Assessment of Linear and Non-linear Feature Projections for the Classification of 3-D MR Images on Cognitively Normal, Mild Cognitive Impairment and Alzheimer’s Disease

Abstract

Alzheimer’s disease (AD) is an age-related neurodegenerative disease and the most common form of dementia. It is a brain disorder that impacts the daily life of the patient due to memory loss and cognitive changes. Due to population aging and the fact that dementia incidence rising sharply at ages greater than 75, AD has become a major public health problem. Currently, hippocampal atrophy assessed on structural magnetic resonance (MR) images is the most used imaging biomarker of AD. Among many methods applied for automated classification of cognitively normal (CN), mild cognitive impairment (MCI), and AD, the linear PCA projection method, also known as eigenbrain, has shown effectiveness in AD subject prediction even with its restriction of using only linear projections. This study investigates both linear (PCA) and non-linear (kernel PCA) projection method performances to classify 3-D structural MR images on the CN, MCI, and AD classes. Support vector machines (SVMs) were trained using feature vectors with different space dimensions obtained by projecting MR study images into the previously created “eigenbrain spaces.” We tested the method using different kernel functions for both cases, the eigenbrain projections and SVM classifiers. We also conducted our analyses separately using the whole-brain and the gray-matter (GM) regions. Comparison results between PCA and kernel PCA methods showed that non-linear projections improved the classification of MR images in both class groups, particularly when processing the GM region - CN \(\times \) MCI (PCA: AUC = 0.63 versus KPCA: AUC = 0.76), MCI \(\times \) DA (PCA: AUC = 0.67 versus KPCA: AUC = 0.74), and CN \(\times \) AD (PCA: AUC = 0.86 versus KPCA: AUC = 0.89).