A mutual-information scale-space for image feature detection and feature-based classification of volumetric brain images

This paper proposes a novel information theoretic scale-space for salient feature detection, based on the mutual information (MI) of image measurement and location. The MI scale-space is designed to identify image regions whose measurements are maximally informative regarding spatial location. A framework for computing the MI scale-space is proposed, based on combining information theory with Gaussian scale-space theory, where uncertainty in spatial location is explicitly defined by the heat equation. Experiments investigate the use of MI features for feature-based classification of Alzheimer's subjects in volumetric magnetic resonance imagery from a public data set, where MI features result in higher classification accuracy than features selected according to the established difference-of-Gaussian (DOG) criterion.