[BUG] NiftiLabelsMasker missing signal

[andreifoldes]

Is there an existing issue for this?

• I have searched the existing issues

Operating system

• Linux
• Mac
• Windows

Operating system version

• Windows 10

Python version

• 3.12
• 3.11
• 3.10
• 3.9
• 3.8

nilearn version

0.10.2

Expected behavior

I have a mask with certified 8 non-zero values and would expect 8 numbers from NiLabelsMasker, that is indeed the case for all but one of my participants:

 right_subfield_mask = f"xxx/hippunfold/sub-{subject_label}/anat/sub-{subject_label}_hemi-R_space-T1w_desc-subfields_atlas-bigbrain_dseg.nii.gz"
    
    tmp = nib.load(right_subfield_mask)
    tmp = tmp.get_fdata()
    # get unique values nonzero values
    unique_values = np.unique(tmp[tmp != 0])
    print(unique_values)
    # print unique values    
    if len(unique_values) != 8:
        # raise an actual warning using warnings.warn
        # print unique values
        print(unique_values)
        warnings.warn("WARNING: unique values in right_subfield_mask is not equal to 8")
    
    right_subfield_labels = [ label + "_right" for label in subfield_labels]
    right_subfield_labels = [right_subfield_labels[i-1] for i in unique_values.astype(int)]

    # Instantiate the masker with label image and label values
    masker = NiftiLabelsMasker(
        right_subfield_mask,
        labels=right_subfield_labels,
        standardize="zscore_sample",
    )

    right_subfield_signals = masker.fit_transform(betamap)
    print(right_subfield_signals)

Current behavior & error messages

This is what I got:

8 non-zero values go in, 7 values come out. No warning or error message...

>> right_subfield_mask = f"/xxx/hippunfold_v2/hippunfold/sub-{subject_label}/anat/sub-{subject_label}_hemi-R_space-T1w_desc-subfields_atlas-bigbrain_dseg.nii.gz"
>>> tmp = nib.load(right_subfield_mask)
p.get_fdata()
# get unique values nonzero values

>>> tmp = tmp.get_fdata()
unique_values = np.unique(tmp[tmp != 0])
print(unique_values)
# print unique values

if len(unique_values) != 8:
    # raise an actual warning using warnings.warn
    # print unique values
    print(unique_values)
    warnings.warn("WARNING: unique values in right_subfield_mask is not equal to 8")

right_subfield_labels = [ label + "_right" for label in subfield_labels]
right_subfield_labels = [right_subfield_labels[i-1] for i in unique_values.astype(int)]
# Instantiate the masker with label image and label values

masker = NiftiLabelsMasker(
    right_subfield_mask,
    labels=right_subfield_labels,
    standardize="zscore_sample",
)

right_subfield_signals = masker.fit_transform(betamap)
print(right_subfield_signals)
>>> # get unique values nonzero values
>>>
>>> unique_values = np.unique(tmp[tmp != 0])
>>> print(unique_values)
**[1. 2. 3. 4. 5. 6. 7. 8.]**
>>> # print unique values
>>>
>>> if len(unique_values) != 8:
...     # raise an actual warning using warnings.warn
...     # print unique values
...     print(unique_values)
...     warnings.warn("WARNING: unique values in right_subfield_mask is not equal to 8")
...
>>> right_subfield_labels = [ label + "_right" for label in subfield_labels]
>>> right_subfield_labels = [right_subfield_labels[i-1] for i in unique_values.astype(int)]
>>> # Instantiate the masker with label image and label values
>>>
>>> masker = NiftiLabelsMasker(
...     right_subfield_mask,
...     labels=right_subfield_labels,
...     standardize="zscore_sample",
... )
>>>
>>> right_subfield_signals = masker.fit_transform(betamap)
>>> print(right_subfield_signals)
**[[-0.52435863 -0.40006825 -0.34511584 -0.5896911  -0.44941693 -0.23059142
  -0.57147044]]**

Steps and code to reproduce bug

No response

[bthirion]

Thx for reporting.
Can you provide the mask and betamap in some way ?

[ymzayek]

@andreifoldes I've seen this happen before due to resampling. The labels masker by default resamples the labels_img to the data and this could result in small regions getting wiped out. See resampling_target paramter of NiftiLabelsMasker. We could add a warning when this happens. WDYT @bthirion ?

[bthirion]

Isn't it addressed by #3761

[ymzayek]

No explicitly but yes. These new returns will help the user map the output back to the input

[andreifoldes]

Yes, nice catch, that's probably it! A warning raised would be nice!

[andreifoldes]

Is sending the files still necessary?

Actually upon second thought maybe its something else because but files are in the native space so resampling shouldn't be doing much of anything :)

[bthirion]

If you still don't understand what's going on, it's a goo idea to share the files. Best,

[ymzayek]

@andreifoldes actually a warning should be raised, see

nilearn/nilearn/maskers/nifti_labels_masker.py

Lines 562 to 567 in e4dc5ad

	warnings.warn("After resampling the label image to the "
	"data image, the following labels were "
	f"removed: {labels_diff}. "
	"Label image only contains "
	f"{len(labels_after_resampling)} labels "
	"(including background).")

Can you double check if that is raised?

If that's not the case then sending the files or providing a minimal reproducible example to raise the same error might be necessary

[andreifoldes]

Yes, thank you, will check this week and also send data!