Deformation based feature selection for Computer Aided Diagnosis of Alzheimer's Disease

Deformation-based Morphometry (DBM) allows detection of signi cant morphological di erences of brain anatomy, such as those related to brain atrophy in Alzheimer's Disease (AD). DBM process is as follows: First, performs the non-linear registration of a subject's structural MRI volume to a reference template. Second, computes scalar measures of the registration's deformation eld. Third, performs across volume statistical group analysis of these scalar measures to detect e ects. In this paper we use the scalar deformation measures for Computer Aided Diagnosis (CAD) systems for AD. Speci cally this paper deals with feature extraction methods over ve such scalar measures. We evaluate three supervised feature selection methods based on voxel site signi cance measures given by Pearson correlation, Bhattacharyya distance and Welch's t-test, respectively. The CAD system discriminating between healthy control subjects (HC) and AD patients consists of a Support Vector Machine (SVM) classi er trained on the DBM selected features. The paper reports experimental results on structural MRI data from the cross-sectional OASIS database. Average 10-fold cross-validation classi cation results are comparable or improve the state-of-the-art results of other approaches performing CAD from structural MRI data. Localization in the brain of the most discriminant deformation voxel sites is in agreement with ndings reported in the literature.