Abstract
Deep convolutional neural networks (DCNNs) have achieved great success for image classification in medical research. Deep learning with brain imaging is the imaging method of choice for the diagnosis and prediction of Alzheimer's disease (AD). However, it is also well known that DCNNs are "black boxes" owing to their low interpretability to humans. The lack of transparency of deep learning compromises its application to the prediction and mechanism investigation in AD. To overcome this limitation, we develop a novel general framework that integrates deep leaning, feature selection, causal inference, and genetic-imaging data analysis for predicting and understanding AD. The proposed algorithm not only improves the prediction accuracy but also identifies the brain regions underlying the development of AD and causal paths from genetic variants to AD via image mediation. The proposed algorithm is applied to the Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset with diffusion tensor imaging (DTI) in 151 subjects (51 AD and 100 non-AD) who were measured at four time points of baseline, 6 months, 12 months, and 24 months. The algorithm identified brain regions underlying AD consisting of the temporal lobes (including the hippocampus) and the ventricular system.