Meeting Banner
Abstract #3747

Atlas-Based Quantification with Machine-Learning Based Characterization of DTI from Patients with Mild Cognitive Impairment and Alzheimers Disease

Kenichi Oishi1, Michelle M. Mielke2, 3, Michael I. Miller4, Marilyn S. Albert5, Constantine G. Lyketsos2, Susumu Mori, 6

1Radiology, Johns Hopkins University, Baltimore, MD, United States; 2Psychiatry and Behavioral Sciences, Johns Hopkins University; 32Division of Epidemiology, College of Medicine, Mayo Clinic; 4Center for Imaging Science, Johns Hopkins University; 5Neurology, Johns Hopkins University; 6F.M. Kirby Research Center for Functional Brain Imaging, Kennedy Krieger Institute

We applied a machine-learning framework to characterize anatomical alterations of early-stage Alzheimers disease (AD), and to investigate a classifier that can predict conversion from mild cognitive impairment (MCI) to AD within 36 months. The Eve atlas was used to measure the fractional anisotropy (FA) and mean diffusivity (MD) of 148 brain structures, followed by principal component analysis (PCA) and support vector machine-based classification. PCA detected subtle but widespread FA&MD alterations related to AD. The trained classifier could differentiate MCI-converters from non-converters with sensitivity of 0.67 and specificity of 1, supporting the potential of DTI in identifying early-stage AD patients.

Keywords

ability accuracy affine alterations among amount anatomical anisotropy applicable applied approaches atlas automated available behavioral brain calculations called care cause characterization classification clinical cognitive cognitively college combined complement component components conversion converted corrected correlated correlation cross dataset detecting detects developed diagnosis differentiate differentiated differentiates diffusion diffusivity disease distributed division early effective elderly enough epidemiology equipped explained extracted features field findings focused fold future highlight hyper impairment impairments indicate indicated individual individuals institute investigate johns learning likelihood location machine manageable many maps matrix metric mild million modalities necessarily neurology neuronal next options palliative pathology patients positive possibility previously principal projections psychiatry pursued quantification quantitative radiology rank recent reduce reductions related remained replication resultant rule scanner sections sensitive sensitivity smith spatial specificity stable stages started statistics structure structures studies suggest suitable support supported systematically targeting tensor therapies tract trained transformation transformed treatment trials units validation variance vast vector view whole widely widespread years