(1) With a limited number of subjects, there isn't much you can do other than analyze subjects with complete data.
(2) The issue with looking at the correlation for all subjects at T1 versus the 12 subjects at T3 is that the correlation may not change, but could be different in the subjects that came back, so you are limited to the 12 subjects.
If you had many more subjects, there are more advanced statistical models that you could use that allow for missing data at later timepoints (e.g. mixed-efffect models and parallel growth models).