MACHINE LEARNING-BASED CLASSIFICATION OF FDG-PET IMAGES FROM ADNI DATABASE

Wenlu Yang, Fangyu He, Nigel H. Greig, Debomoy K. Lahiri, Jack T. Rogers and Xudong Huang

Department of Electrical Engineering, College of Information Engineering, Shanghai Maritime University, Shanghai 201306, China; Laboratory of Neurosciences, National Institute on Aging, NIH, Baltimore, MD 21224, USA; Department of Psychiatry, Institute of Psychiatric Research, Neuroscience Research Center, Indiana University School of Medicine, Indianapolis, IN 46202, USA; Neurochemistry Laboratory, Departments of Psychiatry, Massachusetts General Hospital and Harvard Medical School, Charlestown, MA 02129, USA

Abstract

Numerous studies and many clinical trials strongly suggest that early diagnosis of the Alzheimer's disease (AD), the most common form of age-related dementia, is the key for more effective therapeutic intervention and clinical management of this devastating disease. In particular, early detection of Mild Cognitive Impairment (MCI), the potential AD precursor, may be essential to delay or even prevent the AD onset. Herein, we propose a novel Machine Learning (ML)-based method combining voxel of interest in FDG-PET images and the neuropsychiatric test results from ADNI (Alzheimer’s Disease Neuroimaging Initiative) database for automatic classification of AD or MCI subjects vs healthy control (HC) subjects in ADNI database. This classification procedure includes four steps: pre-processing images, extracting independent components using Independent Component Analysis (ICA) method, selecting voxel of interest, and classifying images using a Support Vector Machine (SVM) classifier. FDG-PET image data were obtained from ADNI database and they include 91 HC subjects, 50 AD subjects, and 105 MCI subjects. The neuropsychiatric test results for these subjects were also included in the data matrix for classification. We thus achieved excellent discrimination of AD patients from HC subjects- accuracy: 97.49%; sensitivity: 93.52%; specificity: 99.65%, and good discrimination of MCI patients from HC subjects- accuracy: 94.48%; sensitivity: 92.71%; specificity: 96.52%.

We conclude that our ML-based method can successfully distinguish between AD or MCI subjects and HC subjects. Therefore, our proposed method may be adequate for automatic and accurate classification of FDG-PET images from ADNI database.

Supported by grants from Ellison Medical Foundation, Alzheimer’s Association, and NIA/NIH.