with GRF. I've emailed the FSL listserv to see what they think, but I haven't heard back yet. http://www.jiscmail.ac.uk/cgi-bin/webadmin?A2=ind0807&L=FSL&P=R25769&D=0&I=-3
A potential temporary solution, if this is indeed because of the altered Feat code, might be to use my way of fixing the bad masks since it bypasses mask generation by Feat (http://www.biac.duke.edu/forums/topic.asp?TOPIC_ID=1163).
David
|
| petty |
Posted - Jul 14 2008 : 4:02:53 PM
quote:
Originally posted by dvsmith
And just out of curiosity, absolutely no one else is having this problem in BIAC?
is your data being written to goldman?
i know of some other people having the same issue, but the only common link i've seen is data written to that particular server.
my intuition is that there are some intermittent connectivity issues while writing the output, which is causing the problem.
i've never had an issue other than with a study writing to goldman. |
| dvsmith |
Posted - Jul 12 2008 : 4:32:40 PM
And just out of curiosity, absolutely no one else is having this problem in BIAC? |
| dvsmith |
Posted - Jul 09 2008 : 10:17:31 PM
So no NaNs anywhere before the second-level analyses (e.g., check all the image files from the first level)? Are the NaNs only showing up around the edge of the brain? Bethany had the same problem last week, but I'm not sure if she figured out what was causing the NaNs in the first place. You can remove them using fslmaths, but it would still be nice to know if there's a deeper problem that producing the NaNs.
(I think this is the right command)
fslmaths INPUT -nan OUTPUT |
| yaxley |
Posted - Jul 09 2008 : 2:41:40 PM
No, the two EVs that are in this model are not empty for any run. |
| dvsmith |
Posted - Jul 09 2008 : 2:29:28 PM
Are you trying to use empty-EVs? |
| yaxley |
Posted - Jul 09 2008 : 1:19:54 PM
I just checked the zstat image and there are indeed voxels above the standard threshold of 2.3.
|
| josh.bizzell |
Posted - Jul 09 2008 : 10:03:54 AM
Rich,
For the first example, did you check to see if your zstat1 image has any voxels above the Z-threshold you've set?
And secondly, I've seen NaN values in P-value when the Z-threshold was 1.0 or lower, but I don't think it was in the COG columns.
Josh |
