Post-registration spatial filtering to reduce noise in functional MRI data sets.

Image registration is commonly performed in the analysis of functional magnetic resonance imaging data sets. However, the nature of artifacts introduced in the process of alignment has not been well described. In this manuscript, high-frequency losses inherent in image registration are discussed, together with a conceptual division into acquisition- and resampling-related artifacts. Simulated and experimental data are presented to illustrate these artifacts. In simulations comparing corrected and reference images, root mean square (RMS) difference errors of 0.74% and 2.62% were observed following the correction of one degree of rotation for images registered with frequency regridding and bilinear interpolation, respectively. In human experiments, regression of RMS difference error as a percentage of mean brain signal yielded slopes of 0.69 to 1.31% per degree corrected by regridding. A post-registration spatial filtering technique is presented to reduce noise introduced during registration by selectively attenuating high frequencies near the corners of k-space. Filtering following regridding resulted in reductions in RMS error of 49.6% for simulated data and of 17.4% to 32.5% in human experiments, demonstrating the effectiveness of the filtering technique.