[BUG] incorrect masking of null array in stats.compare_images

[alicehodapp]

Issue summary

Hi!

It seems like the neuromaps.stats.comapre_images function uses the zero/nan mask that is based on the two input arrays on permutated data with different locations of zeros/nans.

Detailed issue description

Thank you very much for developing this super useful toolbox!

I used the neuromaps.alexander_bloch() function to create rotated data. I have a few nan values in my vertex vise data, so I wanted to see if and how neuromaps.stats.compare_images() is handling such a case. I might be missing something obvious, but it seems like the mask that is created based on zeros/nans in src and trg is then applied to the null distributions (if provided). At least when using the alexander_bloch() rotated output, the nans are also rotated and therefore the mask doesn't fit anymore.

I started fixing this and realized that a fix needs adjustments in multiple locations in the code. In case you want to adapt the masking approach in compare_images like I did (rather than e.g., changes in the alexander_bloch() function), I'd be happy to make a pull request.

This is the section of code in neuromaps.stats.compare_images() I am referring to:

def compare_images():
...
# drop NaNs (if nan_policy==`omit`) and zeros (if ignore_zero=True)
zeromask = np.zeros(len(srcdata), dtype=bool)
if ignore_zero:
    zeromask = np.logical_or(np.isclose(srcdata, 0),
                                np.isclose(trgdata, 0))
nanmask = np.logical_or(np.isnan(srcdata), np.isnan(trgdata))
if nan_policy == 'raise':
    if np.any(nanmask):
        raise ValueError('Inputs contain nan')
elif nan_policy == 'omit':
    mask = np.logical_and(np.logical_not(zeromask),
                            np.logical_not(nanmask))
elif nan_policy == 'propagate':
    mask = np.logical_not(zeromask)
srcdata, trgdata = srcdata[mask], trgdata[mask]

if metric in methods:
    if metric == 'spearmanr':
        srcdata = sstats.rankdata(srcdata)
        trgdata = sstats.rankdata(trgdata)
    metric = partial(efficient_pearsonr, return_pval=False)

if nulls is not None:
    n_perm = nulls.shape[-1]
    nulls = nulls[mask]
...

Steps to reproduce issue

I came across this bug in the context of running:


from neuromaps import stats, nulls

rotated = nulls.alexander_bloch(src, atlas='fsaverage', density='10k',
                                n_perm=100, seed=1234)
corr, pval = stats.compare_images(src, trg, nulls=rotated, ignore_zero=False, nan_policy='omit', metric="pearsonr")

Software version

3.9.6 (default, Nov 11 2024, 03:15:38)
[Clang 16.0.0 (clang-1600.0.26.6)]
0.0.5

Code of Conduct

• I agree to follow the neuromaps Code of Conduct