Wavelet shrinkage versus Gaussian spatial filtering of functional MRI data

Objective The comparison between wavelet-based simultaneous multiscale denoising/hypothesis testing and single scale Gaussian spatial filtering followed by statistical testing of brain activation maps is carried out for a simple block-type visual paradigm EPI MRI experiment. Probabilistic wavelet shrinkage provides means to consistently combine multiresolution denoising and hypothesis testing in a single process. The discrete wavelet transform (DTW) performs scale-varying decomposition of spatial statistic maps, so denoising in the wavelet domain is more adaptive to some spatial features in the true image than applying Gaussian filtering in the spatial domain. DTW also exhibits decorrelating properties, which amounts to mutually independence of the hypothesis tests on the wavelet coefficients and conferring potential benefits in the optimal control of type I error (false positives) [1]. The multiple comparison problem was handled using false discovery rate (FDR) control in wavelet denoising, while familywise error (FWE) control was used for isotropic symmetric Gaussian spatial smoothing required by thresholding the statistical parametric maps with the random field theory (RFT).