Functional brain mapping with locally smoothed regression

High-resolution functional magnetic resonance imaging (hi-res fMRI) methodology offers an opportunity for neuroscientists to gain insight about brain activities at a finer scale, and is thus becoming increasingly common. Traditional voxel-wise general linear model (GLM) is not suitable for hi-res functional brain mapping because local averaging may lose valuable fine-grained information boasted by hi-res fMRI. The searchlight approach may be more suited for this situation, but it can be improved to integrate multi-voxel information more completely and effectively. We propose a locally smoothed regression (LSR) to find the spatial organizations of neural activities, especially for hi-res data. LSR is a flexible model whereby the traditional voxel-wise regression can be seen as a special case of LSR. Further, LSR can be integrated into Mahalanobis-distance-based searchlight framework. This new approach promises to provide improved and reliable activation mapping as illustrated here by applying it to analyze a real set of data using hi-res fMRI imaging.