Dementia classification from magnetic resonance images by machine learning

Dementia is a threatening condition that affects communication, thinking, and memory skills, being Alzheimer its most common type. The early detection of this disease allows for better care of the patient. Recently, Machine Learning (ML) methods have been developed to support the finding and forecast of Alzheimer’s disease through the analysis of Magnetic Resonance Images (MRI). Existing ML methods present some limitations: (i) require an expert to extract relevant features from MRI, (ii) depend on multistep image preprocessing, or (iii) need complex architectures and several images to train them. To surpass these limitations, in the present work, we analyze different Convolutional Neural Networks (CNNs) for Alzheimer’s classification, formulated to learn from a set of representative MRI sagittal images available in the Open Access Series of Imaging Studies (OASIS-2, 72 non-demented and 64 demented subjects, with ages from 60 to 96 years) and the Alzheimer’s Disease Neuroimaging Initiative (ADNI, 200 early Alzheimer and 200 control patients, with ages from 55 to 90 years) datasets. All CNNs were compared with state-of-the-art ML methods, being the VGG-16 variant the best performed architecture with an average validation accuracy of 56