Import NiLearn HRF convolution code directly

[adelavega]

Replace our custom copy of Nilearn's HRF code

[adelavega]

Let's compare NiLearn and our HRF code to identify potential differences, and if looks safe to import directly, try that and see if tests pass

[jmumford]

Notebooks and code in this repo summarize my findings in comparing the nilearn functions to pybids.

Briefly

• I feel like there aren't any reasons why we cannot use the nilearn code directly, but it would be good if a fresh set of eyes looked over the side-by-side code comparison.
• If we don't start using nilearn directly, we need to at least update the HRF code in pybids to repair some bugs.
• This does not fix the issue identified with the FIR model by Michael. Within the same repo above, I've recreated the problem in this notebook. I believe the issue is created by DenseRunVariable after hrf.compute_regressor is run. Below is image of an FIR-based regressor just after the hrf.compute_regresor function is run in the Convolve class vs the regressor that is created in X within the output from node.run()

[adelavega]

Looked into it and there are two distinct issues:

1. The aliasing seen here is actually introduced when resampling DenseRunVariable back down to TR. Downsampling is still subject to aliasing, which is particularly noticeable with box car like functions (of which FIR is one). Sparse variables get oversampled to get convolved and then have aliasing when downsampled to TR.

2. FIR delays are being computed in the oversampled space, not the original one. Did NiLearn fix this issue?