Cyrus Raji MD, PhD, Presenter: Nothing to Disclose

Weiying Dai PhD, Abstract Co-Author: Nothing to Disclose

Oscar Lopez MD, Abstract Co-Author: Nothing to Disclose

H. Michael Gach PhD, Abstract Co-Author: Nothing to Disclose

Lewis H. Kuller MD, Abstract Co-Author: Nothing to Disclose

Paul Thompson PhD, Abstract Co-Author: Nothing to Disclose

Michael D. Kuo MD, Abstract Co-Author: Consultant, Boehringer Ingelheim GmbH Consultant, Confluence Life Sciences, Inc

James T. Becker PhD, Abstract Co-Author: Nothing to Disclose

Alzheimer's disease (AD) is the most common cause of dementia and early accurate diagnosis is of great interest. Perfusion MR imaging with arterial spin labeling (ASL) quantifies regional cerebral blood flow that can alter very early in the course of neurodegenerative disease when symptoms of memory loss are often subtle. Perfusion ASL MR imaging therefore holds promise to identify AD before symptoms are clinically expressed.  Machine learning methods such as support vector machine (SVM) offer a robust approach to quantitatively delineate normal individuals from AD. The purpose of this study was to test SVM for improved AD diagnosis on perfusion ASL MR imaging with structural MR for comparison.  

Study subjects were recruited from the population based Cardiovascular Health Study-Cognition study: 24 subjects, 12 controls and 12 persons with AD  in 2002-2003 by NINCDS-ARDA Criteria with average age of 78. All MRI data were acquired using a 1.5 T GE Signa system (Milwaukee, WI, LX Version), after each subject provided informed consent either directly or by their caregiver per with institutional review board approval.  Multi-slice continuous ASL was acquired. T1-weighted spoiled gradient-recalled echo (SPGR) images covering the whole brain were also acquired in orthogonal planes. SVM was applied on all structural and perfusion MR images using the Probid software (KC, London, http://tinyurl.com/l6frtdd).

Figure 1 shows screen shots from the Probid Graphical User Interface displaying results of an SVM analysis in both perfusion ASL (Figure 1a) and structural SPGR (Figure 1b) MR imaging. Class 1 (red circles) represents persons with AD and Class 2 (blue Xs) depict controls. Machine learning with SVM of perfusion ASL MR imaging is able to separate AD from control with 92% sensitivity, 92% specificity, and 92% accuracy. For SPGR MR structural imaging, classification was less effective with 42% sensitivity, 75% specificity, and 58% accuracy. 

Machine learning SVM methods in perfusion MR imaging are able to separate AD from control with high sensitivity, specificity, and accuracy. Applying the same methodology to SPGR images is comparatively less effective for the same purpose. 

Fully automated machine learning algorithms can be applied to perfusion ASL MR images for highly accurate identification of Alzheimer's dementia. Such methods may be readily applied in clinical environments for improved diagnosis. 

Raji, C, Dai, W, Lopez, O, Gach, H, Kuller, L, Thompson, P, Kuo, M, Becker, J, Application of a Support Vector Machine Learning Algorithm towards the Accurate Identification of Alzheimer's Dementia with Perfusion Arterial Spin Labeled MR Imaging.  Radiological Society of North America 2014 Scientific Assembly and Annual Meeting, - ,Chicago IL. http://archive.rsna.org/2014/14045576.html