Support TFCE thresholding in CBMA algorithms

[tsalo]

Closes #145.

Changes proposed in this pull request:

• Add a tfce option to CBMAEstimator.correct_fwe_montecarlo that calculates TFCE-based FWE corrected maps.
• Add a tfce option to MKDAChi2.correct_fwe_montecarlo as well.
• Add a function, nimare.meta.utils.calculate_tfce, which calculates TFCE on a 3D numpy array.

[codecov]

Codecov Report

Merging #655 (cb7fcfc) into main (6cd751e) will increase coverage by 0.24%.
The diff coverage is 100.00%.

@@            Coverage Diff             @@
##             main     #655      +/-   ##
==========================================
+ Coverage   85.28%   85.53%   +0.24%     
==========================================
  Files          40       40              
  Lines        4513     4589      +76     
==========================================
+ Hits         3849     3925      +76     
  Misses        664      664              

Impacted Files	Coverage Δ
nimare/meta/cbma/ale.py	94.08% <ø> (ø)
nimare/meta/cbma/base.py	96.80% <100.00%> (+0.27%)	⬆️
nimare/meta/cbma/mkda.py	97.32% <100.00%> (+0.29%)	⬆️
nimare/meta/utils.py	57.38% <100.00%> (+4.58%)	⬆️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 6cd751e...cb7fcfc. Read the comment docs.

[tsalo]

Actually not going to touch IBMA any time soon.

[tsalo]

One question I have is whether the maximum statistic null distributions should be separated by sign for two-sided tests. For example, for the MKDAChi2, which can have positive and negative chi-squared values, should we retain maximum negative cluster size and maximum positive cluster size distributions separately?

I'm thinking about this primarily because of what I read in Albajes-Eizagirre et al. (2019):

SDM-PSI can currently save four different maximum statistics from the image of z-values: the largest z-value (i.e., voxel-based statistics), the maximum cluster size or mass after thresholding with a used-defined z- value (i.e., cluster-size or mass statistics) (Bullmore et al., 1999), and the maximum TFCE (Smith and Nichols, 2009). We have implemented TFCE in our software so that users who do not have FSL installed will still be able to use TFCE in SDM-PSI.

This procedure theoretically only tests hypothesis in one direction (e.g., patients > controls), but not the hypothesis in the other direction (e.g., patients < controls), and thus, the procedure should be conducted twice, one for each direction. However, given that the hypotheses are complementary and a permutation test is computationally consuming, SDM-PSI saves two numbers from each permutation: one for the positive hypothesis (e.g., the highest z-value) and one for the negative hypothesis (e.g., the lowest z-value), in two separate null distributions.

[tsalo]

This test now takes ~90 seconds locally. 😬

[jdkent]

relevant publication from Simons group: https://pubmed.ncbi.nlm.nih.gov/35535616/

[tsalo]

Thanks! I knew Simon was working on TFCE in ALE, but I didn't realize there was a paper out already. Based on the abstract, it may not be worth it to merge this PR, but on the other hand I do want NiMARE to be useable for that kind of study in the future...

[tsalo]

I'm thinking that maybe I should close this PR, but also open a separate one that adds a note to the documentation about TFCE and points to this one. That way, if someone does want to use TFCE in NiMARE, they can refer to code that (at least at one point) worked, and they'll also see why we chose not to integrate it into the package.

I am also hopeful that we will be able to implement a much faster probabilistic TFCE in the future (like MNE), and that approach might not have the same issues noted in Simon's paper.