Efficient selection of non-redundant features for the diagnosis of Alzheimer'S disease

Recently, a large research effort has been made on the development of discriminative techniques for the computer-aided diagnosis (CAD) of both Alzheimer's disease (AD) and Mild Cognitive Impairment (MCI) using neuroimages as the main source of information. Often, such systems use the Voxel Intensities (VI) directly as features, and a feature selection procedure is needed in order to tackle the curse of dimensionality. In this paper, we will propose an efficient selection algorithm based on Mutual Information which, unlike the procedures typically used within this research field, is able to avoid the redundancy existing between brain voxels that are typically highly dependent. The proposed approach was able to join a higher amount of relevant information in a feature vector of fixed dimension and, therefore, was able to improve the classification performance attained when using a typical selection procedure.