Searchlight Based Feature Extraction

A multi voxel pattern analysis classi cation framework suitable for neuroimaging data is introduced. The framework includes a novel feature extraction method that uses local modeling based on domain speci c knowledge, and therefore, can produce better whole-brain global classi cation performance using a smaller number of features. In particular, the method includes spherical searchlights in combination with local SVM modeling. The performance of the framework is demonstrated on a challenging fMRI classi cation problem, and is found to be superior to the performance of state-of-the-art feature selection methods used in neuroimaging.