Numerous deep learning approaches have been proposed to automatically classify Alzheimer’s disease (AD) from medical images. However, common approaches, such as convolutional neural networks (CNNs), lack interpretability and are prone to over-fitting when trained on small datasets. As an alternative, significantly less work has explored applying deep learning approaches to region-based features commonly obtained from atlas partitions of known regions of interest (ROI). In this paper, we propose a self-attention mechanism to jointly learn a graph of ROI connectivity as a prior for learning meaningful features for AD prediction. We apply our method to both the classification of AD subjects from healthy controls and to predict whether mild cognitive impaired (MCI) subjects will progress to AD (pMCI) or not (sMCI). We systematically show that our model’s performance compares well with other common ML approaches for ROI-based methods, such as neural networks and support vector machines. Finally, we perform exploratory graph analysis to illustrate the interpretability properties of the attention graphs and how they can provide insight for scientific discovery.
|