Optimization of MDL-based wavelet denoising for fMRI data analysis

Denoising is an important preprocessing step to remove the signal noise with minimum effect on informative part. Wavelet transform is usually used for denoising through some criteria such as Minimum Description Length (MDL) which provides a suitable thresholding value for denoising. In this paper, the wavelet denoising via MDL is optimized in terms of wavelet function, decomposition level and noise type for HRF estimation as well as activation detection in vision region of task-based fMRI data. Simulations show that the MDL-based denoising performance is independent from the noise type for both Refined- and Crude- MDLs. According to simulations, it is necessary to select a scaling function being the most similar to Hemodynamic Response Function (HRF) involved in the experimental fMRI data. Besides, R-MDL can lead to optimum denoising at lower decomposition level compared to C-MDL. Applying MDL-based denoising to fMRI data as a preprocessing step, a larger set of activated voxels for vision tasks has been obtained which appear to be more realistic in comparison to earlier works.