Alzheimer's Disease (AD) is a neurological disorder affecting the functioning of central nervous system. It can lead to poor coordination, seizures and paralysis. Neuroimaging modalities such as Magnetic Resonance Imaging (MRI) and Positron Emission Tomography (PET) can provide important information about AD and will continue to do so in the future as far as clinical manifestations of this disease are concerned. Information from neuroimaging modalities can be combined with deep learning (DL) approaches to diagnose AD in its early stages, reducing the burden on neuropathologists. In this study, we compared the performances of six data augmentation methods -ellipsoidal averaging, Laplacian of Gaussian (LoG), local Laplacian, local contrast, Prewitt-edge emphasising, and unsharp masking -on AD diagnosis. We studied three binary problems: AD-Normal Control (NC), AD-Mild Cognitive Impairment (MCI), and MCI-NC, and one multiclass (3-classes) classification problem: AD-MCI-NC. We also combined these data augmentation methods and tried a strided convolution architecture for these tasks. We find that Prewitt-edge emphasising augmentation yields the best performance for AD-MCI-NC and AD-MCI classification tasks. In contrast, local Laplacian augmentation performs the best for the MCI-NC classification task, while LoG augmentation yields the best results for the AD-NC classification task.
Athar et al. (Wed,) studied this question.