Explainable Deep Learning for Alzheimer Disease Classification and Localisation

Abstract

Alzheimer’s disease is an irreversible neurological brain disorder that causes nuero-degenerative cognitive function like memory loss and thinking abilities. The accurate diagnosis of Alzheimer’s disease at an early stage is very crucial for patient care and conducting future treatment. Deep learning can help to reach the diagnosis: for this reason we propose a method aimed to distinguish and properly classify four Alzheimer disease’s stages. Two different deep learning models are exploited: Alex_Net and a model designed by authors, obtaining an average accuracy equal to 0.97 with the deep learning network developed by authors applying a colormap to brain magnetic resonance images. Our method provides also the localization areas used by the model to perform the classification (by adopting the heatmap overlapping provided by Gradient-weighted Class Activation Mapping algorithm) in order to ensures the explainability of the method.