If you look at most studies that came out before it did, they were already using statistical approaches that were robust to this sort of problem.
These include Bonferroni correction, cluster correction for spatial distributions, and Monte Carlo simulations and permutation testing.
These techniques analyze what you would expect to happen if there were no signal but a lot of random noise. You can then look for signals that are still significantly stronger than predicted by the null hypothesis. This is what most fMRI studies do, and this is not done (intentionally) in the dead salmon study. If you use 'em, you find nothing in the salmon.
[1] http://en.wikipedia.org/wiki/Multiple_comparisons_problem
Edit: Indeed fMRI allows us to do things we had no hope of doing before, e.g. communication with those who are otherwise in permanent vegetative states [2,3] (that guy was my co-supervisor).
[2] http://www.sciencemag.org/content/313/5792/1402.abstract
And IIRC they were VS until such studies demonstrated that actually not all of them were (which was my point).
[edit] The above is a link to a blog post. Is this the paper: http://prefrontal.org/files/posters/Bennett-Salmon-2009.pdf