A Robust Normalization Method for fMRI Data for Brain Decoding

Functional Magnetic Resonance Imaging (fMRI) methods produce high dimensional representation of cognitive processes under heavy noise due to the limitations of hardware and measurement techniques. In order to reduce the noise and extract useful information from the fMRI data, a sequence of pre-processing techniques, such as smoothing with spatial filters and z-scoring, are used. In this study, we suggest an additional normalization technique based upon a statistical property of fMRI data. We, first, define a random variable V(t) as the average voxel intensity value of a brain volume measured at a time instant t. Then, we measure the Pearson correlation between V(t) and 1/V(t) for all time instances. We observe that the Pearson correlation values are very close to -1, indicating that V(t) and 1/V(t) have a strong negative correlation. We show that one explanation for this property is V(t) being almost surely constant and the small fluctuations on V(t) caused by noise. The proposed method removes these fluctuations on the data resulting in almost surely constant brain volumes V(t) for all values of t. The effectiveness of the proposed normalization method is tested with well-known brain decoding algorithms and voxel selection methods. It is observed that the suggested normalization method improves the performance 1-2 percent on the average. The method also improves the signal to noise ratio.