Permutation Tests for Classification: Towards Statistical Significance in Image-Based Studies

Estimating statistical significance of detected differences between two groups of medical scans is a challenging problem due to the high dimensionality of the data and the relatively small number of training examples. In this paper, we demonstrate a non-parametric technique for estimation of statistical significance in the context of discriminative analysis (i.e., training a classifier function to label new examples into one of two groups). Our approach adopts permutation tests, first developed in classical statistics for hypothesis testing, to estimate how likely we are to obtain the observed classification performance, as measured by testing on a hold-out set or cross-validation, by chance. We demonstrate the method on examples of both structural and functional neuroimaging studies.