Abhishek,
Instead of using 3dmaskave to find out the active clusters for each subject and dealing the issue of different number of voxels, you might consider some alternative approaches. One way is to run voxel-wise group analysis with 3dANOVA2 -type 3 (or 3dttest) on the percent signal changes of the 3 conditions. Another approach is to extract the average or peak percent signal change of each cluster for each subject, and then run paired t tests.
Gang