view_img always masks bg_img with MNI mask

[dbrennan44]

when plotting with view_img, it appears the bg image is being "masked" by a different brain (perhaps MNI?)

problem persists even when using same image as bg_image and plotted img

[bthirion]

Thx for reporting. Can you provide a snippet of code (and possibly an image) to reproduce this ? Best,

[dbrennan44]

malf_Labels.nii.gz

import nilearn as ni
from nilearn import image, plotting

malf=image.load_img('/Users/danielbrennan/Downloads/c001/malf_Labels.nii.gz')
plotting.view_img(malf,bg_img=malf,cmap="cold_hot",threshold="99%",black_bg=False)

[dbrennan44]

from messing around, it seems to be some kind of center-of-mass-like calculation which offsets the viewpoint. removing the threshold="99%" fixes the issue. But if I threshold outside of the plotting function, I get the same result

[bthirion]

Looking quickly at it, it seems to amount to the fact that bg_img is handled differently than stat_map_img: in particular it is not resampled. When these images are not in MNI space this indeed leads to spurious plots.
The point is that in most cases we have dealt with so far, the bg_img would be an MNI template, or at least in MNI space.

[dbrennan44]

right, I'll see if I can change that behavior; I do most of my analysis in subject-space. Seems strange to allow supplying of bg_img but resample automatically into MNI.

Should also be emphasized again that this behavior goes away (i.e. plots correctly) when omitting the threshold argument

[NicolasGensollen]

@dbrennan44 view_img resamples the stat map to the background image automatically as can be seen here:

nilearn/nilearn/plotting/html_stat_map.py

Lines 535 to 536 in 78b0986

	stat_map_img, mask_img = _resample_stat_map(stat_map_img, bg_img, mask_img,
	resampling_interpolation)

In addition, if you don't explicitely set bg_img=False in view_img, then the MNI152 template is used and the stat map is therefore resampled to the template.

nilearn/nilearn/plotting/html_stat_map.py

Lines 444 to 447 in 78b0986

	%(bg_img)s
	If nothing is specified, the MNI152 template will be used.
	To turn off background image, just pass "bg_img=False".
	Default='MNI152'.

By setting bg_img to False, or to the stat map like you did, view_img shouldn't resample the stat map.

I'm not sure to understand why the threshold would change this though. It is only used to select the cut slices here:

nilearn/nilearn/plotting/html_stat_map.py

Line 537 in 78b0986

cut_slices = _get_cut_slices(stat_map_img, cut_coords, threshold)

[jeromedockes]

• the bug is here; when preparing the "brain sprite" for the background a mask is computed (not sure why the background needs to be masked, in fact I don't think it should). the problem is that this is done with compute_brain_mask which just thresholds the mni152 template after resampling it to its first argument's resolution -- this doesn't work if the image isn't in mni space
• the problem also happens when there is no thresholding, but in that case we just don't see it because the background map is completely hidden behind the foreground map
• it is not a resampling issue, if the passed bg_img and stat map (as is the case here) have the same affine and shape the stat map isn't modified
• setting bg_img=None (or False) doesn't really solve the problem because (i) we do want the background image to be plotted and (ii) the masking of the background still happens so we see the outline of the mni template:

[jeromedockes]

During core dev meeting yesterday it was decided that:

• the default bg_img (MNI template) would continue to be masked, ie background is transparent outside of the brain
• any other background image will not be masked (so black_bg parameter may not be very useful in that case)