Comparison of filtering methods for fMRI datasets

When studying complex cognitive tasks using functional magnetic resonance imaging (fMRI) one often encounters weak signal responses. These weak responses are corrupted by noise and artifacts of various sources. Preprocessing of the raw data before the application of test statistics helps to extract the signal and can vastly improve signal detection.

Artifact sources and algorithms to handle them are discussed. In an empirical approach targeted to yield an optimal recovery of the hemodynamic response, we implemented a test bed for baseline correction and noise-filtering methods. A known signal is modulated onto foreground patches obtained from event-related fMRI experiments.

Quantitative performance measures are defined to optimize the characteristics of a given filter and to compare their results. Marked improvements in the sensitivity and selectivity are achieved by optimized filtering. Examples using real data underline the usefulness of this preprocessing sequence.