ADAMAEX—Alzheimer’s disease classification via attention-enhanced autoencoders and XAI

Abstract

To bring a new contribution in the area of classification of Alzheimer’s Disease (AD) we introduce a deep learning model, ADAMAEX, which is based on a convolutional autoencoder with four convolutions in the encoder part and a Squeeze and Excitation block for channel attention applied after each convolution. Additionally, we utilised fully connected layers (dense layers) for AD image classification. To conduct our study, we specifically chose axial brain scans acquired through sMRI in T2-weighted mode from the ADNI database, which were augmented using colour jitter, rotations, and flipping techniques. Before feeding the images to the model, we applied pre-processing steps such as re-sampling, normalisation, Contrast-Limited Adaptive Histogram Equalisation (CLAHE), and sharpening using the Unsharp Mask technique. For visualisation, we integrated Grad-CAM, an Explainable AI (XAI) technique, to highlight the brain regions responsible for the model’s classification decisions, a method underutilised by other authors in the context of AD classification. This model achieved an impressive accuracy of 96.2% and shows great promise for adoption in the medical sector, providing valuable assistance to doctors in validating their predictions based on brain scans.