Local discriminative characterization of MRI for Alzheimer's disease
Ahmad Chaddad, Christian Desrosiers, Matthew Toews · 2016
A novel method is proposed for characterizing Alzheimer's disease (AD) in brain MRI using local image texture features. Texture features are computed from local sub-volumes of T1-weighted MR images, and automatic classification performance is used to identify the (brain region, texture feature) combinations that are most discriminative regarding subject groups, i.e. AD vs healthy subjects. Experiments include MRI data of 124 subjects from the public OASIS database, three commonly-used texture feature types including the 3D-GLCM, 3D-DWT and LoG filters and random forest classification. The method identifies numerous (brain region, texture feature) combinations leading to high classification accuracy (> 70%), including several regions not traditionally linked to AD. These may indicate novel computational biomarkers for computer-assisted diagnosis or characterization of AD. The approach is generally applicable to other 3D data and disease contexts.