First of all, the two expressions are essentially the same thing: (A1B1-A1B2)-(A2B1-A2B2) = (A1B1-A2B1)-(A1B2-A2B2). Secondly, they are for the interaction between factors A and B in the case of a 2 x 2 design.
> does (A1B1-A2B1)-(A1B2-A2B2), for example, mean I do the subtraction of A1B1-A2B1 and the subtraction
> A1B2-A2B2 for each subject and the run a paired sample t-test between the resultant maps?
Yes, you can do two subtractions and then perform a paired t-test. Or, you can get (A1B1-A2B1)-(A1B2-A2B2) for each subject, and then run a one-sample t-test. The results would be the same.
> in which of these ttests should I run multiple compairson correcting permutation tests? Running on all of
> them seems wrong, but so does running it on jus the final ttest.
There are two aspects of multiple comparisons involved here. One is across the voxels in the brain for each statistical inference, and the other is the number of tests (interaction, main effects, and other comparisons). The first one is the major aspect everyone focuses on. There is no easy solution currently available for the second one.
Gang