Hi Yisheng,
The ARMA modeling is for the noise in the residual. I am not sure what you mean by " works on covariates and takes care of the latency-variation problem ...".
A cautionary note: Removing WM mean, or CSF mean, can end up being similar to removing the average GM signal, to the extent that your masks have partial voluming with the GM tissue.
How different are the ventilation patterns, and do you have respiration belt measurements acquired during the scans?
cheers,
ziad