Multiclass classification of initial stages of Alzheimer's Disease through Neuroimaging modalities and Convolutional Neural Networks

Alzheimer's Disease (AD) is the most common form of dementia that causes memory related brain changes which impair the thinking patterns of its subjects. Magnetic Resonance Imaging (MRI) and Positron Emission Tomography (PET) are widely used modalities to capture the structural changes in the brain caused by AD in its early stages. Early diagnosis of AD is important from clinical perspective to improve the life of an individual who is at the risk of developing memory deficits. Deep learning architectures such as 2D and 3D Convolutional Neural Networks (CNNs) have shown promising performances in extracting features and building useful representations of data for computer vision tasks. This study is geared towards understanding the performance differences between these architectures. We used transfer and non-transfer learning approaches to study the underlying disease phenomenon. In our experiments on three class classification of early stages of AD, we found the performance of 3D architectures to be better in comparison to their 2D counterparts. We found the performance of 3D architecture trained on PET neuroimaging modality data to be the best in terms of performance metrics which shows superior diagnostic power of this type of architecture.