Transformed domain convolutional neural network for Alzheimer's disease diagnosis using structural MRI

Abstract

Structural magnetic resonance imaging (sMRI) has become a prevalent and potent imaging modality for the computer-aided diagnosis (CAD) of neurological diseases like dementia. Recently, a handful of deep learning techniques such as convolutional neural networks (CNNs) have been proposed to diagnose Alzheimer's disease (AD) by learning the atrophy patterns available in sMRIs. Although CNN-based techniques have demonstrated superior performance and characteristics compared to conventional learning-based classifiers, their diagnostic performance still needs to be improved for reliable classification results. The drawback of current CNN-based approaches is the requirement to locate discriminative landmark (LM) locations by identifying regions of interest (ROIs) in sMRIs, thus the performance of the whole framework is highly influenced by the LM detection step. To overcome this issue, we propose a novel three-dimensional Jacobian domain convolutional neural network (JD-CNN) to diagnose AD subjects and achieve excellent classification performance without the involvement of the LM detection framework. We train the proposed JD-CNN model on the basis of features generated by transforming the sMRI from the spatial domain to the Jacobian domain. The proposed JD-CNN is evaluated on baseline T1-weighted sMRI data collected from 154 healthy control (HC) and 84 Alzheimer's disease (AD) subjects in the Alzheimer's disease neuroimaging initiative (ADNI) database. The proposed JD-CNN exhibits superior classification performance to previously reported state-of-the-art techniques.