P-065: Characterization of MCI using patterns of atrophy recognized by a support vector machine

Abstract

Amnestic mild cognitive impairment (aMCI) is a prodromal stage for Alzheimer's disease (AD). Clinically and psychometrically, aMCI subjects occupy an intermediate position between control and AD subjects. Libraries of MRI scans from AD and control subjects can in theory be used to predict the clinical status of individual subjects who lie at any point along the entire cognitive continuum. Characterizing patterns of atrophy in subjects predisposed and pre-symptomatic to AD can aid clinical characterization. To develop a support vector machine (SVM) learning tool for characterization of aMCI patients using libraries of structural MRI scans of clinically well characterized AD patients and normal controls. Three-dimensional MRI scans were obtained for 100 patients with clinically diagnosed probable AD and 100 age- and gender-matched controls. Smoothed, modulated, segmented and normalized GM, WM and CSF tissue densities for all 200 image volumes were obtained. Concatenation of the tissue densities gives a single vector which is the signature of each subject. A SVM was trained using the signatures and diagnoses of the 200 subjects and it learns to predict the patient class (+1 for AD and -1 for controls) with maximum possible accuracy. For each new incoming scan, the signature is extracted and input to the existing SVM model. The output represents an “abnormality magnitude score” and if it lies between +1 and -1 it indicates that the subject has a degree of abnormality intermediate between control and AD and can be classified as aMCI. The patterns identified by SVM to differentiate AD subjects from controls with maximum accuracy (in our case 93%) are consistent with known distribution of fibrillary AD pathology (Fig. 1). The SVM output for the aMCI subjects lie in an intermediate position between control and AD subjects and can be identified quite well (Fig. 2). Classifying an individual's MRI scan in relation to a library of scans can provide useful diagnostic information. SVM has significant advantages over traditional region-of-interest (ROI) approaches as it utilizes the diagnostic information embedded in the entire 3D MRI data array in each subject.