TriAD: A Deep Ensemble Network for Alzheimer Classification and Localization
Francesco Mercaldo, Marcello Di Giammarco, Fabrizio Ravelli, Fabio Martinelli, Antonella Santone, Mario Cesarelli
Abstract
<italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Background and Objective</i> : Alzheimer’s disease is a hereditary neuro-degenerative disorder whose typical form causes amnestic cognitive impairment and less common variants cause non-amnestic cognitive impairment. The aim of the paper is to classify and localize magnetic resonance images of patients with Alzheimer’s disease, Mild Cognitive Impairment, and Cognitively Normal. <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Method</i> : We propose a method for the automatic detection and localization of Alzheimer’s disease, by proposing a neural network designed by authors i.e., TriAD. One of the distinctive features of the proposed network is its ability to process three simultaneous input MRI images of the brain corresponding to the three reference planes (axial, coronal and sagittal). <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Result</i> : The assembly of these three classifications has led to excellent quantitative results (95% in accuracy, precision, and recall) and punctual heatmaps for the localization of the region of interest in the images. Moreover, we show how visual explainability can increase the trustworthiness and reliability in the adoption of deep learning by the medical staff. <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Conclusions</i> : The main novelty, the TriaAD network, guarantees good performances from a quantitative and qualitative point of view, providing a very precise classification and localization of Alzheimer’s disease images.