Removing outliers can be problematic due to the arbitrariness involved. A more principled way is to handle it through the modeling process. If you obtain the contrast between the two conditions for each subject, you can try robust regression by feeding the contrast together with the covariate to 3dMVM with its option -robust
Gang