Jérôme Dockès

> Is there a reason the `maskers.labels_` was kept the way it is? I think its main problem is its name. it gives, for each output dimension, the index of...

> @jeromedockes could you explain your idea for PR2 a bit more? ATM if I want to give names to the atlas regions I must pass an array (`labels`) where...

thanks @mtorabi59 ! I agree with both points. for the second one I think it will still be useful to have a mapping from integer values contained in the atlas...

As @mtorabi59 and I just discussed orally: There are 3 important attributes for each atlas region: - the (optional) region human-readable name, such as "Parietal lobe" - the region id...

> 3\. To solve the background problem, I suggest that we fill the background and empty voxels in `region_maps` with `np.nan` instead of `0`, which makes more sense, so that...

> Also @mtorabi59 after looking at the API more closely, I think we can already do what you suggest in the new function with I'm not sure that is the...

> @bthirion I think the argument for the second option is: why let users give a ROI mask (the label or maps img) and then another mask on top of...

In any case I think the step of storing the final (masked and resampled) atlas in an image and inspecting it cannot be avoided, because masking the atlas could result...

FWIW the `mask_img` parameter is used in none of the 12 examples that use either `NiftiLabelsMasker` or `NiftiMapsMasker` in the gallery

> So `img_to_signals_labels` is not aware of them to include them in `region_signals` right -- what I meant is that some regions from the original atlas provided by the user...