The aim of this work was to identify ventricular shape-based biomarkers in MR images to discriminate between patients with Alzheimer's disease (AD) and healthy elderly. Clinical MR images were collected for 58 patients and 28 age-matched healthy controls. After normalizing all the images the ventricular cerebrospinal fluid was semiautomatically extracted for each subject and an innovative technique for fully automatic shape modeling was applied to generate comparable meshes of all ventricles. The search for potential biomarkers was carried out with repeated permutation tests: results highlighted well-defined areas of the ventricular surface being discriminating features for AD: the left inferior medial temporal horn, the right medial temporal horn (superior and inferior), and the areas close to the left anterior part of the corpus callosum and the head of the right caudate nucleus. The biomarkers were then used as features to build an intelligent machine for AD detection: a Support Vector Machine was trained on AD and healthy subjects and subsequently tested with leave-1-out experiments and validation tests on previously unseen cases. The results showed a sensitivity of 76% for AD, with an overall accuracy of 84%, proving that suitable biomarkers for AD can be detected in clinical MR images.
(c) 2008 Wiley-Liss, Inc.