Anisotropic regularization of posterior probability maps using vector space projections. application to MRI segmentation

Abstract

In this paper we address the problem of regularized data classification. To this extent we propose to regularize spatially the class-posterior probability maps, to be used by a MAP classification rule, by applying a non-iterative anisotropic filter to each of the class-posterior maps. Since the filter cannot guarantee that the smoothed maps preserve their probabilities meaning (i.e., probabilities must be in the range [0, 1] and the class-probabilities must sum up to one), we project the smoothed maps onto a probability subspace. Promising results are presented for synthetic and real MRI datasets.