Parameter estimation in the magnitude-only and complex-valued fMRI data models

In functional magnetic resonance imaging, voxel time courses are complex-valued data but are traditionally converted to real magnitude-only data ones. At a large signal-to-noise ratio (SNR), the magnitude-only data Ricean distribution is approximated by a normal distribution that has been suggested as reasonable in magnitude-only data magnetic resonance images for an SNR of 5 and potentially as low as 3. A complex activation model has been recently introduced by Rowe and Logan [Rowe, D.B., and Logan, B.R. (2004). A complex way to compute fMRI activation. NeuroImage, 23 (3):1078-1092] that is valid for all SNRs. The properties of the parameter estimators and activation statistic for these two models and a more accurate Ricean approximation based on a Taylor series expansion are characterized in terms of bias, variance, and Cramer-Rao lower bound. It was found that the unbiased estimators in the complex model continued to be unbiased for lower SNRs while those of the normal magnitude-only data model became biased as the SNR decreased and at differing levels for the regression coefficients. The unbiased parameter estimators from the approximate magnitude-only Ricean Taylor model were unbiased to lower SNRs than the magnitude-only normal ones. Further, the variances of the parameter estimators achieved their minimum value in the complex data model regardless of SNR while the magnitude-only data normal model and Ricean approximation using a Taylor series did not as the SNR decreased. Finally, the mean activation statistic for the complex data model was higher and not SNR dependent while it decreased with SNR in the magnitude-only data models but less so for the approximate Ricean model. These results indicate that using the complex data model and not approximations to the true magnitude-only Ricean data model is more appropriate at lower SNRs. Therefore, since the computational cost is relatively low for the complex data model and since the SNR is not inherently known a priori for all voxels, the complex data model is recommended at all SNRs.