Estimation and removal of spatiotemporally structured noise infMRI data

The structured noise within functional magnetic resonance imaging (fMRI) data is a complex combination of multiple noise sources and is nonstationary in time. Because of the temporal nonstationarity of the structured noise, it is difficult to fully model using temporal characteristics. While the structured noise is temporally nonstationary, its sources are physiological and, therefore, have spatial characteristics that remain largely fixed in time. The goal of this research was to estimate and remove structured noise by exploiting its fixed spatial structure. The first step in this process was developing a filter using spatial correlations within the data to produce a noise estimate. This method required no special pulse sequences or monitoring equipment, needing only the collection of a brief baseline period before the fMRI protocol was begun.; While the spatially correlated filter was effective at removing structured noise, there were two drawbacks to this method. First, it required a low-threshold pretest to exclude potentially active pixels from the model. Second, the method was univariate and required approximately 4 s per time course. To remove these constraints, independent component analysis (ICA) was employed as a multivariate tool for spatiotemporal filtering.; Before ICA was used, its temporal and spatial characteristics were examined. It was found that, despite producing nonstationary component time courses, the spatial components ICA produced were consistent over time. Therefore, ICA provided a tool to extract the structured noise in all pixels while separating the structured noise from the activation.; ICA was used in a manner similar to the spatially correlated filter. The baseline period and the entire dataset were decomposed into an equal number of components. The components in the baseline and entire dataset were then matched and combined to form an estimate of the noise in the dataset. The filter produced results on par with the spatially correlated filter but did so in a fraction of the time. For an entire slice, the ICA filter only required 30 s to remove the structured noise from the data; the spatially correlated filter would have taken over 30 min. In addition to its speed, the ICA filter also required no prior knowledge of the response timing or shape.