Spatially Adaptive Kernels for Adaptive Spatial Filtering of fMRI Data

Making use of neighborhood time series is an effective way of noise reduction in fMRI data. However the conventional averaging methods blur activated areas. In this paper, a filter with adaptive kernel is designed such that its kernel size and direction are defined at each voxel. First, for finding the optimum size of kernel, a linear combination of some isotropic Gaussian filters with different variances is used, and optimum value of variance is specified. Then, the appropriate kernel direction is determined by a linear combination of some anisotropic basic filters with various directions. The weights of these linear combinations can be calculated by using the restricted canonical correlation method. The proposed method is compared with a similar method based on the steerable filters and the results show that the proposed method improves the ROC curve and prohibits false spread of the activation areas.