Introduction
Alzheimer’s disease (AD), the most common form of dementia, is a major challenge for healthcare in the twenty-first century. An estimated 5.5 million people aged 65 and older are living with AD, and AD is the sixth-leading cause of death in the United States. The global cost of managing AD, including medical, social welfare, and salary loss to the patients’ families, was $277 billion in 2018 in the United States, heavily impacting the overall economy and stressing the U.S. health care system (Alzheimer’s Association, 2018). AD is an irreversible, progressive brain disorder marked by a decline in cognitive functioning with no validated disease modifying treatment (De strooper and Karran, 2016). Thus, a great deal of effort has been made to develop strategies for early detection, especially at pre-symptomatic stages in order to slow or prevent disease progression (Galvin, 2017; Schelke et al., 2018). In particular, advanced neuroimaging techniques, such as magnetic resonance imaging (MRI) and positron emission tomography (PET), have been developed and used to identify AD-related structural and molecular biomarkers (Veitch et al., 2019). Rapid progress in neuroimaging techniques has made it challenging to integrate large-scale, high dimensional multimodal neuroimaging data. Therefore, interest has grown rapidly in computer-aided machine learning approaches for integrative analysis. Well-known pattern analysis methods, such as linear discriminant analysis (LDA), linear program boosting method (LPBM), logistic regression (LR), support vector machine (SVM), and support vector machine-recursive feature elimination (SVM-RFE), have been used and hold promise for early detection of AD and the prediction of AD progression (Rathore et al., 2017).
In order to apply such machine learning algorithms, appropriate architectural design or pre-processing steps must be predefined (Lu and Weng, 2007). Classification studies using machine learning generally require four steps: feature extraction, feature selection, dimensionality reduction, and feature-based classification algorithm selection. These procedures require specialized knowledge and multiple stages of optimization, which may be time-consuming. Reproducibility of these approaches has been an issue (Samper-Gonzalez et al., 2018). For example, in the feature selection process, AD-related features are chosen from various neuroimaging modalities to derive more informative combinatorial measures, which may include mean subcortical volumes, gray matter densities, cortical thickness, brain glucose metabolism, and cerebral amyloid-β accumulation in regions of interest (ROIs), such as the hippocampus (Riedel et al., 2018).
In order to overcome these difficulties, deep learning, an emerging area of machine learning research that uses raw neuroimaging data to generate features through “on-the-fly” learning, is attracting considerable attention in the field of large-scale, high-dimensional medical imaging analysis (Plis et al., 2014). Deep learning methods, such as convolutional neural networks (CNN), have been shown to outperform existing machine learning methods (Lecun et al., 2015).
We systematically reviewed publications where deep learning approaches and neuroimaging data were used for the early detection of AD and the prediction of AD progression. A PubMed and Google Scholar search was used to identify deep learning papers on AD published between January 2013 and July 2018. The papers were reviewed and evaluated, classified by algorithms and neuroimaging types, and the findings were summarized. In addition, we discuss challenges and implications for the application of deep learning to AD research.