Article Preview
Top1. Introduction
Alzheimer’s disease (AD) is an incurable central neurodegenerative disorder that tends to occur in old age (Mayeux, 2010). The number of people with AD has increased dramatically due to the aging population. It is expected that there will be 60 million AD patients worldwide in 50 years (Alzheimer’s Association, 2019). AD is incurable; therefore, the best strategy to control AD is to identify people at higher risk of developing the disease as early as possible and intervene to prevent its effects. Due to the unclear etiology of AD, clinical diagnostic methods are limited to neuropsychological tests (Kowoll et al., 2015; Zhang et al., 2019).
Mild cognitive impairment (MCI) is a precursor to dementia. Patients with MCI present with moderate symptoms like mild memory loss. The symptoms do not affect independent living, which makes it more difficult to detect and, in turn, more likely to develop into AD. A growing number of studies have shown that magnetic resonance imaging (MRI) can observe progressive brain loss in patients with MCI to complete AD (Whitwell et al., 2008).
Higher imaging parameters and high soft tissue resolution of the MRI compares to computed tomography (CT) images. MRI is the most used diagnostic image of the brain due to its accurate display of the 3D structure of the brain (see Figure 1). MRIs are widely used in clinical applications as an adjunct to neuropsychological testing for organic lesions. However, in clinical applications, the judgment of patients’ MRI is highly dependent on the physician's consulting experience. This leads to the judgment of the patients’ condition being limited to the ability of the physician.
Figure 1. Presentation of 3D MRI slices from three directions
The convolutional layer is the most critical component of Convolution Neural Networks (CNNs). As a feature extractor, the convolutional kernel works by sweeping through the input features in a regular manner, summing the matrix elements and superimposing the bias amount within the perceptual field. This architecture introduces the inductive bias of locality and spatial invariance to CNNs, allowing CNNs to excel in image processing tasks. The sliding feature extraction approach uses convolutional kernels smaller than the feature map size, which allows the convolutional kernels to extract local features at one time. In addition, key feature changes to any position within the feature map can be sensed during the sliding process. Such an operation allows neurons to only connect to a small region of the previous layer of the feature map. This greatly reduces the number of parameters in the network (Albawi et al., 2017).
Several approaches have been used to classify AD via convolutional neural networks. Zhang et al. (2019) used a classification network with a deep small convolutional kernel to classify AD via sliced images. Li et al. (2017) proposed a multimodal network with a combination of 2D and 3D convolution for the classification of Alzheimer’s patients and normal individuals. However, the convolution operation focuses on local feature extraction, and the MRI global association information in the image is easily lost. Therefore, this article attempts to implement the classification of AD using the transformer model.