No improvemetn with periodic or cyclic design

[Sébastien Proulx]

Hello !

I just started to use GLMdenoise on my data. I have a cyclic design, with multiple 2-min runs where large visual stimulus are presented for 6-sec ON and 6-sec OFF repeatedly. I am using spin-echo with 13 coronal slices at 1x1x1mm, positioned perpendicular to the calcalrine sulcus, at the end of the occipital pole. Running GLMdenoise with default parameters except that HRF estimation is turned off ('assume') yield no improvement with the 1 PC that the algorithm chose to use. Median Rsquared in PCselection.png is ~0.08 at 0 and 1 PC, and then decrease with more PC, down to 0.04 at 20 PCs, and in the scatter plot comparing 0 PC to 1 PC (PCscatter01.png), all data points fall on the line. Could that be due to my particular design? Any fix? Any other idea on what could possibly be wrong?

Thanks a lot for any help!

Sébastien

[Kendrick Kay]

Hi Sebastien,

I'm not sure anything is "wrong".  Whether denoising is possible for a given dataset / experiment depends on a lot of factors (including the MR pulse sequence you chose, the number of trials / experimental design, etc.).

If you can show us what your design matrix looks like, that could be insightful.  Also, if you have a picture of one of your raw MR volumes, that might be helpful.  It might be that you are dominated by thermal noise, which is a type of noise that GLMdenoise can't really help you with.  Also, block designs with many trials tend to have high estimation power, so it is less likely that denoising will substantially change your estimates.

Kendrick

--
Kendrick Kay, PhD
Assistant Professor
Center for Magnetic Resonance Research (2-116)
University of Minnesota, Twin Cities
   Web: http://cvnlab.net
E-mail: k...@umn.edu
  Cell: 510-206-1059
 Skype: kendrickkay

--
You received this message because you are subscribed to the Google Groups "GLMdenoise" group.
To unsubscribe from this group and stop receiving emails from it, send an email to glmdenoise+...@googlegroups.com.
To post to this group, send email to glmde...@googlegroups.com.
Visit this group at https://groups.google.com/group/glmdenoise.
For more options, visit https://groups.google.com/d/optout.

[Sébastien Proulx]

Hello Kendrick,

Thanks for your input. It seems like it is indeed due to the already high estimation power. Now that I ran GLMdenoise on many more subjects and that I look at SNRamountofdatagained.png (see attachment), I see that although most voxels are unaffected, voxels with very small SNR to start with do show some benefit.

For the sake of completness, I also attached my raw volume and my design.

Thanks for your help and for your awesome method!

Sébastien

[Kendrick Kay]

Hi Sebastien,

Your volume and design matrix look reasonable.  Good luck with your experiment!

Kendrick

<SNRamountofdatagained.png><MeanVolume.png><design.png>