Localization of cerebral regions involved in the coding of a stimulus category ?

[JeanLucAnton]

How can we obtain accurate information on the localization of cerebral regions involved in the coding of a stimulus category ?


More precisely, we have tested 2 methods (for each subject, normalized in a same space): ∞ Searchlight : for each voxel, we run an MVPA analysis in a sphere that is centered on the voxel. We thus obtain an accuracy map for each voxel tested. We can then test the value of this accuracy across all the subjects of our population. ∞ MVPA analysis on the whole brain. For instance, in nilearn : support vector classifier - svc, with penalty = tv-l1, or SpaceNetClassifier with penalty=graph-net). Here we obtain a weight map for each subject. We can then test the value of these weights accross all the subjects of our population.

Which is the most rigorous method ? What biases do they each bring ? What differences in interpretations do we get from one approach or the other ? 


[JeanLucAnton]

RESPONSE OF JO ETZEL  Neither of these is particularly great; I actually prefer ROI-based analyses: carry out the analyses in predefined (independent) regions, either chosen from the hypotheses (e.g., visual or motor regions) or from a full brain parcellation (e.g., Schaefer 2018) https://github.com/ThomasYeoLab/CBIG/tree/master/stable_projects/brain_parcellation/Schaefer2018_LocalGlobal).
 Very briefly, searchlights can identify focal areas well, but get hard to interpret when areas vary across people or are not focal, e.g. my paper https://www.ncbi.nlm.nih.gov/pubmed/23558106 (and associated blog posts http://mvpa.blogspot.com/search/label/SA%3APPP ).
 Weight maps can also be incredibly hard to interpret; perhaps even more so than searchlights. I listed some references in a reply on a previous thread NILEARN: Are SVM weights can be useful for group analysis ?

[JeanLucAnton]

Weight maps : Sylvain thinks it is not appropriate : when you fit a classifier if the algo finds 2 features (eg. 2 voxels) that do exactly the same : it will give a weight in the best voxel and will undermine the other. More exactly, it could very well give a non-null weight to one of the two voxels, and a null weight to the other one, in an arbitrary manner...



Confusion matrices : you can get them from each voxel, from each ROI and even look at them after your searchlight. Do whatever you want (as always). You can then do the stats like you would do with neurons : compare the accuracies for one condition/decoding (one box of your confusion matrix) to another. 

Parcellisation method : you run your MVPA on a specific parcellation pattern of your brain and you do that 1000 times. You get a weight map for each parcellation. Then you average them all and get a global map of weights that is finer spatially : you get blobs where the weights are higher. The parcellations are chosen randomly : you take the full brain without any a priori.  Here is the paper : 
https://doi.org/10.1016/j.neuroimage.2017.10.005
 The Thirion team is aiming at including this FREM method in nilearn... Should be available in a few months if everything goes as planned...


[JeanLucAnton]

Difference in the sizes of the chosen ROI ?

Sylvain: If you want to compare two regions in terms of decoding, you must have the same number of voxels.