TDT: ROI mask used for the whole brain searchlight analysis

Hi everyone,

I have a question on what kind of mask images should be set when performing the whole brain searchlight analysis.

Following “decoding_template.m”, I’m using a “mask.nii” as a mask image for each participant (cfg.files.mask), which was generated when performing SPM 1st-level specify. I’ve performed the whole brain searchlight RSA using those mask images. But in some results, I observed “leaky” signals (i.e., signals observed outside the cortex) in the group-level statistic as shown in the attached picture. I guess that this would be due to the results of my preprocessing (I’m wondering to try segmentation or change some normalization parameters. If there are any recommended methods for better segmentation, please let me know). Or, I’m also wondering whether I should use different images instead of mask.nii as ROI masks. I would be very happy if I could get some tips on this topic.

TDT is really helpful for my research. Thanks a lot!


I generally suggest against restricting searchlight analysis to a tight mask, both for the logic of searchlight analysis (e.g., to minimize edge effects) and to ease QC. If the searchlight analysis is detecting something sensible it should report more signal (accuracy, similarity, whatever) in grey matter than white, ventricles, and outside the brain. A bit of blurring around the edges is expected (e.g., because of imperfect alignment, searchlights partly in and partly out of the brain), but there should be an obvious difference between grey matter and everything else.

I suggest checking the individual results images, not just the group, and also the SPM GLM results, or whatever you ran the searchlight on. If everything looks good (e.g., successful normalization, more BOLD in grey matter) except for the searchlight results, there’s likely a bug in the searchlight analysis setup, or the effect you were hoping for isn’t present/detectable with this analysis.

Hi Jo,

I really appreciate your helpful comments on the searchlight analysis! I clearly understand what I have to do. First I will check individual results images as well as SPM GLM results carefully. I also consider using a tight mask.

Thanks a lot,