fMRI Noise Reduction Based on Canonical Correlation Analysis and Surrogate Test

In this paper, we proposed a noise-reduction method for functional magnetic resonance imaging (fMRI). We classified noise into structured and unstructured ones. Canonical correlation analysis was exploited to extract the underlying components among which the structured ones were recognised. Furtherly, The task-related components were detected among the structured ones by using surrogate test based on reduced autoregression model. The low degree of temporal correlation of the unstructured residuals was reduced by using randomization technique. The task-related components and the randomly permuted unstructured residuals were used to generate the reconstructed data. With application of our method, SNR of data can be significantly improved. In addition, the temporal correlation of unstructured background noise can be efficiently reduced. Twenty sets of true fMRI data for finger tapping task were processed. Some task-related areas which cannot be detected from the original data set were discerned.