We have seen several feature selection approaches for ultrahigh dimensional data recently. However, most of the approaches are applicable for n*1, or at most n*q response variables. If we encounter a more complex data structure when the response is tensor-shaped, e.g. human facial shape, multi-omics gene data, Electroencephalography (EEG), or functional magnetic resonance imaging (fMRI), current feature selection approaches will have difficulty to handle the data. We propose a simple yet useful feature selection method that can be applied to tensor response data, and show the selection consistency of our method.