Joint Multiclass Classification of the subjects of Alzheimer's and Parkinson's Diseases through Neuroimaging Modalities and Convolutional Neural Networks

Alzheimer's disease (AD) is the most widespread type of dementia defined by an accumulation of amyloid-\beta proteins, the formation of tau plaques as well as the loss of neurons. On the other hand, Parkinson's disease (PD) is defined by the loss of dopaminergic neurons in the substantia nigra pars compacta within the midbrain. AD and PD are affecting millions of elderly people worldwide which highlights the need for their early diagnosis for the welfare of subjects diagnosed with these neurodegenerative disorders. A large number of autopsy confirmed PD subjects have sufficient postmortem plaque and tangle pathology to meet criteria for a second diagnosis of AD. Neuroimaging modalities such as Positron Emission Tomography (PET) and Single Photon Emission Computed Tomography (SPECT) are routinely used by clinicians to diagnose the initial stages of these two diseases. Together with these neuroimaging modalities, deep learning techniques are widely used in medical settings and have the potential to aid clinicians in the early diagnosis of these two diseases. In this work, we deployed a 3D Convolutional Neural Network (CNN) forjoint feature extraction and multiclass classification of both AD and PD brain images in the spatial and frequency domains using PET and SPECT imaging modalities discriminating between AD, PD and Normal Control (NC) classes. We used Discrete Cosine Transform (DCT) as the frequency domain method and deployed random weak Gaussian blurring and random zooming in/out augmentation methods in both spatial and frequency domains. Based on our experiments and deployment of cross-validation approach for optimal hyperparameters selection, we found the performance of AD/NC(SPECT)/PD classification with random weak Gaussian blurred augmentation in the spatial domain to be the best and that of AD/NC(SPECT)/PD classification with combined augmentations in the frequency domain to be the worst while spatial domain methods outperformed their frequency domain counterparts.