Nonlinear denoising of functional magnetic resonance imaging time series with wavelets

In functional magnetic resonance imaging (fMRI) the blood oxygenation level dependent (BOLD) effect is used to identify and delineate neuronal activity. The sensitivity of a fMRI-based detection of neuronal activation, however, strongly depends on the relative levels of signal and noise in the time series data, and a large number of different artifact and noise sources interfere with the weak signal changes of the BOLD response. Thus, noise reduction is important to allow an accurate estimation of single activation-related BOLD signals across brain regions. Techniques employed so far include filtering in the time or frequency domain which, however, does not take into account possible nonlinearities of the BOLD response. We here evaluate a previously proposed method for nonlinear denoising of short and transient signals, which combines the wavelet transform with techniques from nonlinear time series analysis. We adopt the method to the problem at hand and show that successful noise reduction and, more importantly, preservation of the shape of individual BOLD signals can be achieved even in the presence of in-band noise.