Hello
I am using the MVPA-Light toolbox for fmri searchlight on a 3D volume unsmoothed, therefore there are some voxels whose raw value is NaN
When applying searchlight to it, it returns some dvals that are NaN, but then the part that calculates the AUC transforms them into performances at 0 ; also when a NaN voxel is present within a searchlight sphere it gives weird performances like 0.3 or so
I think i should apply a mask beforehand (for each subject) but i worry this would screw up the 3D structure that is necessary for the searchlight
So far i use a 'manual' mask just before doing the cluster permutation correction that just removes the values at 0 (i made sure they're all non-brain and near or at the border of the field of view) and it seems to be working okay but is not very rigourous methodologically
More generally, how do you deal with searchlight in the case of non-brain / NaN-voxels in the volume? especially for the voxels at the limit brain / non brain ? we would expect the voxels at the center of a searchlight including non brain voxels to be 'polluted' and have an underestimated performance right?
At what step should the mask be added and how to preserve the 3D structure ?
Thanks in advance, sincerely
An thank you for the very nice toolbox and easy-to-use functions
Julie Boyer, PhD student, Paris
Hello
I am using the MVPA-Light toolbox for fmri searchlight on a 3D volume unsmoothed, therefore there are some voxels whose raw value is NaN
When applying searchlight to it, it returns some dvals that are NaN, but then the part that calculates the AUC transforms them into performances at 0 ; also when a NaN voxel is present within a searchlight sphere it gives weird performances like 0.3 or so
I think i should apply a mask beforehand (for each subject) but i worry this would screw up the 3D structure that is necessary for the searchlight
So far i use a 'manual' mask just before doing the cluster permutation correction that just removes the values at 0 (i made sure they're all non-brain and near or at the border of the field of view) and it seems to be working okay but is not very rigourous methodologically
More generally, how do you deal with searchlight in the case of non-brain / NaN-voxels in the volume? especially for the voxels at the limit brain / non brain ? we would expect the voxels at the center of a searchlight including non brain voxels to be 'polluted' and have an underestimated performance right?
At what step should the mask be added and how to preserve the 3D structure ?
Thanks in advance, sincerely
An thank you for the very nice toolbox and easy-to-use functions
Julie Boyer, PhD student, Paris