DICOVERY OF "BIOMARKERS" FOR ALZHEIMER'S DISEASE PREDICTION FROM STRUCTURAL MR IMAGES

We propose a computational framework for learning predictive image features as "biomarkers" for Alzheimer's disease discrimination using high-resolution magnetic resonance (MR) brain images. We focus on the exploration of a very large (>500 million) feature space derived extensively from the deformation and tensor fields. In such a huge space, our computational tool supports an automatic search for discriminative feature subspaces and the corresponding anatomical regions in human brains, which can be used to discriminate previously unseen, individual structural MR images from Alzheimer' disease (AD) and normal control (CTL) subjects. Our aggressive leave-ten-out cross-validations on 40 subjects demonstrate higher than 90% sensitivity and specificity. In addition, we demonstrate intriguing anatomical locations as automatically discovered "biomarkers" and the spatial distributions of 20 mild cognitive impairment (MCI) subjects in the discriminative feature space automatically learned for AD and CTL separations. Our results illustrate a truly complementary effort of human and computers for early diagnosis of AD from MR images.