The option to remove the small "clusters" in the source maps is only a display option. This should not be used for statistical analysis, the results of the statistical tests should not be edited to remove "small clusters" after the computation of the p-values, otherwise the correction for multiple comparisons is not valid anymore.
If you want to be able to explore your results by spatio-temporal clusters, and keep only the significant clusters, you need to use cluster-based statistics:
https://neuroimage.usc.edu/brainstorm/Tutorials/Statistics#Example_3:_Cluster-based_correction