Add ability to re-run ICA when no/few BOLD components are found

[tsalo]

Summary

When no BOLD components are detected by the decision tree, it is possible that the ICA simply failed to produce components that adequately dissociated BOLD and non-BOLD signal sources by chance. Changing the random seed can help with this. It would be awesome to be able to feed information from the decision tree back into the decomposition (combining with the maxit and maxrestart parameters) to restart the decomposition with an updated seed.

This would require a substantial refactor, and would probably best work with an object-oriented approach...

Additional Detail

This stems from #659. In that case, changing the random seed was helpful in some subjects.

Next Steps

1. Discuss.
2. Come up with an implementation, in particular deciding between our current functional approach and an object-oriented one.
3. Implement.

[eurunuela]

Not the fanciest implementation but we could add a while or a for loop (I'm thinking 2 iterations max) that only runs for a second time if the number of BOLD components is below a certain threshold.

[tsalo]

Good point. We could put that in the workflow function, although the ICA function would need to return the seed and/or number of restarts. Something like....

bad_decomp = True
n_restarts = 0
while bad_decomp:
    ica_results, n_restarts = ica_function(seed, maxrestarts=(maxrestarts - n_restarts))
    comptable, n_bold_comps = decision_tree(ica_results)
    seed += n_restarts
    if n_bold_comps > 2:
        bad_decomp = False

[eurunuela]

That looks good but I would limit the number of restarts. We don't want tedana to run for hours (I know it is an extreme case). So, I would restrict it to something like 2 or 3 restarts.

[tsalo]

We already have a limit on restarts (maxrestarts), so I'd leverage that if possible. In most cases (especially our five-echo test dataset), the main overhead is the ICA restarts, rather than the decision tree. If your main concern is overhead, I think the optimal solution would be to propose reducing our default value for maxrestarts.

[eurunuela]

Oh, I didn't see maxrestarts. Then yes, LGTM!