Lattice independent component analysis feature selection on diffusion weighted imaging for Alzheimer's disease classification

Abstract Diffusion weighted imaging (DWI) provides information on the diffusion of water molecules which can be useful to determine structural properties in the brain. Specifically, fractional anisotropy (FA) is a scalar measure computed from each voxel's diffusion tensor giving information about the existence of a privileged diffusion direction. The FA volume is the raw data in our classification approach. We apply lattice independent component analysis (LICA) across volumes for feature selection on FA data to perform classification of healthy control (HC) subjects and Alzheimer's disease (AD) patients. Feature selection is done on the basis of Pearson's correlation between the LICA residuals at each voxel site and the data indicative variable. Voxel sites having an absolute value Pearson's correlation above a given percentile of its empirical distribution are selected as feature variables for classification. We compare the LICA based feature selection with (a) a Pearson's correlation approach on the raw FA data, and (b) a voxel based morphometry (VBM) approach. We apply relevance vector machines (RVM), nearest-neighbor (1NN) and linear support vector machines (LSVM) to build classifiers on these feature vectors. LSVM reach very high accuracy, specificity and sensitivity for some feature selection percentile parameter values. We provide results of the approach on data coming from an on-going study in Hospital de Santiago Apostol collecting anatomical T1-weighted MRI volumes and DWI data of HC and AD patients. Results point to the validity of FA data as an image-marker for AD.