Denoising of event-related fMRI data based on a Rician noise model for robust analysis using

Given the fundamental constraints in ICA of having the number of sources less than that of signals, the number of independent noise signals must be lowered to enable this technique to work To ths end, the new scheme uses spectrum subtraction denoising a s s u m g a Rician noise model We derive the model characteristxs and present the method for its implementation We demonstrate the potential of this method as a preprocessing step to enable the ICA algorithm to converge and to provide efficient separation of signal components Inntroduction The Door sirmal-to-noise ratio ( S N R) of event-related fMRI., \, data triggered some work that addressed the denoising of such data (1-2). Denoising is essential for subsequent analysis steps to work. In particular, in ICA, the main assumption is that the number of signals available is higher than the number of underlying independent sources. Given that each signal must contain an independent noise component in addition to the possible other physiological fluctuations and true activation components, denoising must be employed to reduce the number of effective noise components to satisfy the fundamental condition of ICA. In this work, we propose a strategy that estimates in the noise from the acquired data for denoising and demonstrate its effectiveness as a preprocessing step to ICA. Theow The technique is based on spectrum subtraction (3). In this technique, a denoised power spectrum is obtained by subtracting the noise power spectrum model from the power spectrum of the original signal. Then, the denoised signal is computed using magnitude reconstruction from the square root of this power spectrum. Simpler reconstruction can be obtained by using a phase estimated from phase of the Fourier transform of the original signal. Unlike a technique described earlier (3), we rely on the more realistic Rician noise model rather than the generic Gaussian noise model that does not apply to fMRI data in general. The use of such denoising makes the subsequent use of ICA more robust computationally and enables the separation of different signal sources more efficiently