Dear Thierry,

Here SPM has a Weighted Graph Laplacian (WGL) spatial prior over each regression coefficient with a different inferred level of smoothness for each. I see from the images that beta1 is also much smoother than beta3. Are there very different numbers of events in columns 1 versus 3 ? 

You've also used the subvolumes option - so that the algorithm is applied separately to each subvolume, as described in

http://www.fil.ion.ucl.ac.uk/~wpenny/publications/graph08.pdf

So the combination of being very smooth and having subvolumes gives you these large edge effects. I don't think there's an error in processing however - this is what the algorithm does. The only thing that concerns me are the large values in the SD_beta1 images - am not entirely sure of the origin of this. It could be that the SD is just larger at the edges because there are fewer voxels to base estimates on (in combination with a small number of events ?). Whatever the origin, this will introduce noise into your MVPA (i know the SD's don't enter directly, but its a sign that betas have not been estimated well).

So perhaps its best to not choose the subvolumes/WGL combination here and just go with unweighted (U)GL (WGL would take a very long time on the whole brain as a single volume).

Best,

Will.

---

From: SPM (Statistical Parametric Mapping) <[log in to unmask]> on behalf of Thierry Chaminade <[log in to unmask]>
Sent: 26 April 2017 09:48
To: [log in to unmask]
Subject: [SPM] fMRI First level Bayesian estimation: some weird CBeta maps

 

Dear SPMers,

For reasons explained in reference below, I am using Bayesian first level estimation of an event-related fMRI experiment (with AR(6)). 

I would like to understand why the CBeta images for the first trial (the data is processed for MVPA analysis with 1 column per trial) are strikingly different from the other trials (cf image hereunder). In particular it seems that the limits of the subvolumes used for the estimate are clearly visible in CBeta_0001, but invisible from CBeta_0003 onward. This is true for all of the 6 acquisition sessions.

My questions for Bayesian experts are 1) is it normal?, 2) Can (should) it be corrected? and 3) Is it correct to continue an MVPA analysis with such images?

Thanks in advance for any help you could bring me on this issue!
Best regards, Thierry C.

CBeta_0001

CBeta_0003


For information, SDBeta_0001



Reference:

Gilron, R., Rosenblatt, J.D., Mukamel, R., 2016. Addressing the ‘problem’ of Temporal Correlations in MVPA Analysis. Proceeding of the In: Proceedings of the 6th International Workshop on Pattern Recognition in Neuroimaging.