Parameterization of motion artifacts in fMRI time series using autoregressive models for the construction of computer-generated phantoms

We explore the use of scalar and multivariate autoregressive models to parameterize motion artifacts in fMRI time series. To do so, we acquire real fMRI data sets, measure rigid body motion in these data sets, and classify the type of observed motion in several categories such as random motion or motion correlated with activation. The measured motion sequences are then modeled and used to generate realistic image phantoms that can be used to validate fMRI data analysis packages. We compare phantoms generated with the original motion sequences and phantoms generated with simulated sequences. We show that both scalar and multivariate autoregressive models can be used to generate realistic motion sequences. An important difference between the two is the fact that multivariate models can capture correlations between motion parameters, which cannot be done with scalar models.