Blind estimation of fMRI data for improved BOLD contrast detection

Variations due to noise about the baseline MR signal make detection of BOLD contrast in fMRI data difficult for voxels with weak activation. We present a new wavelet- and Fourier-based estimation technique that improves the ability of a t-test to detect BOLD contrast in fMRI data. Our scheme approximates the optimal linear estimator for an fMRI dataset using a 3-D discrete wavelet transform to decorrelate in space and the discrete Fourier transform to decorrelate in time. In contrast to the optimal estimator, which is useful only in theory as it requires second-order signal and noise statistics, the proposed technique is able to achieve blind estimation of fMRI data. Applying this estimator to fMRI data improves the ability to correctly detect BOLD contrast, especially for voxels with contrast levels between 1% and 2%. In addition, the proposed method produces increased confidence (lower p-value) in active voxels of both synthetic and experimental fMRI data (compared to an unestimated version of the same voxels).